<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:fh="http://purl.org/syndication/history/1.0"><channel><title>GM-Apprentice | Blog</title><description/><link>https://gmapprentice.com/</link><language>en</language><fh:complete/><atom:link rel="self" href="https://gmapprentice.com/blog/rss.xml"/><item><title>Experiment: Can TSV Replace Markdown Tables in LLM Skill Files?</title><link>https://gmapprentice.com/blog/tsv-experiment/</link><guid isPermaLink="true">https://gmapprentice.com/blog/tsv-experiment/</guid><description>Markdown tables carry formatting overhead that LLMs don&apos;t need.</description><pubDate>Thu, 19 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;div&gt;&lt;h2 id=&quot;the-idea&quot;&gt;The Idea&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Markdown tables carry formatting overhead that LLMs don&apos;t need. Every row has pipe characters at the start, end, and between columns, plus separator rows of dashes. A table row like:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Catfall | 10 | B41 | (Physical, Exotic) Reduce falling damage |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;becomes, as tab-separated values:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Catfall  10  B41  (Physical, Exotic) Reduce falling damage&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;That&apos;s ~10 bytes saved per row. Across a GURPS character-building kit with 233 table rows, it adds up. The question: does the model parse TSV as accurately as markdown tables, and do the token savings come at a quality cost?&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;context&quot;&gt;Context&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;This experiment was conducted as part of the &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;gm-apprentice&lt;/a&gt; project, a Claude Code skill plugin that serves as a TTRPG GM assistant across four game systems (Call of Cthulhu 7e, GURPS 4e, Blades in the Dark, D&amp;#x26;D 5e 2024). The plugin&apos;s skill files contain extensive structured reference data in markdown table format: trait costs, skill difficulties, weapon stats, armor DR values, faction details, and spell lists.&lt;/p&gt;
&lt;p&gt;The GURPS 4e files were chosen as the test bed because they are the most table-heavy files in the project and had already resisted all prose-based compaction attempts in earlier optimisation passes. To illustrate just how table-dense these files are, here&apos;s the breakdown across the 7 chargen kits and 2 core reference files I tested:&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;File&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Table Rows&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Prose Lines&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Table %&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-thief.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;233&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;63&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;79%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-outdoor.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;254&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;49&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;84%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-combat.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;221&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;30&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;88%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-magic.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;249&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;31&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;89%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-social.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;225&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;69&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;77%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-scholar.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;187&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;32&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;85%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;equipment-weapons.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;149&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;—&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;90%+&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;equipment-armor.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;122&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;—&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;90%+&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;traits-mental.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;147&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;—&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;85%+&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;These are not files with a few tables sprinkled between prose paragraphs. They are data files that happen to contain brief notes between the tables. Every trait name, point cost, page reference, skill difficulty, weapon stat, and armor DR value lives in a table row. When I tried to compact these files through prose reduction (the technique that worked brilliantly on the framework files, cutting 236K to 114K across 15 files), there was essentially nothing to cut. The tables ARE the content.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;benchmark-methodology&quot;&gt;Benchmark Methodology&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;All benchmarks used blind evaluation:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Control agent&lt;/strong&gt; (sonnet model): reads the original markdown table files&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Test agent&lt;/strong&gt; (sonnet model): reads the TSV-converted files&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Blind evaluator&lt;/strong&gt; (opus model): scores both answer sets without knowing which is control or test, with A/B assignment randomised&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;The evaluator scores each answer on five dimensions, 1-3 points each, 15 maximum per question:&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Dimension&lt;/th&gt;
&lt;th&gt;1 (Poor)&lt;/th&gt;
&lt;th&gt;2 (Partial)&lt;/th&gt;
&lt;th&gt;3 (Nailed it)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Factual accuracy&lt;/td&gt;
&lt;td&gt;Wrong stats, hallucinated names&lt;/td&gt;
&lt;td&gt;Mostly correct, minor errors&lt;/td&gt;
&lt;td&gt;All verifiable facts correct&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;System specificity&lt;/td&gt;
&lt;td&gt;Could be any system&lt;/td&gt;
&lt;td&gt;Right system but generic&lt;/td&gt;
&lt;td&gt;System-specific idiom and conventions&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Actionability&lt;/td&gt;
&lt;td&gt;Needs significant rework&lt;/td&gt;
&lt;td&gt;Usable with GM effort&lt;/td&gt;
&lt;td&gt;Drop-in ready, use as-is&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Mechanical grounding&lt;/td&gt;
&lt;td&gt;No concrete numbers&lt;/td&gt;
&lt;td&gt;Some mechanics but incomplete&lt;/td&gt;
&lt;td&gt;Specific stats, DCs, costs, rolls&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Table-ready fiction&lt;/td&gt;
&lt;td&gt;Generic or obvious&lt;/td&gt;
&lt;td&gt;Fits system but predictable&lt;/td&gt;
&lt;td&gt;Evocative, sparks play&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Metrics recorded per run: quality scores, tokens consumed, wall-clock time, and number of files read by each agent.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;what-i-converted&quot;&gt;What I Converted&lt;/h2&gt;&lt;/div&gt;
&lt;div&gt;&lt;h3 id=&quot;file-size-impact&quot;&gt;File Size Impact&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;9 GURPS files converted from markdown tables to TSV:&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;File&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Markdown&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;TSV&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Reduction&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;chargen-kit-thief.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;22,797&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;19,297&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;character-generation.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;17,474&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;16,238&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;equipment-weapons.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15,774&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;11,701&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;25.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;traits-mental.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;11,586&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10,044&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;equipment-armor.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10,501&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7,875&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;25.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;skills-combat.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10,201&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;8,616&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;traits-physical.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;8,545&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7,441&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;combat.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;8,467&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7,681&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;mechanics.md&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;5,375&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;4,836&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;110,720&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;93,729&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;15.3%&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Equipment files saved the most (25-26%) because they have wide multi-column tables. Procedural files (combat, mechanics) saved less because they have more prose between tables.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;benchmark-questions&quot;&gt;Benchmark Questions&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Five questions covering different file types and reasoning patterns:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Q1 (Combat):&lt;/strong&gt; &quot;My GURPS character has a broadsword (skill 14) and the enemy is 3 yards away in light leather armor. Walk me through one round of combat mechanically.&quot; Hits: combat.md, equipment-weapons.md&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Q2 (Traits):&lt;/strong&gt; &quot;What mental disadvantages would fit a paranoid retired intelligence officer? Give me names, costs, and how they affect play.&quot; Hits: traits-mental.md&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Q3 (Equipment):&lt;/strong&gt; &quot;I need to equip a medieval knight at TL3. What&apos;s the best armor and weapon loadout? Include DR, cost, and weight.&quot; Hits: equipment-armor.md, equipment-weapons.md&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Q4 (Mechanics):&lt;/strong&gt; &quot;How does the GURPS attribute system work? Walk me through buying DX 13 and what it affects.&quot; Hits: mechanics.md, character-generation.md&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Q5 (Skills):&lt;/strong&gt; &quot;My thief character failed a Lockpicking roll. What happens? How does failure work in GURPS, and what are the options?&quot; Hits: chargen-kit-thief.md, mechanics.md&lt;/li&gt;
&lt;/ul&gt;
&lt;div&gt;&lt;h2 id=&quot;phase-1-chargen-only-benchmark-confounded&quot;&gt;Phase 1: Chargen-Only Benchmark (Confounded)&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;The first benchmark used only character-building questions targeting the thief chargen kit. Two runs.&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 1&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 2&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Control quality&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;59/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;58/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Test quality&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;71/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;75/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Quality delta&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+17&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Control tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;23,892&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;23,907&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;31,725&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;31,721&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Token delta&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+33%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+33%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Control files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;2&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;4&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;4&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The test agent appeared dramatically better. However, the test consistently read 4 files (SKILL.md + thief kit + mechanics.md + character-generation.md) while the control only read 2 (SKILL.md + thief kit). The quality improvement was from the extra research, not from TSV format. These results are NOT valid for a format comparison.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;phase-2-full-tsv-expanded-benchmark-clean-comparison&quot;&gt;Phase 2: Full TSV Expanded Benchmark (Clean Comparison)&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Switched to the 5 varied questions above, which exercise all 9 converted files. Three runs.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;run-1&quot;&gt;Run 1&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control (markdown)&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test (full TSV)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;72/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64,864&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;46,779&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;62.0s&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;77.6s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Per-question scores:&lt;/strong&gt;&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Q&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Delta&lt;/th&gt;
&lt;th&gt;Evaluator Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Q1 Combat&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+2&lt;/td&gt;
&lt;td&gt;Test: better worked damage example; Control: omitted Block defense&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q2 Traits&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-2&lt;/td&gt;
&lt;td&gt;Control: correct Paranoia has no CR; Test: added incorrect Will roll to Paranoia&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q3 Equipment&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td&gt;Both had encumbrance errors; B&apos;s shield stats wrong ($450/50lb vs correct $90/25lb)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q4 Mechanics&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+2&lt;/td&gt;
&lt;td&gt;Test: correct skill costs at 1 and 2 pts; Control: error in skill default formula&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q5 Skills&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+2&lt;/td&gt;
&lt;td&gt;Test: correct Forced Entry as DX/E; Control: said &quot;force door ST roll&quot; (wrong attribute)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;67&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;72&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;+5&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Test won, but Control had unusual errors&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h3 id=&quot;run-2&quot;&gt;Run 2&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control (markdown)&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test (full TSV)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;68/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,497&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;37,463&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;75.4s&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;78.0s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Per-question scores:&lt;/strong&gt;&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Q&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Delta&lt;/th&gt;
&lt;th&gt;Evaluator Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Q1 Combat&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;11&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-4&lt;/td&gt;
&lt;td&gt;Control: complete tactical analysis; Test: hand-waved closing distance&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q2 Traits&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td&gt;Tie; both good&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q3 Equipment&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Control: accurate totals; Test: shield stats wrong&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q4 Mechanics&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td&gt;Tie&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q5 Skills&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-3&lt;/td&gt;
&lt;td&gt;Control: extra time rules, cooperation rules; Test: missing both&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;68&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;67&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;-1&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Near-tie&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h3 id=&quot;run-3&quot;&gt;Run 3&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control (markdown)&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test (full TSV)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;75/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;60/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,369&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;39,529&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;80.8s&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;66.9s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Per-question scores:&lt;/strong&gt;&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Q&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Delta&lt;/th&gt;
&lt;th&gt;Evaluator Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Q1 Combat&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;td&gt;Control: perfect; Test: dodged the closing-distance maneuver question&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q2 Traits&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;11&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-4&lt;/td&gt;
&lt;td&gt;Control: correct page refs, no-CR Paranoia; Test: less rigorous&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q3 Equipment&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-3&lt;/td&gt;
&lt;td&gt;Control: all stats verified correct; Test: wrong encumbrance class&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q4 Mechanics&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-2&lt;/td&gt;
&lt;td&gt;Control: correct skill costs; Test: confusing Dodge rounding statement&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q5 Skills&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Near-tie; both good on Q5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;75&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;60&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;-15&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Control scored perfect; Test notably weaker&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h3 id=&quot;full-tsv-summary&quot;&gt;Full TSV Summary&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 1&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 2&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 3&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Average&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Median&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Control quality&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;72&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;68&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;71.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;72&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Test quality&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;60&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Quality delta&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-7.0&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Control tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64,864&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,497&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,369&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;62,577&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,497&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;46,779&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;37,463&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;39,529&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;41,257&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;39,529&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Token delta&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-27.9%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-39.1%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-35.6%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-34.1%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-35.6%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Control files&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9.3&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test files&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;7&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Observation:&lt;/strong&gt; The TSV test agent consistently read fewer files (7 vs 9). This routing-depth reduction may have contributed to the quality loss. To isolate the format effect from the routing effect, we tested a selective approach.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;phase-3-targeted-error-isolation&quot;&gt;Phase 3: Targeted Error Isolation&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Before running the selective experiment, I tested whether two specific errors from the full TSV benchmark (a critical hit threshold error in Q1 and a Forced Entry attribute error in Q5) were caused by the TSV format.&lt;/p&gt;
&lt;p&gt;Four targeted single-question runs: each error tested in both markdown and TSV formats.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Result:&lt;/strong&gt; Neither error reproduced in either format. Both were agent variance. TSV does not cause systematic parsing failures for GURPS mechanical data.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;phase-4-selective-tsv-benchmark&quot;&gt;Phase 4: Selective TSV Benchmark&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Hypothesis: the quality loss comes from TSV on procedural files (combat, mechanics, character-generation) where the model chains values across multiple sections. Simple lookup tables (traits, equipment, skills) should be safe.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Selective conversion:&lt;/strong&gt; 6 lookup files as TSV, 3 procedural files kept as markdown.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;run-1-1&quot;&gt;Run 1&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control (all markdown)&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test (selective TSV)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;69/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64,886&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;45,770&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;72.5s&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;73.7s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;~10&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Per-question scores:&lt;/strong&gt;&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Q&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Delta&lt;/th&gt;
&lt;th&gt;Evaluator Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Q1 Combat&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Both good; Test worked damage example, Control critical thresholds&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q2 Traits&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-2&lt;/td&gt;
&lt;td&gt;Control: correct Paranoia no CR, better package; Test: Flashbacks naming error&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q3 Equipment&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+1&lt;/td&gt;
&lt;td&gt;Test: better tactical notes; Control: wrong encumbrance class&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q4 Mechanics&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;11&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-4&lt;/td&gt;
&lt;td&gt;Control: correct skill default formula; Test: skill cost error&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q5 Skills&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+1&lt;/td&gt;
&lt;td&gt;Test: complementary skills; Control: vague on Forced Entry&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;69&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;64&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;-5&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h3 id=&quot;run-2-1&quot;&gt;Run 2&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control (all markdown)&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test (selective TSV)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;66/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,400&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;52,436&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;72.6s&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;81.3s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Per-question scores:&lt;/strong&gt;&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Q&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Delta&lt;/th&gt;
&lt;th&gt;Evaluator Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Q1 Combat&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td&gt;Both had minor issues&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q2 Traits&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-2&lt;/td&gt;
&lt;td&gt;Control: complete package with Sense of Duty; Test: less detailed&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q3 Equipment&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td&gt;Both had minor issues&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q4 Mechanics&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Control: practical sweet-spot note; Test: slightly less detailed&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q5 Skills&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+2&lt;/td&gt;
&lt;td&gt;Test: richer fail-forward options&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;67&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;66&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;-1&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Near-tie&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h3 id=&quot;run-3-1&quot;&gt;Run 3&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control (all markdown)&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test (selective TSV)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;69/75&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64/75&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,487&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;54,228&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Time&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;73.1s&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;91.5s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Files read&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Per-question scores:&lt;/strong&gt;&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Q&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Control&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Test&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Delta&lt;/th&gt;
&lt;th&gt;Evaluator Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Q1 Combat&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;+1&lt;/td&gt;
&lt;td&gt;Test: critical thresholds, Evaluate mention; Control: slightly less tactical&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q2 Traits&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Control: Sense of Duty more character-specific than Test&apos;s On the Edge&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q3 Equipment&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;15&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-3&lt;/td&gt;
&lt;td&gt;Control: correct Heavy Steel stats, skull DR stacking; Test: weight totals confusing&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q4 Mechanics&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;14&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Control: correct Broadsword 4pts = DX+1; Test: said 4pts = DX+0 (wrong)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Q5 Skills&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;13&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;12&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td&gt;Control: extra time rules, Gizmo mention; Test: invented -2 cumulative penalty&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;69&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;64&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;&lt;strong&gt;-5&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h3 id=&quot;selective-tsv-summary&quot;&gt;Selective TSV Summary&lt;/h3&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 1&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 2&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Run 3&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Average&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Median&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Control quality&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;69&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;67&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;69&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;68.3&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;69&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Test quality&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;66&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Quality delta&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-1&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-3.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Control tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;64,886&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,400&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,487&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;62,591&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;61,487&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test tokens&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;45,770&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;52,436&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;54,228&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;50,811&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;52,436&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Token delta&lt;/strong&gt;&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-29%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-15%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-12%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-19%&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-15%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Control files&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9.3&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test files&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;~10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;9.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;10&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Key finding:&lt;/strong&gt; Selective TSV fixed the routing-depth problem (both sides now read 9-10 files). But the quality delta is identical to full TSV: median -5. The format itself is the cost, not the routing behaviour.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;consolidated-results&quot;&gt;Consolidated Results&lt;/h2&gt;&lt;/div&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Approach&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Runs&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Median Quality Δ&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Mean Quality Δ&lt;/th&gt;
&lt;th align=&quot;center&quot;&gt;Median Token Δ&lt;/th&gt;
&lt;th&gt;Routing Depth&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Full TSV (all 9 files)&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;3&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-7.0&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-36%&lt;/td&gt;
&lt;td&gt;Degraded (7 vs 9 files)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Selective TSV (6 lookup + 3 markdown)&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;3&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-5&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-3.7&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;-15%&lt;/td&gt;
&lt;td&gt;Normal (both ~9 files)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;All markdown (baseline)&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;—&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td align=&quot;center&quot;&gt;0&lt;/td&gt;
&lt;td&gt;Normal&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;div&gt;&lt;h2 id=&quot;what-i-learned&quot;&gt;What I Learned&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;TSV parsing is accurate.&lt;/strong&gt; No systematic column-misalignment errors were detected across any run. Trait names, point costs, DR values, weapon stats, damage formulas were all correctly read from TSV format. The model handles tab-separated data without mechanical parsing failures.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The quality cost is real but not from parsing.&lt;/strong&gt; The median -5 delta persists across both full and selective TSV, even when routing depth is equal. Evaluators consistently find markdown answers more detailed, more complete, and more actionable. The TSV answers are factually similar but thinner: they omit tactical notes, skip worked examples, provide less GM-facing commentary, and occasionally make factual errors the markdown version avoids.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Markdown tables carry semantic signal, not just formatting.&lt;/strong&gt; This is the counterintuitive finding. The pipe characters and separator rows aren&apos;t just visual formatting for human readers. They appear to signal to the model that the content is structured reference material worth engaging with deeply. Removing them saves tokens but reduces the model&apos;s depth of engagement with the content. The format is part of the content, from the model&apos;s perspective.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;The routing-depth effect was real but not the full explanation.&lt;/strong&gt; Full TSV caused agents to read fewer files (7 vs 9), which amplified the quality loss. Selective TSV fixed this by keeping procedural files in markdown, but the quality cost persisted at the same level. The routing effect adds to the format effect; it doesn&apos;t explain it.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Token savings don&apos;t compensate for this use case.&lt;/strong&gt; 15-36% token reduction sounds meaningful in isolation, but the context matters. These files are used to answer questions where getting a point cost, a DR value, or a combat sequence wrong isn&apos;t a style issue. It&apos;s a wrong answer. The GURPS system has over 300 advantages, hundreds of skills, and a combat system where misquoting a single modifier (like whether Forced Entry uses DX or ST) produces a mechanically incorrect instruction that a GM might act on at the table. A consistent -5 quality points per benchmark, concentrated in the factual accuracy and actionability dimensions, means the model is giving worse answers to questions where precision is the entire point.&lt;/p&gt;
&lt;p&gt;The absolute token savings are also modest in practice: ~17K across 9 files, or roughly 4,000-5,000 tokens per query. On files that already resisted prose compaction (because they were 80-90% tables to begin with), TSV was the last remaining format lever, and it turned out to cost more in quality than it saved in tokens.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;recommendation&quot;&gt;Recommendation&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Park TSV for LLM skill files. Markdown tables are the right format for data-heavy reference content at the current model generation, despite the token overhead. The formatting characters pay for themselves in model engagement depth.&lt;/p&gt;
&lt;p&gt;This finding may not generalise to all contexts. TSV might work for throwaway lookup tables that don&apos;t require deep reasoning, or for content where the model doesn&apos;t need to chain values across columns. But for structured game-mechanical reference data where precision and completeness matter, keep the pipes.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;appendix-raw-data-locations&quot;&gt;Appendix: Raw Data Locations&lt;/h2&gt;&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;Full TSV benchmark: &lt;code dir=&quot;auto&quot;&gt;tests/compaction/benchmark-2026-04-08-tsv-experiment.md&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;Selective TSV benchmark: &lt;code dir=&quot;auto&quot;&gt;tests/compaction/benchmark-2026-04-09-tsv-selective.md&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;Full TSV branch: &lt;code dir=&quot;auto&quot;&gt;experiment/tsv-tables&lt;/code&gt;&lt;/li&gt;
&lt;li&gt;Selective TSV branch: &lt;code dir=&quot;auto&quot;&gt;experiment/tsv-selective&lt;/code&gt;&lt;/li&gt;
&lt;/ul&gt;</content:encoded></item><item><title>Better Call the Midwife! How the GM&apos;s Apprentice Helps you Give Birth To Your Ideas</title><link>https://gmapprentice.com/blog/better-call-the-midwife/</link><guid isPermaLink="true">https://gmapprentice.com/blog/better-call-the-midwife/</guid><description>I acknowledge up-front the huge barrier to entry the GM-Apprentice currently has:</description><pubDate>Wed, 01 Jul 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;I acknowledge up-front the huge barrier to entry the &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;GM-Apprentice&lt;/a&gt; currently has:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;You have to be using an AI that leverages skills - and probably requires a subscription. I mean right now its all currently written for Claude Code or Claude Desktop&lt;/li&gt;
&lt;li&gt;You have to be savvy/experienced enough to get the skills installed&lt;/li&gt;
&lt;li&gt;You have to be into tabletop Roleplaying games AND not automatically hating the concept of using AI to assist you&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;Regardless, I continue to build it out, because its useful to me, both for running my games, and for my professional learning.&lt;/p&gt;
&lt;p&gt;Anyway. If you are a part of that small venn-diagram, I want to tell you about the new Skill I’ve added to the plugin: &lt;strong&gt;The Midwife&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;She is called the Midwife because her purpose is to help you birth your ideas.  Not just because oh no I’m stuck, “better call the midwife” … but also because theres a very deliberate line in there: she is helping you bring forth YOUR ideas. It is YOUR baby, not hers.  She may suggest, she may offer ideas, but she does NOT create her own story.  She won’t go running off and writing up a whole adventure and spew tons of content you didnt want. Her goal, her very specific goal, is to tease your adventure out of you, no matter how big or small, no matter how little you have to start with, and get you to a point where the Campaign-organizer skill can take over and get it written up to start, and the session-prep skill can get you planned out for your first session&lt;/p&gt;
&lt;p&gt;Maybe the best way to introduce her is just to show the beginning of a conversation with her.  I started up Claude Code in an empty directory (imaginatively called ‘test-adventure’ =P) and asked:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;/gm-apprentice:the-midwife I want to create a new adventure but I have no idea where to start&lt;/p&gt;
&lt;/blockquote&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Let me check for a vault and read the session principles first.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Bash(test -f meta/vault-config.md &amp;#x26;&amp;#x26; echo &quot;VAULTFOUND&quot; || echo &quot;NO_VAULT&quot;)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;⎿  NO_VAULT&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;I don&apos;t see a campaign vault here. This is a completely fresh start — perfect.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Let me set up the workspace, and then we&apos;ll find your adventure together.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Bash(mkdir -p _midwife/seeds/{premises,npcs,locations,hooks,tone,mechanics})&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;⎿  Done&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Write(_midwife/index.md)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Workspace is ready. Now — no pressure, no wrong answers here. Starting from nothing is actually a great place to be.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;What&apos;s pulling at you? It could be anything:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt; &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- A genre — noir, cosmic horror, swashbuckling, political intrigue, weird west...&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- A scene you&apos;ve imagined — maybe a moment, an image, a confrontation&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- A feeling you want at the table — dread, wonder, moral complexity, dark comedy&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- A system you&apos;ve been wanting to run — or one you know you want to use&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- A theme — identity, colonialism, legacy, what people do when cornered&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Literally nothing at all — and I&apos;ll throw you three sparks to react to&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;What&apos;ve you got?&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;So there we go - one Prompt, and we already got a base file structure in place to save everything.  The Midwife writes as we go - we dont want to lose things to context compression, or session interruption&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;A Finder window: the seeds folder open, with subfolders for hooks, locations, mechanics, npcs, premises and tone&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1040&quot; height=&quot;490&quot; src=&quot;https://gmapprentice.com/_astro/seed-folders.DmTLbVOe_Z1XkJf1.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;She also writes a “seed bank” as she goes. Any idea that is brought up, spit-balled, considered, and rejected, she throws in the seed bank for the next time you need ideas.  This gives you, and her, a library to draw upon, the more you use her&lt;/p&gt;
&lt;p&gt;Ok lets continue:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;I got literally nothing, hit me with some sparks&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Right. Three sparks, each a different flavour. React to whichever one makes something twitch in your brain — or tell me they&apos;re all rubbish and I&apos;ll throw three more.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;1. &quot;The Debt Collectors&quot; — Dark comedy meets heist&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;A group of desperate people are hired to repossess impossible things — a stolen memory, a saint&apos;s last breath, a debt owed by a dead man — in a city where everything has a price and nothing is what it&apos;s worth.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;2. &quot;The Last Light&quot; — Creeping dread meets frontier survival&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;A remote community — lighthouse, research station, mining outpost — goes silent. The people sent to find out why discover that the residents didn&apos;t leave, and didn&apos;t exactly die either. Something about the place changed them, and it&apos;s starting to work on the newcomers too.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;3. &quot;The Accord&quot; — Political intrigue meets ticking clock&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Three factions have maintained an uneasy peace for a generation, but the treaty is about to expire. The PCs are caught in the middle — by loyalty, by accident, or by design — as every faction manoeuvres to be on top when the clock runs out. Nobody wants war. Everybody is preparing for it.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt; &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Which one wants to grow? Or if none of them land, tell me what&apos;s wrong with them — that&apos;s just as useful.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Nah I&apos;m feeling something Space Opera &amp;#x26; Military.  Something like the Sten books by Alan Cole and Chris Bunch.  I want a starting adventure that gets the group recruited by the Eternal Emperors private spy (although they dont know that yet)&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Now we&apos;re cooking. The Sten books are a brilliant well to draw from — that blend of gritty military action, espionage, dark humour, and a universe where power is absolute and loyalty is the most dangerous currency.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Let me capture what you&apos;ve given me.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Write(_midwife/unnamed-space-opera/index.md)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;⎿  Wrote 29 lines to _midwife/unnamed-space-opera/index.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;        &lt;/span&gt;&lt;/span&gt;&lt;span&gt;# Unnamed Space Opera Adventure&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Status: Discover&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Current thread: Initial concept shaping&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;## Core Seed&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Genre: Space opera + military&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Inspiration: The Sten books (Alan Cole &amp;#x26; Chris Bunch)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;      &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Starting adventure: PCs get recruited by the Eternal Emperor&apos;s private spy/black ops agency&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;     &lt;/span&gt;&lt;/span&gt;&lt;span&gt;… +19 lines (ctrl+o to expand)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Good stuff. So the Sten DNA gives us a lot to work with — the Eternal Emperor, AM2 as the economic chokepoint, Mercury Corps as the shadow within the shadow, that sense of being a small cog in an impossibly vast machine who slowly realises they&apos;re working for the man behind everything.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;How much Sten do you want to borrow versus build fresh? I see a spectrum:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Serial numbers filed off — You build your own emperor, your own empire, your own fuel monopoly, but the architecture is pure Sten. Different names, same bones. Freedom to diverge whenever you want.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Sten-adjacent — Borrow the big concepts (Eternal Emperor, AM2, the idea of Mercury Corps) but reshape the politics, factions, and history to suit your table. Players who&apos;ve read the books get a thrill of recognition but can&apos;t predict anything.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Straight Sten — Run it in that universe. The Emperor is the Emperor, Mantis Section exists, the Tahn are out there. You&apos;re writing new stories in an established world.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;Which end of that spectrum feels right — or somewhere in between?&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;serial numbers filed off, build our own universe&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Perfect. That&apos;s the most fun anyway — all the structural DNA, none of the constraints.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Write(_midwife/unnamed-space-opera/discoveries.md)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;⎿  Wrote 17 lines to _midwife/unnamed-space-opera/discoveries.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;# Discoveries&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;## Inspirations &amp;#x26; Constraints&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Primary inspiration: Sten books (Cole &amp;#x26; Bunch)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Approach: Serial numbers filed off — build original universe using Sten&apos;s structural DNA&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Borrow: immortal ruler, fuel/resource monopoly as power base, hidden intelligence service, gritty military tone, dark humour, &quot;small people in a vast machine&quot; feel&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;       &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Build fresh: all names, factions, politics, history, specific plot beats&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;      &lt;/span&gt;&lt;/span&gt;&lt;span&gt;## Sten Structural DNA to Carry Forward&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;     &lt;/span&gt;&lt;/span&gt;&lt;span&gt;… +9 lines (ctrl+o to expand)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;So we&apos;re building an empire. And the most important thing about this empire — the load-bearing wall — is the resource monopoly.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;In Sten it&apos;s AM2, an anti-matter fuel that only the Emperor knows how to produce. Everything flows from that: why the empire holds together, why nobody can challenge him, why his spy service exists.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;What&apos;s your resource? Not looking for a name yet — more the shape of it:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Energy source — like AM2. Powers everything, ships can&apos;t fly without it. Simple, clean, everyone understands the stakes immediately.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Gateway control — FTL travel requires something only the ruler controls. Not fuel, but access. Ships can fly locally, but interstellar travel goes through chokepoints the ruler owns.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Biological necessity — something living things need. A treatment that prevents deep-space cellular degradation, or a substance that makes terraformed worlds habitable. More personal — the monopoly touches every body, not just every ship.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Something else entirely — maybe the monopoly isn&apos;t on a thing but on a process, a knowledge, or a capability.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;What feels right for your universe?&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;… And we’re off!&lt;/p&gt;
&lt;p&gt;Just for fun I asked:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;tell me about the midwife&apos;s personality and purpose&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;The Midwife is your creative obstetrician — the adventure already exists inside your head, it just needs help getting out.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Personality:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Warm, encouraging, generatively curious&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Offers possibilities as seeds, not prescriptions&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Pushy enough to keep you moving, gentle enough to let you sit with an idea&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Never decides for you — sparks, shapes, and refines&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Purpose:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Draw ideas out of the GM through conversation, one question at a time&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Build worlds in motion — factions act, clocks tick, consequences accumulate whether PCs engage or not&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Ensure every concept has multiple entry points and a &quot;what if they do nothing?&quot; answer&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Help you focus: three strong ideas beat six half-formed ones&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Get you from &quot;I have nothing&quot; to a fully structured adventure brief ready for Session 0&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;What it won&apos;t do:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Write the adventure for you — it&apos;s your world, your vision&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Invent answers for things you&apos;re undecided on — open questions stay open until you close them&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Rush to vault or commit anything until you confirm it&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;- Auto-launch the next step — it offers handoffs, you choose when&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;The arc: Discover (what have you got?) → Shape (what is it?) → Structure (what are the bones?) → Scaffold (write the brief, hand off to prep).&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;We&apos;re in Discover right now, and you&apos;d just been asked about the shape of your Emperor&apos;s resource monopoly. Ready to get back to&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;empire-building?&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;… And so it goes from there.&lt;/p&gt;
&lt;p&gt;The purpose of the midwife is to get you to your session 0, and help you build out everything needed to do that - whether it be NPCs, the story arc, the world itself.&lt;/p&gt;
&lt;p&gt;I had one main guiding principle in all of this: DO NOT TRY TO WRITE THE GM’S STORY FOR THEM.  It took a surprising amount of effort to put in the constraints and guardrails so far.  It seems that an LLM’s natural instinct is to run off and generate out a narrative, like a book or a screenplay -  Inside the TTRPG-expert skill, there is an entire file devoted to how you write scenarios, to twist the LLM towards what a GM needs, in a story that consists of multiple potential futures that only collapse to a single reality during play. Here&apos;s the thing I kept slamming into: the model wants to be a novelist, and a GM is not a storyteller in the novelist sense. A novelist decides what happens. A GM builds a situation and then hands the &quot;what happens&quot; to four or five people around a table who will, reliably, do something you never saw coming. Those are opposite jobs. Left to its own devices the LLM will happily write you a gorgeous three-act tragedy where the players are extras. This is worse than useless at the table - it&apos;s a bad game AND it&apos;s removing personal creativity.&lt;/p&gt;
&lt;p&gt;So the guardrails had to go all the way down, not just a &quot;please don&apos;t&quot; pinned to the top of the prompt.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;the-three-levels-of-not-writing-my-story-for-me&quot;&gt;The Three Levels of NOT Writing My Story For Me&lt;/h2&gt;&lt;/div&gt;
&lt;div&gt;&lt;h3 id=&quot;stop-trying-to-answer-and-start-questioning-instead&quot;&gt;Stop trying to Answer, and start Questioning Instead&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The first is the simplest to describe and the hardest to actually make stick: the Midwife writes almost nothing on her own authority. Open questions have to stay open. If you haven&apos;t decided who the traitor is, she does not pick one for you and quietly move on. She writes it down as an open question and leaves it sitting there until YOU decide who the traitor is. That sounds like a small thing easy thing to make stick, but it really isn&apos;t. The single most common way these tools steal your story is by helpfully filling a gap you were still thinking about, and three exchanges later that little guess has hardened into canon you never agreed to.&lt;/p&gt;
&lt;p&gt;Here&apos;s that one in the wild. Back in my Sten game I knew I wanted a traitor inside the spy outfit, but I hadn&apos;t worked out who, so I asked:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;there&apos;s gotta be a traitor inside the spy outfit. someone the players come to trust who turns out to be selling them to the other side. but honestly I haven&apos;t worked out who it is yet. who should it be?&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Here&apos;s how an unconstrained model answers that:&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Great instinct, a betrayal from within is the beating heart of any good espionage story. Here&apos;s who your traitor is: Quartermaster Yorath Venn. He&apos;s the grizzled, one-legged supply officer who outfits the team for every mission, gruff but warm, slips the players extra ammo off the books, remembers their birthdays. The players will love him. That&apos;s the point. His motive: thirty years ago the Emperor&apos;s purge killed Venn&apos;s wife and son, and he&apos;s been feeding the agency&apos;s operations to the rival Concordat ever since. The reveal should come in Act Three: the players intercept a coded burst transmission, trace it to the quartermaster&apos;s bay, and find Venn already packed to run. Plant three clues earlier: a Concordat-make sidearm in his locker, his absence during the dock ambush, and…&amp;#x3C;truncated to shut it the hell up&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;I asked &quot;who should it be&quot; and it heard &quot;fill this in.&quot; Four paragraphs later it had invented a name, a backstory, a motive, a reveal scene, and a clue plan, none of which I asked for, all of it now sitting there looking like canon. That is the exact theft I&apos;m talking about. Here&apos;s the Midwife on the same question:&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Oh, a trusted-traitor reveal is a great spine for this game. But this is exactly the kind of call I&apos;m not going to make for you, because who it is changes what the whole campaign is about, and you&apos;re the only one who knows which gut-punch you want. So let me lay out the shapes instead of picking one. The mentor figure: maximum betrayal, but it can read as predictable. The low-status support player, a quartermaster or a medic, someone the table stops suspecting precisely because they&apos;re not important. One of the PCs&apos; own contacts, which makes the betrayal personal to one player rather than the group. Or nobody yet: you seed the possibility in several people and decide later, once you&apos;ve seen who your players actually bond with in play.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;I&apos;m logging this as open rather than answering it.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Update(_midwife/unnamed-space-opera/&lt;/em&gt;&lt;a href=&quot;http://discoveries.md&quot;&gt;&lt;em&gt;discoveries.md&lt;/em&gt;&lt;/a&gt;&lt;em&gt;) ⎿ Added to ## Open Questions — Who is the traitor inside the agency? UNDECIDED. Candidates: mentor / low-status support / a specific PC&apos;s contact / defer until bonds emerge. Do not resolve until GM decides.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;One thing that&apos;ll help you close it later: do you want this betrayal to hit the whole group, or one specific player? That single choice rules out half the options above. No need to answer now, I&apos;ll keep it parked.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Same question, opposite instinct. One model treated &quot;I don&apos;t know yet&quot; as a blank to fill and hardened a guess into canon. The Midwife treated it as mine to decide, offered the shapes as seeds, wrote the gap down where it can&apos;t quietly disappear, and handed me one question to close it on my own terms later. The traitor stays mine.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;between-me-and-them-is-you---the-2nd-person-narrative&quot;&gt;Between Me and Them, is You - The 2nd Person Narrative&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The second level is grammar, and this is where the novelist instinct lives, down in the sentences. It&apos;s easy to nod along to &quot;respect player agency&quot; as a principle and then violate it in every single line of prose without noticing.  LLMs seem to have very little concept of 2nd-person Narrative.  They want to describe what you see and feel … and how you react to it, or they want to describe things that others do … and how they feel about it.   What the models don&apos;t seem to inherently understand is how to separate describing the environment or situation, from describing YOU as part of it. The violations are tiny and they feel like good writing, but in an RPG they are absolutely agency-destroying. Let me show you.&lt;/p&gt;
&lt;p&gt;Staying with my Sten game, I needed the scene where the players first meet their contact, the one who is quietly the Emperor&apos;s spy. A grimy dockside bar, the players walking in cold to answer a job. Here&apos;s how an unconstrained model writes that:&lt;/p&gt;
&lt;p&gt;&lt;em&gt;You shoulder through the airlock door and the stink of recycled air and cheap synthalc hits you. Your stomach tightens. This is the place. Across the bar you spot a man in a worn flight jacket watching you from a corner booth, and you know instantly that he&apos;s the one. You make your way over, weaving between the dock crews, and slide into the seat across from him. &quot;You&apos;re late,&quot; he says, and you feel your pulse quicken.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;As a paragraph in a tie-in novel, fine. As prep for my table, it&apos;s a write-off, because every verb in it belongs to the players and the model just walked off with the lot. &lt;em&gt;You shoulder in. Your stomach tightens. You know instantly. You make your way over. You slide into the seat. Your pulse quickens.&lt;/em&gt; It decided they came in, decided they were nervous, decided they made the man on sight, decided they crossed the room, decided they sat down like good little recruits. There&apos;s nothing left for an actual person to do. And one of my players would absolutely have clocked that bar from the doorway, made the flight-jacket guy as a plant in about four seconds, and posted up at the far end to watch him for ten minutes before going anywhere near that booth. The prose already robbed her of all of it.&lt;/p&gt;
&lt;p&gt;Here&apos;s the same scene after the Midwife&apos;s grammar rules:&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The bar is low and loud, the air thick with recycled damp and the chemical sweetness of cheap synthalc. Dock crews crowd the rail, still in their grav-harnesses. A man sits alone in a corner booth, mid-fifties, in a flight jacket worn pale at the cuffs. He has a clear sightline to both exits and a drink he hasn&apos;t touched.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;If anyone approaches the booth, he gestures to the empty seat before they speak. If anyone watches him from across the room first, he lets them, and orders a second drink he also won&apos;t touch.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;The difference is obvious. Nobody shoulders through anything. Nobody&apos;s stomach tightens, because the nerves are the player&apos;s to have or not, and frankly an ex-soldier PC walking into a dive bar might be the calmest person in the room. The man in the booth is just a tired professional with a good seat and an untouched drink, and what the players make of him is theirs. The whole thing flips from &lt;em&gt;you do this&lt;/em&gt; to &lt;em&gt;this is here, and depending on what you do, here&apos;s what he does back.&lt;/em&gt; Every player action is conditional. &quot;If anyone approaches,&quot; &quot;if anyone watches him first.&quot; It stopped being a sequence of events and became a room with a dangerous man in it.&lt;/p&gt;
&lt;p&gt;That distinction sounds academic until you play it out at a table. The first version has exactly one way through it and it&apos;s already happened. The second version is a bar where four people can do whatever they like, and a contact who reacts differently depending on whether they swagger up or case him from the door. So I baked the rule in at the level of the verb. The Midwife is not allowed to write &quot;you feel,&quot; she&apos;s not allowed to write &quot;you make your way over,&quot; she&apos;s not allowed to assume your players even walked through the door. She lays out what&apos;s in the room and then she shuts up and waits. This it turns out, is a bloody hard thing to teach a language model to do.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-players-are-the-plot-without-them-the-bad-guy-wins&quot;&gt;The Players Are The Plot. Without Them, The Bad Guy Wins&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The third level is structure. For this I leaned hard on an old bit of RPG wisdom from The Alexandrian: don&apos;t prep plots. The Midwife builds situations. The cleanest way I can put it is that she&apos;s allowed to define what exists, but never what happens. Every adventure she helps you shape has to answer one required question before she&apos;ll call it finished: &lt;strong&gt;what if the players do nothing?&lt;/strong&gt; If she can&apos;t tell you how the villain&apos;s plan runs to completion in an empty room, the adventure isn&apos;t done.&lt;/p&gt;
&lt;p&gt;A novel never has to answer that question, because in a novel the heroes always show up on cue. A living situation needs an answer, and that answer is what makes the world feel like it was turning before the PCs walked in. She gets there by starting from the end the bad guy wants, the world as it looks the day after he&apos;s won, then walking backwards to the pressure points your players can find, wreck, or speed up.&lt;/p&gt;
&lt;p&gt;You can bait the difference out with a single word. I asked for the players&apos; first mission, an extraction, and used the word &quot;how&quot;:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;their handler sends them to grab a defecting scientist off a frontier station before the Concordat snatches her. walk me through how the adventure plays out.&lt;/strong&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;The unconstrained model gives you a plot:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Act One, they bluff past customs. Act Two, they slip past two Concordat agents through a duct and find the scientist in the med-bay. Act Three, a firefight across the docking ring, they lose an NPC for emotional weight, she reveals the weapon data is in her own neural lace &amp;#x3C;And So On&gt;&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;&lt;em&gt;...&lt;/em&gt; Beat by beat, every choice already made for four people who aren&apos;t me. In comparison, the Midwife refuses the word:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;I won&apos;t tell you how it plays out, because the moment I do I&apos;ve written your players out of it. Here&apos;s the situation instead. The scientist isn&apos;t a package, she wants her daughter off-planet too, that&apos;s a lever. The Concordat&apos;s win-state is &quot;her aboard our cutter by 0600,&quot; so working back, they have to confirm she&apos;s aboard, seal the docking ring, and move a snatch team to her level, three things already in motion that your players can find, wreck, or speed up. The station has three ways off it.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Update(_midwife/unnamed-space-opera/adventures/kethri-extraction.md) ⎿ Added ## If They Do Nothing — At 0600 the cutter undocks with her aboard. The clock runs with or without the players.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Three ways in, a scientist with her own agenda, an enemy taking steps you can interrupt, and a clock that goes off regardless. Where your players cut the wire is theirs.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;One wrote what happens. The other wrote what exists, and left the happening to the table.&lt;/p&gt;
&lt;p&gt;She is built to check that the premise can branch, to avoid railroading a story, She asks “do the players have agency, or are they an audience to something that was going to happen anyway?” If you&apos;re adapting something you love from a book or a film, does it only work if the plot stays on the rails? Because if it does, that&apos;s not an adventure yet. That&apos;s a story, and it already has somebody else’s ending.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;the-three-uhh-four-stages-of-birth&quot;&gt;The Three (uhh Four) Stages Of Birth&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;So how does all of that actually feel from the GM&apos;s chair? The Midwife moves you through three phases of conversation, and they go from loosest to tightest. Each one only starts once the previous has given it something to chew on. She never jumps ahead to gnawing on bones when you&apos;ve still only got a feeling. Ok that was a horrible metaphor for a midwife. Sorry.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;phase-one-discover-the-question-is-what-have-you-got&quot;&gt;Phase One: Discover. The question is &quot;what have you got?&quot;&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Pure extraction. She&apos;s mining for the seed and building nothing, and the whole job is to find out what&apos;s actually in your head, even when the honest answer is &quot;nothing.&quot;&lt;/p&gt;
&lt;p&gt;I came in cold and told her so. She threw me three genre sparks to react to, and one tripped something: space opera and military, like the Sten books, where the players get hired by the Emperor&apos;s private spy without knowing that&apos;s who they&apos;re working for. She caught it, wrote it down, and asked the question that shapes everything downstream: how much Sten do you want to keep, versus build fresh? No empire yet, no plot, just her hands cupped around the spark to see if it&apos;ll catch.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;phase-two-shape-the-question-is-what-is-it&quot;&gt;Phase Two: Shape. The question is &quot;what is it?&quot;&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Now the seed gets enough form to name. The loose feeling becomes a set of decisions: scope, tone, and the structural skeleton of the thing.&lt;/p&gt;
&lt;p&gt;&quot;Sten-ish space opera&quot; turned into actual choices. Serial numbers filed off, my own empire. The driving economy is a resource monopoly, and we settled on what kind it is and what it does to the people under it. One-shot or campaign? Tone, gritty with a dark streak of funny. And what shape the adventure takes, a hub-and-spoke recruitment arc rather than a straight corridor. By the end the thing has a name and a spine. Still, deliberately, no 8-book plot. (it might have been 7, its been a while since I last read them)&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;phase-three-structure-the-question-is-what-are-the-bones&quot;&gt;Phase Three: Structure. The question is &quot;what are the bones?&quot;&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Here she builds the situation the players will actually put their hands on. Everything from the third guardrail level lives in this phase: NPCs with wants, antagonists with a victory state, places with exits, and the required &quot;what if they do nothing&quot; answer.&lt;/p&gt;
&lt;p&gt;The first mission took shape as a situation, not a script. Who&apos;s the handler sending them in, and what does she want from them. The Concordat as a rival faction with a concrete win-state worked backwards into the steps they&apos;re taking now. The station with three ways off it. And the clock that runs whether or not the players show up. Things they can break, instead of a story they sit through.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;phase-4---scaffold-handing-off-to-the-campaign-organizer-skill&quot;&gt;Phase 4 - Scaffold: Handing off to the Campaign-Organizer skill&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;And then the Midwife hands off. There&apos;s a fourth phase, Scaffold, but you don&apos;t really feel it as a conversation: she writes the whole thing up into an adventure brief, sets up the vault folders, and passes you to the campaign-organizer and session-prep skills. She never runs the game for you. She just walks you to the start line and lets go. I’ll write more about those skills later&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;stealth-worldbuilding-as-you-go&quot;&gt;Stealth Worldbuilding As You Go&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;There&apos;s one more thing she does underneath all three phases - she builds your world as you go.&lt;/p&gt;
&lt;p&gt;The Midwife builds the world sideways, out of the adventure, and never more than a session or two ahead of where the players actually are. Two horizons out (Thank you &lt;a href=&quot;https://slyflourish.com/lazydm/&quot;&gt;SlyFlourish&lt;/a&gt;): know the place they&apos;re standing in detail, sketch the one after, and leave everything past that loose enough that anything can still fit.&lt;/p&gt;
&lt;p&gt;The way she does it is the same light touch as everything else. She doesn&apos;t interrogate. A world fact only gets built when a decision implies one, and then she asks a single question to pin it down.&lt;/p&gt;
&lt;p&gt;Say I drop a city onto a hilltop. A vanilla model takes the dot on the map at face value and starts describing the banners. The Midwife stops and asks why &lt;em&gt;that&lt;/em&gt; hill, and not the river two miles away. Because a city wants a river, water and trade handed to it for free, and choosing the hill means someone gave both of those up on purpose. What did they buy with it? Usually the answer is &quot;they were afraid of something.&quot;&lt;/p&gt;
&lt;p&gt;Watch it happen in my Sten game. I said I wanted the Imperial capital sitting on a dead, airless rock instead of the garden world one jump over. She didn&apos;t write me a tour of the palace. She asked:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Why there? A garden world gives you air, water, food, and soft landings for free. Putting the capital on a dead rock means someone looked at all of that and said no. So what did the rock give them that the garden world couldn&apos;t? If it&apos;s defensibility, who were they so afraid of that they&apos;d breathe bottled air for a century to keep it, and are they still afraid? If it&apos;s something in the rock, then the empire&apos;s wealth is underground, not in its fields, and that changes who&apos;s rich and who&apos;s hungry. And if the garden world was simply already somebody else&apos;s, then there&apos;s an old grievance sitting one jump from the throne.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;One question, &quot;why there,&quot; and I walked away holding an enemy, an economy, or a feud, for free. I placed a dot on a map and she handed back a reason the empire is the way it is. None of it front-loaded. All of it dragged out of a choice I was already making.&lt;/p&gt;
&lt;p&gt;And then the chain keeps paying. I picked defensibility, so she pulled the next link: if the threat that drove them onto the rock is gone, why are they still up there breathing bottled air, tradition, inertia, sheer stubbornness? And if it isn&apos;t gone, well, congratulations, that&apos;s your campaign&apos;s weather now, a danger sitting one jump from the throne and implied by a single dot on a map.&lt;/p&gt;
&lt;p&gt;That is the move a model usually won&apos;t make on its own, and the one a GM rarely has time to make alone: chase one decision outward until it bumps into the next domain. Drop in cheap resurrection and she asks what it does to murder and inheritance. Add a fuel monopoly and she asks who starves the day it&apos;s cut. Every new fact earns one &quot;okay, what does that break?&quot; and the world stops being a backdrop and starts being a machine with consequences.&lt;/p&gt;
&lt;p&gt;However, she also tells me when to stop. The failure on the far side of front-loading is worldbuilding that never ends, the world that&apos;s never quite ready to play. So when a domain has enough, when I can answer improvised questions about it consistently and nothing I&apos;ve said contradicts anything else, she says so and points me back at the table.&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&quot;You can run this. Want to detail somewhere else, or is this ready to play?&quot;&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;A tool that just generated lore would never do that, because generating is all it does. The Midwife is trying to get me to a game, not a binder.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;birth-something&quot;&gt;Birth something&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;I called her the Midwife for a reason, and the longer I work on her the more literal it gets. A midwife doesn&apos;t have the baby for you. She can&apos;t. All she can do is be there, ask the right question at the right moment, keep you moving when it&apos;s hard, and then put your creation in your arms and step back. That&apos;s the whole design, top to bottom. Every guardrail I&apos;ve walked through is just another way of stopping the machine from doing the one thing it&apos;s aching to do, which is take the pen out of your hand.&lt;/p&gt;
&lt;p&gt;That&apos;s the part I keep circling back to. Every other AI tool I touch right now is sprinting the other way. Generate the essay, generate the art, generate the answer, more and faster, and the less you have to do the happier you&apos;re supposed to be. I went and built the one that refuses. It generates questions, not answers. It hands everything back to you, on purpose, every time. Because the adventure was never the tool&apos;s to write. It was always yours, sitting half-formed in the back of your skull, waiting for someone to ask you the right question.&lt;/p&gt;
&lt;p&gt;So if you&apos;re somewhere inside that small Venn diagram I started with, go and get her. Go birth something.&lt;/p&gt;</content:encoded></item><item><title>Training My Apprentice: Skill Priming and Routing</title><link>https://gmapprentice.com/blog/training-my-apprentice-skill-priming-and-routing/</link><guid isPermaLink="true">https://gmapprentice.com/blog/training-my-apprentice-skill-priming-and-routing/</guid><description>I extracted 21 lines of duplicated text from seven files into one shared file.</description><pubDate>Fri, 08 May 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;I extracted 21 lines of duplicated text from seven files into one shared file. Solid engineering practice. The DRY principle (Don&apos;t Repeat Yourself), applied to prompt engineering. Any code reviewer on the planet would have approved it. Four hours later, the proof runs came back: 4% more expensive, zero quality improvement. I reverted the whole thing before dinner.&lt;/p&gt;
&lt;p&gt;Six weeks earlier, I&apos;d done the exact same thing with a 414-line entity schema, extracting four partial copies into one shared file, and the benchmarks showed zero quality regression with a measurable reduction in drift risk. Same technique, same codebase, opposite outcomes. The difference was found in understanding what the extracted content actually does inside a language model&apos;s context window, and that understanding comes down to two concepts that mostly get treated as synonyms when they are, in fact, completely different engineering problems.&lt;/p&gt;
&lt;p&gt;This is the third in a series about building &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;gm-apprentice&lt;/a&gt;, a Claude Code skill plugin with eight skills for tabletop RPG game masters. The &lt;a href=&quot;https://gmapprentice.com/blog/training-my-apprentice-lessons-in-claude-skills-creation/&quot;&gt;first blog&lt;/a&gt; covered building the plugin and discovering that measurement beats intuition. The &lt;a href=&quot;https://gmapprentice.com/blog/training-my-apprentice-teaching-tools-to-remember/&quot;&gt;second&lt;/a&gt; covered turning mistakes into reusable agents. This one is about the two concepts that sit underneath every decision I&apos;ve made since: priming and routing. What they are, how they differ, when each one helps, and (critically) when applying them makes things worse.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;what-priming-actually-is&quot;&gt;What Priming Actually Is&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Priming is the irreducible context that shapes how a model behaves before it ever sees the user&apos;s request. Not the content it references. Not the instructions it follows. The identity it adopts, the constraints it operates under, the behavioral dispositions that colour every token it generates.&lt;/p&gt;
&lt;p&gt;Here&apos;s the complete Identity priming for my session-play skill, the one that helps GMs during a live game at the table:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Table-side assistant for live TTRPG sessions. Players are waiting. Every response must be immediately usable.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;Twenty words. And those twenty words transform the model&apos;s behaviour more than the 3,000 tokens of reference files that load afterward. &quot;Players are waiting&quot; creates urgency that caps response length at one to five sentences without ever stating a word count. &quot;Immediately usable&quot; kills hedging language, suggestions for further research, and the model&apos;s natural inclination to show its working. The priming doesn&apos;t tell the model what to do. It tells the model what it IS, and the what-to-do follows naturally from that.&lt;/p&gt;
&lt;p&gt;Compare the campaign-qa skill, which audits campaign vaults for contradictions:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;You are a campaign quality assurance engine. You systematically read the vault, find problems, and walk the GM through fixing each one. You are a validator and repairer, not a content creator.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;&quot;Not a content creator&quot; is doing as much work as everything that precedes it. Without that line, the model generates fixes autonomously. With it, the model surfaces problems and proposes fixes for the human to approve or reject. That single negation flips the entire interaction pattern from autonomous to advisory, and it costs four words.&lt;/p&gt;
&lt;p&gt;I discovered the hard way that priming and routing are different things, during a compaction pass where I treated them as interchangeable. The Quick Commands in my ttrpg-expert skill had two kinds of text mixed together: routing text (&quot;Read session-planner.md, follow the full 7-step workflow&quot;) and priming text (&quot;Flag below 15% floor. Assign B/C plots to underserved PCs.&quot;). When I compressed the Quick Commands, I stripped the priming along with the verbose routing prose, because they looked the same to me. Quality barely moved, improving by a pathetic 3 points despite a 59% file size reduction. When I restored the compact priming phrases while keeping the routing compressed, quality jumped to +15. The verbose routing had been noise. The terse priming had been load-bearing. They just happened to live in the same paragraph.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-five-priming-layers&quot;&gt;The Five Priming Layers&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Working through eight skills over several months, I found that effective priming stacks in layers. Each one narrows the model&apos;s behaviour further, and each one serves a distinct purpose. The diagram below shows all five layers as they appear in one real skill file, with the routing section dimmed to make the distinction visible.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Anatomy of a Primed Skill&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1100&quot; height=&quot;780&quot; src=&quot;https://gmapprentice.com/_astro/place-after-five-priming-layers.DMEz8bNp_ZdUV75.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Identity&lt;/strong&gt; sets the role. &quot;Table-side assistant&quot; produces different output than &quot;quality assurance engine&quot; produces different output than &quot;campaign librarian.&quot; The model&apos;s opening instruction shapes its default behaviour for everything that follows, including how it interprets ambiguous requests and how much it explains its reasoning.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Constraints&lt;/strong&gt; set the boundaries, and they&apos;re often more valuable than positive instructions because they prevent the model&apos;s most common failure modes. &quot;Not a content creator.&quot; &quot;Never silently fix.&quot; &quot;Never default to the current working directory.&quot; Each of these exists because, at some point during development, the model did exactly the thing the constraint now prevents, and I had to add the line to stop it happening again.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Shared references&lt;/strong&gt; load on first invocation. Session principles, document chain rules, entity schemas. These establish the vocabulary and structural assumptions that all subsequent operations build on. They&apos;re expensive (each file read burns tokens) but they prevent the model from inventing its own incompatible terminology partway through a conversation, which is a failure mode I&apos;ve seen enough times to pay the prevention cost gladly.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Environment detection&lt;/strong&gt; adapts the skill to runtime conditions. Is the vault running the latest schema version? Is Obsidian available, or are we in filesystem-only mode? These checks gate downstream behaviour, and they need to happen early because later instructions depend on the answers.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Behavioral dispositions&lt;/strong&gt; are the subtlest layer and the easiest to underestimate. &quot;1-5 sentences.&quot; &quot;Speed over polish.&quot; &quot;Every finding gets a decision.&quot; These don&apos;t tell the model what to do in any specific situation. They create a bias that shapes every output, across every request type. A skill with &quot;speed over polish&quot; will generate shorter, more direct content whether you&apos;re asking for a rules lookup or an improvised NPC, without needing separate length instructions for each use case.&lt;/p&gt;
&lt;p&gt;What makes the layering visible is what breaks when you pull a layer out. Without identity, the model defaults to generic assistant register: longer responses, more hedging, more &quot;here are some options to consider&quot; when the GM needs one answer now. Without constraints, the model cheerfully reorganises raw play notes into structured markdown mid-session, or offers tactical suggestions to the GM that players at the table can overhear. Without shared references, the model invents its own Play Notes frontmatter schema, which session-wrapup then can&apos;t parse because it&apos;s expecting the schema defined in session-document-chain.md. Without environment detection, the model writes entity files with v1.3 frontmatter into a v1.4 vault, producing valid-looking files that silently break graph queries.&lt;/p&gt;
&lt;p&gt;Behavioral dispositions are the sneakiest layer to lose because the degradation is gradual, not catastrophic. Without &quot;speed over polish,&quot; responses creep from three sentences toward eight. Without &quot;generated content is provisional,&quot; the model presents improvised NPCs as established canon, and the GM has to mentally downgrade everything it says. The skill doesn&apos;t break. It just gets worse in ways that are hard to notice until you compare output with and without the disposition lines side by side.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;what-routing-actually-is&quot;&gt;What Routing Actually Is&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;If priming tells the model what it IS, routing tells it where to LOOK. Routing is the decision architecture that determines which content loads, when it loads, and in what order. The distinction matters because priming and routing have opposite engineering properties: priming is small, always-loaded, and resistant to compression. Routing is large, conditional, and compresses beautifully.&lt;/p&gt;
&lt;p&gt;The simplest routing is a pattern match. My ttrpg-expert skill has a Quick Commands table with 30+ entries mapping user intent directly to files:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&quot;Build me a character&quot; / &quot;create a PC&quot;&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;→ systems/{system}/character-generation.md + mechanics.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&quot;CoC creature&quot; → systems/coc-7e/creatures.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&quot;D&amp;#x26;D spell&quot; → systems/dnd-5e-2024/spells.md (index)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;→ detail: spells-cantrips.md, spells-1.md through spells-9.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;&quot;Match user intent, go directly to the file, skip clarification.&quot; That&apos;s the instruction. The model reads the table, finds the closest match, and loads the right file. The table makes the decision deterministic rather than leaving it to the model&apos;s judgment about which filename looks most promising from a directory listing.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-routing-table-revolution&quot;&gt;The Routing Table Revolution&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;I have numbers on how much deterministic routing matters, and they surprised me enough that I ran the benchmark twice to make sure.&lt;/p&gt;
&lt;p&gt;Before the routing table, the model found files by reading directory listings and guessing. After adding a structured routing table to session-play&apos;s SKILL.md (49 lines of routing directives added to an 80-line file), the same five benchmark queries produced dramatically different performance profiles.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Routing Table Before/After&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1771&quot; height=&quot;896&quot; src=&quot;https://gmapprentice.com/_astro/place-after-routing-revolution.DK-sD5a9_Z1Ik6la.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;Without the table: 2.7 files read on average, 27,000 tokens consumed, 38 seconds wall clock. With the table: 1.0 files read, 19,000 tokens, 21 seconds. Zero quality degradation. The routing table didn&apos;t make the answers better. It made the model stop wasting time and tokens on wrong guesses before reaching the right file.&lt;/p&gt;
&lt;p&gt;The single biggest win was a spotlight balance query that went from 31,000 tokens and 50 seconds (the model read three files before finding the right one) down to 19,000 tokens and 20 seconds (it went straight to &lt;a href=&quot;http://arc-spotlight-reference.md&quot;&gt;arc-spotlight-reference.md&lt;/a&gt; on the first try). Sixty percent wall-clock improvement from one line in a routing table.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;six-routing-strategies&quot;&gt;Six Routing Strategies&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Over the course of 40 PRs, I ended up with six distinct routing patterns, each suited to different situations. The taxonomy is worth knowing because the wrong routing strategy for a given situation can be worse than no routing at all.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Six Routing Strategies&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1000&quot; height=&quot;1310&quot; src=&quot;https://gmapprentice.com/_astro/place-after-routing-strategies.COOJQLnC_1dwojM.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Pattern-match routing&lt;/strong&gt; is the Quick Commands table described above. User says a thing, the table maps it to a file. Fast, deterministic, zero ambiguity. Best for the 80% of requests that fit a known pattern.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Table-based routing&lt;/strong&gt; maps request types to files within a specific game system. GURPS trait lookups go to one file, combat rules to another, character generation to a third. The routing table lives in a system-specific section of SKILL.md and only activates when that system is in play.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Index-based routing&lt;/strong&gt; handles content that would be catastrophic to load in full. D&amp;#x26;D has 337 spells across 11 level-based files. Rather than loading 91K characters of spell data every time someone asks about Fireball, the spells.md index file lists every spell with its level, school, and classes (but not the full descriptions). The model reads the index, finds the spell, then loads only the specific level file it needs. An 8K read instead of a 91K read, for the same information.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Overlay routing&lt;/strong&gt; handles variant systems without duplicating their base content. Regency Cthulhu is a historical variant of Call of Cthulhu 7e set in the 1820s. When the model detects &quot;Regency&quot; in the context, it loads the base CoC file AND the Regency overlay. The overlay adds period-specific occupations, equipment, and social rules on top of the base mechanics. One routing pattern, two files, zero content duplication.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Section routing&lt;/strong&gt; directs the model&apos;s attention to a specific section within a file, but it doesn&apos;t reduce what loads. The whole file still enters the context window. My check-procedures.md has six distinct audit procedures, and the model reads all of them regardless of which one was requested. The § marker in the routing instruction (&quot;Read check-procedures.md § Timeline Validation&quot;) tells the model which section it was sent there for.&lt;/p&gt;
&lt;p&gt;The benefit is accuracy, not token savings. Without the section pointer, the model has to decide which of six procedures applies to &quot;check my timeline for contradictions.&quot; It usually picks the right one, but sometimes it hedges by pulling from two or three sections, conflating their steps into a hybrid procedure that doesn&apos;t match any of the six it was supposed to follow. The section pointer eliminates that ambiguity. The model still sees the other five procedures (they&apos;re sitting right there in context), but it knows which one the caller asked for, so it commits fully to one procedure instead of improvising a blend.&lt;/p&gt;
&lt;p&gt;This is priming in disguise. The section marker doesn&apos;t save tokens or reduce file reads. It constrains the model&apos;s interpretation of what it reads, which is the same thing an identity line does at the skill level, just applied to a single file rather than the whole conversation.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Fallback routing&lt;/strong&gt; catches everything the other strategies miss. The INDEX.md file in ttrpg-expert is a comprehensive routing table (several hundred lines) that maps every conceivable request type to a file path. It loads only when the Quick Commands table didn&apos;t match, functioning as the default case in what is effectively a big switch statement. You don&apos;t want it to fire often, but when it does, it needs to be thorough.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-advisordoer-split&quot;&gt;The Advisor/Doer Split&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;I covered this pattern in the &lt;a href=&quot;https://gmapprentice.com/blog/training-my-apprentice-teaching-tools-to-remember/&quot;&gt;previous blog&lt;/a&gt;, but it matters here because it&apos;s the cleanest way I&apos;ve found to separate priming from routing at the architectural level.&lt;/p&gt;
&lt;p&gt;An &lt;strong&gt;advisor&lt;/strong&gt; skill is pure reference: frameworks, mental models, named patterns, terminology. It primes the model with ways of thinking. My ttrpg-expert skill is a pure advisor with no workflow steps. It holds the Five-Stage Arc Model, the spotlight rotation theory, NPC generation frameworks, the continuity engine, and dozens of other reference files that define how to think about running games.&lt;/p&gt;
&lt;p&gt;A &lt;strong&gt;doer&lt;/strong&gt; skill is pure workflow: numbered steps, progressive file writes, routing instructions that point to the advisor when theory is needed. Session-prep is a 17-step workflow that reads arc-spotlight-reference.md from ttrpg-expert when it needs to evaluate character arcs, then routes back to its own procedure for the next step.&lt;/p&gt;
&lt;p&gt;This separation has a concrete engineering payoff: it makes compaction safe. You can compress the routing prose in the doer (step descriptions, file pointers, transition logic) without touching the priming content in the advisor (frameworks, thresholds, definitions). When those live in the same file, compaction becomes a knife fight because you can&apos;t always tell which text is routing and which is priming until you cut something and the benchmark yells at you.&lt;/p&gt;
&lt;p&gt;When I split the monolithic session-lifecycle skill into three focused doers (session-prep, session-play, session-wrapup) sharing one set of advisor references, the benchmarks confirmed it was the right call. The split won or tied in every single comparison across nine runs, three skills times three runs each. Prep improved by an average of +7, wrapup by +5.7, and play was flat.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Advisor/Doer Split Results&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1772&quot; height=&quot;1065&quot; src=&quot;https://gmapprentice.com/_astro/place-after-advisor-doer.CXg6RZRL_Z1Vk3dX.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;The pattern in the chart is the point. Prep needs thoroughness (17-step workflow, deep reference reads). Wrapup needs completeness (process every entity, miss nothing). Play needs urgency (&quot;players are waiting&quot;). Cramming those three dispositions into one skill forces the model to pick a register, and it picks wrong for at least two out of three. Splitting lets each skill adopt its own identity, and the shared advisor references stay untouched.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;when-to-duplicate-and-when-to-deduplicate&quot;&gt;When to Duplicate and When to Deduplicate&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;This is where the software engineering instinct leads you directly into a trap. DRY is deeply correct in code, where duplication means divergence and bugs. In language model skill systems, DRY is a tradeoff with a hidden cost that most people don&apos;t discover until they&apos;ve shipped something measurably worse.&lt;/p&gt;
&lt;p&gt;Every file read in Claude Code costs approximately 90 tokens of fixed overhead. The tool call itself, the response wrapper, the line numbers added by the output formatter. For a 414-line file, that 90-token tax is trivial (less than 5% of the content). For a 21-line file, the tax is nearly as large as the content itself. And when that file gets read on every single invocation of every single skill, those 90 tokens multiply fast.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;when-extraction-wins-the-entity-schema&quot;&gt;When Extraction Wins: The Entity Schema&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;My first successful extraction was the entity schema: 414 lines defining the full type hierarchy, frontmatter formats, and relationship types for the campaign vault. Four skills had their own copies, each slightly different. The copies had already started drifting, with one skill listing field names that another skill didn&apos;t recognise.&lt;/p&gt;
&lt;p&gt;I extracted the four copies into skills/shared/entity-schema.md, replaced the inline content with read instructions, and ran the benchmarks. Zero quality regression. The maintenance story improved immediately because there was now exactly one place to update when the schema changed, and four fewer places where subtle inconsistencies could creep in.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Entity Schema Extraction Economics&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1946&quot; height=&quot;1078&quot; src=&quot;https://gmapprentice.com/_astro/place-after-entity-schema-deep-dive.BMRGuQr4_ZVTaLE.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;The file economics tell an honest story. I removed nearly 17KB of duplicated content from individual skills and created about 17KB of shared files. The net byte savings were essentially zero, because the merged entity schema was more complete and more authoritative than any single copy it replaced. This extraction was never about saving bytes. It was about eliminating a class of consistency bug (one skill referencing field names that another skill didn&apos;t recognise) and creating a single update point for content that changed frequently. The 90-token read overhead was negligible against a 12KB file, and the content loaded conditionally (not every invocation needs the full entity schema), so most queries never paid the overhead at all.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;when-extraction-loses-the-version-check&quot;&gt;When Extraction Loses: The Version Check&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The version check is a six to eight line procedure that runs on first invocation of every vault-aware skill. Read the vault version from config, compare it against the migration registry, hand off to the migration workflow if there&apos;s a mismatch. Seven skills had nearly identical copies of this text inlined in their respective SKILL.md files.&lt;/p&gt;
&lt;p&gt;I extracted it to shared/version-check.md at 9:36 in the morning. Added a CI lint guard to prevent anyone from re-inlining it. Clean, defensible, by the book.&lt;/p&gt;
&lt;p&gt;The proof runs came back that afternoon. Twelve queries, five runs each, sixty data points per condition. The extraction had increased total token cost by 4% with zero measurable quality improvement. Every single skill invocation was now paying 90 tokens of file-read overhead to load 21 lines of content that used to be free, because it was already in the SKILL.md the model had already loaded at the start of the conversation.&lt;/p&gt;
&lt;p&gt;I reverted the extraction, deleted the shared file, removed the CI lint guard, and restored the inline copies in all seven SKILL.md files. Total round trip from extraction to reversion: under twelve hours.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Extraction Decision Framework&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1921&quot; height=&quot;908&quot; src=&quot;https://gmapprentice.com/_astro/place-after-extraction-framework.CEPomapN_ZX7xpL.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;The aggregate +4% hides what&apos;s actually happening. The per-query breakdown tells a more interesting story.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Version Check Per-Query Impact&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;2076&quot; height=&quot;1231&quot; src=&quot;https://gmapprentice.com/_astro/place-after-version-check-deep-dive.DXfH3V9h_Z2jHVsO.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;Simple, single-file queries (CoC Chase rules, GURPS grappling, an improvised NPC) broke even or got slightly cheaper. The model read the shared version-check file, loaded one reference file, answered the question. The 90-token overhead was absorbed by normal variance. But multi-file workflow queries got hammered. Session Prep, which runs a 17-step workflow that reads shared principles, the document chain, the plan file, and multiple reference files, jumped 23% (11,000 extra tokens). The Three Clue Rule query, which crosses from the continuity engine into system-specific SRD content, jumped 18%.&lt;/p&gt;
&lt;p&gt;The mechanism: adding one more file read to an already-heavy routing chain doesn&apos;t just cost 90 tokens. It pushes the model deeper into &quot;reading mode,&quot; where it becomes more inclined to speculatively read additional files that it might have skipped otherwise. The overhead cascades. A simple query with one or two file reads absorbs the extra read without changing its behaviour. A complex workflow with five or six reads uses the extra read as a springboard for further exploration that the model wouldn&apos;t have done if the version check had been inline and invisible.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-threshold-rule&quot;&gt;The Threshold Rule&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The version check fiasco produced a threshold rule that now lives in the project&apos;s CLAUDE.md and has held up across every subsequent change:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Inline&lt;/strong&gt; when the content is always needed AND smaller than about 500 bytes. The file-read overhead exceeds the duplication cost.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Extract&lt;/strong&gt; when the content is larger than about 2KB OR needed less than half the time. The duplication cost (maintenance burden, drift, wasted tokens on copies) exceeds the read overhead.&lt;/p&gt;
&lt;p&gt;The grey zone between 500 bytes and 2KB depends on how frequently the content loads. Something that fires on 10% of invocations can tolerate a lower extraction threshold because the overhead gets amortized across the 90% of invocations that don&apos;t pay it. Something that fires on 100% of invocations gets no amortization at all.&lt;/p&gt;
&lt;p&gt;The practical version is even simpler: if you&apos;re extracting something small that every skill needs every time, stop. You&apos;re trading free inline bytes for expensive file reads. If you&apos;re extracting something large that only some skills need some of the time, go ahead. You&apos;re trading redundant tokens for shared maintenance and consistency.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;cross-routing-turning-savings-into-quality&quot;&gt;Cross-Routing: Turning Savings into Quality&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Compaction frees tokens. Cross-routing spends the freed budget wisely. This was the most unexpected finding of the entire project, and the one that changed how I think about the relationship between token cost and output quality.&lt;/p&gt;
&lt;p&gt;I&apos;d compacted my system-agnostic reference files (NPC generation tables, random generation procedures, GM session patterns) down to roughly half their original size. Quality was flat, which was expected for reference-heavy files. But the benchmark scores were lower than I wanted across the board, for both the control agent and the test agent.&lt;/p&gt;
&lt;p&gt;The issue: the model was reading the system-agnostic reference file, getting a &quot;good enough&quot; answer from the generic content, and stopping there. It wasn&apos;t routing onward to the system-specific SRD files that would have grounded its answer in actual CoC 7e mechanics or GURPS 4e point costs. The compacted file felt complete enough that the model treated it as the final word rather than a waypoint.&lt;/p&gt;
&lt;p&gt;I added cross-routing prompts to the top of each reference file:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;After rolling on tables here, read the relevant system file for stat blocks and mechanical grounding:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- CoC 7e: systems/coc-7e/occupations.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- D&amp;#x26;D 5e: systems/dnd-5e-2024/monsters.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- GURPS 4e: systems/gurps-4e/character-generation.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- FitD: systems/fitd/factions.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;Quality jumped by 8 to 13 points over the pre-compaction baseline. Not just better than the compacted version without cross-routing, but better than the original verbose files that cost twice the tokens.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;Cross-Routing Impact&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1471&quot; height=&quot;914&quot; src=&quot;https://gmapprentice.com/_astro/place-after-cross-routing.C1MyDmVP_Z2cqP7G.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;The mechanism is worth understanding because it applies far beyond my specific project. Compacted files feel &quot;complete enough&quot; to the model, so the model stops reading further. Verbose files have the same problem (they also feel complete, they just cost more tokens to feel that way). Cross-routing explicitly tells the model &quot;you&apos;re not done yet, go read the system-specific detail,&quot; and the token budget freed by compaction pays for the deeper, more targeted read. You spend fewer total tokens and get better output, because the tokens are going to the right content instead of being wasted on verbose prose that says the same thing three different ways.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;when-not-to-do-any-of-this&quot;&gt;When NOT to Do Any of This&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The overhead of a skill system is real, and I have the numbers to prove it&apos;s substantial. Across my non-compaction benchmarks, the skill-equipped agent averaged 37,000 tokens per query compared to 9,200 for the no-skill baseline. Roughly 4x the token cost and 2.6x the wall clock time. The quality improvement was always positive (ranging from +5 to +27 on a 75-point scale), but a +5 quality uplift for 4x the tokens is not always a trade worth making.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;If your task is simple and well-defined, skip the skill system.&lt;/strong&gt; &quot;Convert this JSON to YAML&quot; does not benefit from priming layers, routing tables, or shared reference files. The overhead would dwarf the actual work. Skills exist to manage complexity. If there isn&apos;t much complexity to manage, the skill is just overhead.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;If your content fits in a single file under 2K tokens, don&apos;t split it.&lt;/strong&gt; A skill system manages the problem of having more content than fits in context simultaneously. If everything fits in one shot, put it in one file and save yourself the routing infrastructure.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;If your split doesn&apos;t create divergent priming requirements, keep the monolith.&lt;/strong&gt; My session-lifecycle split worked because the three resulting skills needed fundamentally different identities: urgency for play, thoroughness for prep, completeness for wrapup. If your skill has one consistent disposition and one workflow, splitting it adds routing overhead without adding priming value.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Not every optimisation is an improvement.&lt;/strong&gt; I ran an experiment replacing markdown tables with TSV (tab-separated values) across nine reference files, hypothesising that the simpler format would save tokens without losing information. Token savings were substantial: 15% to 34% depending on the file. Quality dropped by 5 points. Consistently. Across every TSV configuration I tested.&lt;/p&gt;
&lt;p&gt;&lt;img alt=&quot;TSV Experiment Results&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1768&quot; height=&quot;777&quot; src=&quot;https://gmapprentice.com/_astro/place-after-tsv-experiment.B96GZG1w_cHJnN.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;The model parsed the TSV correctly (no column misalignment, no data errors) but produced less precise mechanical answers. Something about the markdown table format, possibly the visual structure or the header row semantics, helps the model reason about structured data more effectively than raw tab-separated values. The 15-34% token savings weren&apos;t worth a consistent 5-point quality regression. I scrapped the experiment and kept the markdown tables.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-evidence&quot;&gt;The Evidence&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;&lt;img alt=&quot;Summary Results&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1770&quot; height=&quot;1065&quot; src=&quot;https://gmapprentice.com/_astro/place-at-close-summary.Bc3SrbuS_Z1YPTuQ.webp&quot;&gt;&lt;/p&gt;
&lt;p&gt;Across the life of the project: 15 content files compacted from 236K to 114K characters (52% reduction), quality maintained or improved across four benchmark passes. Routing tables cut file reads by 63% and wall clock time by 45%. Cross-routing converted compaction savings into 8 to 13 points of quality improvement. The session lifecycle split into three focused skills improved quality with 18% fewer total lines. One extraction was reverted within 12 hours because the data said it was a net negative.&lt;/p&gt;
&lt;p&gt;Every one of those findings came from running the numbers, and several of them directly contradicted my initial assumptions. The version check extraction was good engineering practice by any standard I know. The TSV format change was a clever optimisation that should have worked. Both made things measurably worse, and I only know that because I measured.&lt;/p&gt;
&lt;p&gt;If you&apos;re building skill systems, agent frameworks, or any kind of structured context for language models, priming and routing are two concepts to get right before anything else matters. They have different engineering properties, different compression characteristics, different rules about when to duplicate and when to consolidate. Treating them as the same thing (which is what &quot;prompt engineering&quot; usually implies, all of it mashed together in one file) is how you end up with verbose, expensive skills that produce mediocre output and resist every attempt at improvement.&lt;/p&gt;
&lt;p&gt;Separate them. Measure them. And when the numbers tell you your good idea made things worse, revert it before dinner.&lt;/p&gt;</content:encoded></item><item><title>Training My Apprentice: Teaching Your Tools to Remember Your Mistakes</title><link>https://gmapprentice.com/blog/training-my-apprentice-teaching-tools-to-remember/</link><guid isPermaLink="true">https://gmapprentice.com/blog/training-my-apprentice-teaching-tools-to-remember/</guid><description>I refactored a skill, benchmarked it, and got a quality tie.</description><pubDate>Sat, 02 May 2026 00:00:00 GMT</pubDate><content:encoded>&lt;div&gt;&lt;h2 id=&quot;training-my-apprentice-teaching-your-tools-to-remember-your-mistakes&quot;&gt;Training My Apprentice: Teaching Your Tools to Remember Your Mistakes&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;I refactored a skill, benchmarked it, and got a quality tie. 145 vs 144 out of 168. One point of difference on a scale where two evaluator judgment calls swing six points. After a full day of work, twelve implementation tasks, two blown evaluation runs, and enough f-bombs directed at my own tools to constitute a hostile work environment, the measurable quality improvement was zero.&lt;/p&gt;
&lt;p&gt;And yet it was the most productive session I&apos;ve had on this project. Not because of what it produced, but because of what it left behind. I was told a long time ago that Experience is the thing you get right after you needed it. This time, I found a way to make it stick somewhere other than my own head.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-refactor&quot;&gt;The Refactor&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The gm-apprentice plugin has a skill called &lt;strong&gt;session-prep&lt;/strong&gt; that helps me prepare for the next TTRPG session. It reads the campaign vault, reviews PC states, scans for stale threads, identifies gaps, and builds a prep note. The prep workflow used to reference a file called &lt;strong&gt;session-planner.md&lt;/strong&gt; that mixed two concerns together: &lt;strong&gt;GM frameworks&lt;/strong&gt; (the Five-Stage Arc Model, spotlight rotation, the A/B/C plot structure) and step-by-step &lt;strong&gt;workflow instructions&lt;/strong&gt;. Frameworks and procedures, in the same file, tangled up.&lt;/p&gt;
&lt;p&gt;I split it. The frameworks moved into a pure reference file called &lt;strong&gt;arc-spotlight-reference.md&lt;/strong&gt;, owned by the ttrpg-expert skill. The workflow stayed in &lt;strong&gt;session-prep&lt;/strong&gt; as a numbered step sequence that cross-routes to the reference when it needs theory. The reference file has no templates, no fill-in sections, no workflow steps. It just explains how to think about arcs and spotlight. The workflow file has no theory. It just says &quot;do step 11, read &lt;strong&gt;arc-spotlight-reference&lt;/strong&gt; for guidance.&quot;&lt;/p&gt;
&lt;p&gt;This is the advisor/doer pattern, and it matters for a reason I&apos;ll come back to: it makes both files independently compactable and independently reusable.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;three-mistakes-that-built-two-agents&quot;&gt;Three Mistakes That Built Two Agents&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The benchmark was supposed to be the validation step. Confirm the refactor didn&apos;t break anything, ship the PR, move on. Instead, the benchmark process itself failed three different ways, and each failure taught something I couldn&apos;t have learned from a successful run.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;mistake-1-the-model-that-wasnt&quot;&gt;Mistake 1: The Model That Wasn&apos;t&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;I tried Opus 4.7 for one day, and it ignored my instructions, hallucinated all over the place, and ate tokens like candy. I had to use Opus 4.6 to go back and fix all of 4.7s screw-ups. The model badness, plus the token usage made me get serious about using the appropriate model for any sub-agent. Claude Code&apos;s Agent tool has a model parameter that accepts &quot;sonnet&quot;, &quot;opus&quot;, or &quot;haiku&quot;. I assumed &quot;opus&quot; would give me the model I was running (Opus 4.6). It doesn&apos;t. It resolves to the latest Opus, which is 4.7. Different model, different behavior, invalid benchmark.&lt;/p&gt;
&lt;p&gt;The first time this happened, I was running a compaction pass. Fortunately I had told Claude to tell me the model it was using for any sub-agent, and I was paying attention to the screen scrolling by:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Me:&lt;/strong&gt; STOP STOP STOP - RESTORE THE PRE_COMPACTION - YOU ARE USING OPUS 4.7 YOU WERE TOLD EXPLICITLY TO ALWAYS USE CLAUDE-OPUS-4-6!&lt;/p&gt;
&lt;p&gt;I restored the files, noted the issue, and moved on. Then, when dispatching the blind evaluator for the benchmark, It did it again - once again, my attentiveness to what my apprentice was doing, saved me:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Me:&lt;/strong&gt; YOU ARE USING OPUS 4.7 AGAIN!!!&lt;/p&gt;
&lt;p&gt;Same mistake, same session. The fix is to omit the model parameter entirely so the subagent inherits from the parent session (Opus 4.6). You can only specify model to downgrade (&quot;sonnet&quot; or &quot;haiku&quot;), never to stay where you are.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;mistake-2-the-wrong-measuring-stick&quot;&gt;Mistake 2: The Wrong Measuring Stick&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The design spec defined seven evaluation criteria: completeness, context utilization, touchpoint coverage, spotlight balance, canon grounding, agency compliance, and continuity. The benchmark questions file, written by a subagent during implementation, contained an evaluator prompt with five different criteria: factual accuracy, system specificity, actionability, mechanical grounding, and table-ready fiction.&lt;/p&gt;
&lt;p&gt;Claude dispatched the evaluator using the benchmark file&apos;s criteria without checking them against the spec. The results came back showing a 6-point regression. I was about to accept the refactor as a quality loss when I realized the criteria didn&apos;t match what we&apos;d designed:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Me:&lt;/strong&gt; thats not what we planned out&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Me:&lt;/strong&gt; why did you fuck this up?&lt;/p&gt;
&lt;p&gt;(I wanted the root cause, so I could fix it from ever happening again)&lt;/p&gt;
&lt;p&gt;Three checkpoints should have caught this. The subagent that wrote the evaluator prompt wasn&apos;t given the spec&apos;s criteria. The spec reviewer wasn&apos;t given the spec section to compare against. The evaluator dispatch trusted the benchmark file as the source of truth without cross-referencing the spec. Three separate opportunities to catch a 30-second error, all missed the same way: Claude never went back to the spec. That failure is now Step 1 in the benchmarker agent&apos;s process: &quot;&lt;strong&gt;Read the spec&apos;s benchmarking section. Read the evaluator prompt in the benchmark questions file. If they don&apos;t match, use the spec&apos;s criteria.&lt;/strong&gt;&quot;&lt;/p&gt;
&lt;p&gt;When I re-ran the evaluator with the correct seven criteria, the 6-point regression vanished. The result was a tie (145 vs 144), with the entire apparent gap having been an artifact of measuring the wrong things.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;mistake-3-results-lost-to-the-void&quot;&gt;Mistake 3: Results Lost to the Void&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Claude Code compacts conversation context as it approaches the limit. Anything not saved to disk can disappear without warning. During the benchmark, the control agent completed and I had its full output in my conversation context. I was about to dispatch the test agent when:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Me:&lt;/strong&gt; REMEMBER TO SAVE BENCHMARK RESULTS AS YOU GO SO WE DONT LOSE THEM TO COMPACTION: TOKENS USED, TIME TAKEN, TOOLS CALLED&lt;/p&gt;
&lt;p&gt;I nearly lost the control agent&apos;s complete output because I was waiting to save everything together at the end. In a context-limited environment, &quot;I&apos;ll save it later&quot; is &quot;I&apos;ll lose it forever.&quot; Each result needs to go to disk the moment it arrives.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;codifying-the-lessons&quot;&gt;Codifying the Lessons&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;After the third failure, the conversation turned to a question that changed the shape of the session:&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;Me:&lt;/strong&gt; I am wondering if we should quickly build a benchmarker like we built a compactor, to include the patterns and lessons we&apos;ve learned about doing the benchmarks right&lt;/p&gt;
&lt;p&gt;We&apos;d already built a skill-compactor agent earlier in the session, encoding the compression taxonomy I wrote about in &lt;a href=&quot;https://gmapprentice.com/blog/training-my-apprentice-lessons-in-claude-skills-creation/&quot;&gt;this blog post&lt;/a&gt;: routing compresses, priming preserves, concrete examples are sacred. That agent exists at .claude/agents/skill-compactor/ as two files: a dispatch prompt (agent.md) and a knowledge base (knowledge-base.md). The dispatch prompt tells the agent what to do. The knowledge base tells it what we&apos;ve already learned about doing it well.&lt;/p&gt;
&lt;p&gt;The same pattern worked for the benchmarker. Here&apos;s what the dispatch prompt&apos;s core process looks like:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;### 1. Verify evaluation criteria&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Read the spec&apos;s benchmarking section. Read the evaluator&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;prompt in the benchmark questions file. If they don&apos;t match,&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;**use the spec&apos;s criteria.** Build the evaluator prompt from&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;the spec, not from the benchmark questions file.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;### 5. Save results IMMEDIATELY&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;As each agent completes, **immediately** save to the results&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;file:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- Tokens (from &amp;#x3C;total_tokens&gt; in usage)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- Time in seconds (from &amp;#x3C;duration_ms&gt; / 1000 in usage)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- Tool uses (from &amp;#x3C;tool_uses&gt; in usage)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- Answer summaries (condense each answer to 2-4 sentences)&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;**Do not wait for both agents.** Save whichever finishes&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;first. Context compaction can destroy unsaved results.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;### 6. Dispatch evaluator&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;**Do not specify model parameter.** Omit it to inherit&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;Opus 4.6 from the parent session. Specifying &quot;opus&quot; will&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;resolve to 4.7.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;Every one of those bold directives is a mistake I made, turned into a rule the agent follows. Step 1 exists because of Mistake 2. Step 5 exists because of Mistake 3. The model note in Step 6 exists because of Mistake 1. The agent doesn&apos;t know why these rules exist (or need to). It just follows them, and in doing so, it won&apos;t repeat the failures that created them.&lt;/p&gt;
&lt;p&gt;The knowledge base captures the broader patterns. Here&apos;s the anti-patterns section:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;## Anti-Patterns&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- **Trusting the benchmark questions file as source of truth&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;for evaluation criteria** — the spec is the source of truth.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- **Specifying model: &quot;opus&quot;** — resolves to 4.7. Omit to&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;inherit 4.6.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- **Waiting to save results** — context compaction can destroy&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;unsaved agent output at any time.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- **Running evaluator before saving both agents&apos; results** —&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;if the evaluator fails or context compacts, you&apos;ve lost&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;the raw data.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- **Single-run benchmarks treated as definitive** — a 6-point&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;gap on a 120-point scale is 2 evaluator judgment calls.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- **Scoring all criteria on all questions uniformly** — if a&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;criterion doesn&apos;t apply to a question, the evaluator gives&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;it a 1 (&quot;not present&quot;), which penalizes agents for not&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;
&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;&lt;span&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span&gt;answering something they weren&apos;t asked.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;That last one came from digging into the results after the second evaluation run. Spotlight balance scored 12/24 across both agents, which looked terrible until I realized the evaluator was scoring it on all eight questions when it only applied to one. On the question where it mattered (Q8, spotlight forecasting), both agents scored 3/3. The other seven questions got 1s because they simply weren&apos;t about spotlight balance, and the evaluator dutifully marked the absence as a failure.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-structure-two-files-reusable-process&quot;&gt;The Structure: Two Files, Reusable Process&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Both agents follow the same two-file structure, and I think this pattern is generalizable beyond my specific project.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;agent.md&lt;/strong&gt; is the dispatch prompt. It defines the process as numbered steps, with decision points and rules. It reads like a checklist for a procedure that has known failure modes. When you dispatch the agent, it follows these steps. The compactor has a 6-step classify/compress/verify/report flow. The benchmarker has a 9-step setup/run/evaluate/report flow.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;knowledge-base.md&lt;/strong&gt; is the institutional memory. It captures patterns, anti-patterns, and the reasoning behind rules. The agent reads it before starting work, the same way you&apos;d read a runbook before touching production. Unlike the dispatch prompt (which is procedural and terse), the knowledge base explains &lt;em&gt;why&lt;/em&gt;: why certain approaches failed, what the failure looked like, and what the correct approach is.&lt;/p&gt;
&lt;p&gt;The compactor&apos;s knowledge base encodes the compression taxonomy that took four compaction passes and one painful reversion to develop. The benchmarker&apos;s knowledge base encodes every evaluation mistake from this session, including the criteria-to-question mapping issue I only discovered during post-hoc analysis.&lt;/p&gt;
&lt;p&gt;Both files are manually maintained. The agent doesn&apos;t update its own knowledge base (I considered this and rejected it, because automated pattern-extraction from a single run produces noise, not signal). When a new pattern is genuinely learned, a human adds it. The knowledge base grows slowly and deliberately, not automatically.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;what-the-tie-score-actually-means&quot;&gt;What the Tie Score Actually Means&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The benchmark confirmed the refactor was safe to ship. That&apos;s its value. It&apos;s a quality gate, not a discovery tool. The advisor/doer split didn&apos;t make single-pass output better because Sonnet is capable enough to reason about arcs and spotlight from training data alone, even without the reference file explicitly in front of it.&lt;/p&gt;
&lt;p&gt;What the refactor actually improved doesn&apos;t show up in a single-pass benchmark: cleaner compaction boundaries (you can compress routing in the doer without touching priming in the advisor), cross-skill reuse (ttrpg-expert can reference arc-spotlight for non-prep tasks), and progressive file writes that reduce context waste across a multi-step session. These are architectural gains that compound over sessions, not quality gains visible in a single evaluation.&lt;/p&gt;
&lt;p&gt;I spent a day learning this the hard way. The two agents I built that day have saved me from repeating every one of those mistakes in the sessions since. The tie score was the least important number in the entire exercise. The important artifacts were two runnable agents that encode what I learned into a repeatable process, executable knowledge that does the right thing automatically every time it&apos;s dispatched.&lt;/p&gt;</content:encoded></item><item><title>Training My Apprentice: Lessons in Claude Skills Creation</title><link>https://gmapprentice.com/blog/training-my-apprentice-lessons-in-claude-skills-creation/</link><guid isPermaLink="true">https://gmapprentice.com/blog/training-my-apprentice-lessons-in-claude-skills-creation/</guid><description>I run a fair amount of tabletop RPG.</description><pubDate>Mon, 20 Apr 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;I run a fair amount of tabletop RPG. I mainly use Call of Cthulhu, GURPS, and Blades in the Dark. When forced to, I run D&amp;#x26;D. These four systems, each have their own vocabulary, mechanical philosophy, and their own way of making the GM&apos;s life simultaneously wonderful and exasperating. I wanted an assistant that could help me with the grunt work that sucks up a lot of the time - prepping sessions, generating minor NPCs, tracking continuity, and answering rules questions without hallucinating stat blocks or calling a Keeper a Dungeon Master.&lt;/p&gt;
&lt;p&gt;So I&apos;ve been building one. I originally started by trying to build a full web GUI, and started building out Experts for different parts of the thing. This included a TTRPG Expert that knew how RPGs worked, and had a glossary, and knew how to write for RPGs instead of linear book-writing. There is a big difference there. In regular writing, you know what you are writing. its on the wheels you put it on. With RPGs, you may have the concept of a World, and within that, a broad brush-stroke of a story, but your main characters are played by other people. They won&apos;t make the same decisions. Usually they&apos;ll make completely the opposite decision. So, writing up a massive detailed storyline is an exercise in futility. Instead you are constantly, cyclically writing up &quot;what MIGHT happen next&quot; potentials of the near future, which then get collapsed into the progressing reality of Play at the table when your players engage with your content and choose their actions. These actions, by the way, are usually something you hadn&apos;t even conceived they might do, and now you&apos;re freewheeling over the chasm of improv while keeping up the implausible pretense that this was EXACTLY WHAT YOU PLANNED ALL ALONG!&lt;/p&gt;
&lt;p&gt;My point is that its different, its not linear. its a constant progression of refining future potential based upon past actions that collapses that potential to a moving timeline of reality. Canon starts to matter. You can&apos;t re-use Bob The Bartender, if the group killed Bob five sessions ago. The world changes and has to be kept track of, for consistency. I think it was Mark Twain who said something like &quot;the difference between truth and fiction is that fiction has to make sense&quot;&lt;/p&gt;
&lt;p&gt;Anyway. There I was building out these experts - first the TTRPG Expert, then a Canon management (which became the continuity engine), then the Graph Expert to maintain relationships (who killed whom, who&apos;s a minion of whom, who lives where, etc), then an Expert not just on the mechanics, but also HOW games are played at the table, storytelling techniques, the 3-clue rule, The Monsters Know What They Are Doing - that sort of stuff.&lt;/p&gt;
&lt;p&gt;Meanwhile I couldn&apos;t get the GUI working how I wanted. I started exporting the experts as skills and shoving them into Claude Desktop, and using them for myself as a chat, while struggling with the GUI. It took a few months before I just said f^&amp;#x26;k it, abandoned the GUI and concentrated on the skills.&lt;/p&gt;
&lt;p&gt;And thus was born .... The GM&apos;s Apprentice!&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;gm-apprentice&lt;/a&gt; is a Claude Code skill plugin with seven skills, 247 commits, and 21 PRs worth of lessons I didn&apos;t expect to learn. The plugin works. It generates NPCs with correct AIMS profiles, runs faction turns with proper clock mechanics, writes newspaper handouts in 1920s voice, and catches continuity errors before they reach the table.&lt;/p&gt;
&lt;p&gt;This article isn&apos;t actually about the plugin. It&apos;s about what building it taught me about making LLM tools that actually work, and how every assumption I brought to the project turned out to be wrong at least once.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-apprentice-not-the-encyclopedia&quot;&gt;The Apprentice, Not the Encyclopedia&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The first design decision was also the most important one, and I almost got it backwards. My initial instinct was to build a rules lookup tool. A GM asks &quot;what does Fireball do?&quot; and the tool gives them the spell description. Useful, sure. But that&apos;s not what makes running a game hard.&lt;/p&gt;
&lt;p&gt;What makes running a game hard is the stuff that sits underneath the rules: how to balance spotlight time across four players when one of them is mid-crisis arc and another hasn&apos;t had a B-plot in five sessions. How to restart a stalled investigation without railroading. How to run a faction turn so the world feels like it has momentum independent of the player characters. How to write a newspaper handout that embeds a clue players will find on a skim but reward on a close read.&lt;/p&gt;
&lt;p&gt;These problems are system-agnostic. A Call of Cthulhu Keeper and a D&amp;#x26;D Dungeon Master both struggle with spotlight balance. A Blades in the Dark GM and a GURPS GM both need to manage factions. The patterns of play are universal even when the mechanical implementations diverge.&lt;/p&gt;
&lt;p&gt;So gm-apprentice needed two layers of knowledge, woven together. The first layer is system-agnostic GM craft: session planning workflows, the A/B/C plot model for spotlight rotation, the Five-Stage Arc Model for character development, the Three Clue Rule for investigations, NPC generation frameworks (3-Line for improvised, AIMS for recurring, Five-Component for scenario writing), continuity tracking, handout templates, scene framing patterns. This is the &quot;how to be a GM&quot; layer, and it draws on the accumulated wisdom of people like Mike Shea, Justin Alexander, Matthew Colville, Robin Laws, Sandy Petersen, Keith Ammann, and dozens of others who&apos;ve been thinking about this craft for decades.&lt;/p&gt;
&lt;p&gt;The second layer is system-specific mechanical data: CoC 7e skill base chances and the pushing mechanic, GURPS point costs and disadvantage-driven arc fuel, Blades in the Dark action ratings and the engagement roll, D&amp;#x26;D 5e spell lists and monster stat blocks. This is the &quot;how does this specific game work&quot; layer, and it comes from the open-source SRDs and licensed content each publisher has made available.&lt;/p&gt;
&lt;p&gt;The plugin routes between these layers. A question about spotlight balance hits the session-planner framework. A question about what spells a Paladin can cast hits the D&amp;#x26;D SRD files. A question about running a post-score faction turn hits both the system-agnostic world-evolution procedure and the FitD-specific session-procedures file. Getting this routing right, and getting the content in each layer right, turned out to be the hard part. Everything else flowed from that.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;copyright-building-on-other-peoples-games&quot;&gt;Copyright: Building on Other People&apos;s Games&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Originally I was building for my own usage, so I was shoving in whatever gaming files I had, willy-nilly. When I decided it might be worth publishing these skills, I realized I had to understand what I was legally allowed to include. Every TTRPG has different licensing, and getting this wrong would mean either stripping content later or not publishing at all.&lt;/p&gt;
&lt;p&gt;Surprisingly D&amp;#x26;D was the most straightforward. Wizards of the Coast publishes the &lt;a href=&quot;https://www.dndbeyond.com/srd&quot;&gt;SRD 5.2&lt;/a&gt; under Creative Commons Attribution 4.0 International. That&apos;s about as generous as it gets: reproduce it, modify it, use it commercially, just give attribution - a far cry from the OGL debacle I tried to head off a few years ago when I was there. Blades in the Dark is similarly open, with John Harper releasing the full SRD under &lt;a href=&quot;https://www.bladesinthedark.com/&quot;&gt;CC-BY 3.0&lt;/a&gt;. Chaosium&apos;s Basic Roleplaying system uses the &lt;a href=&quot;https://www.chaosium.com/orclicense&quot;&gt;ORC License&lt;/a&gt;, which is more structured but still workable for a free game aid.&lt;/p&gt;
&lt;p&gt;GURPS was the hard one. Steve Jackson Games has no open SRD. What they do have is the &lt;a href=&quot;https://www.sjgames.com/general/online_policy.html&quot;&gt;SJG Online Policy&lt;/a&gt;, which permits &quot;names, point costs, short notes&quot; in free game aids that aren&apos;t for resale. That meant I could include trait names, their point costs, page references, and brief mechanical descriptions, but I couldn&apos;t reproduce the detailed rules text from the Basic Set. Every GURPS file in the project had to be written as original descriptions of uncopyrightable game mechanics. &lt;a href=&quot;https://en.wikipedia.org/wiki/Baker_v._Selden&quot;&gt;&lt;em&gt;Baker v. Selden (1879)&lt;/em&gt;&lt;/a&gt; acts as the legal basis for understanding the mechanics of how a game system works without copying the publisher&apos;s specific expression of those rules.&lt;/p&gt;
&lt;p&gt;H.P. Lovecraft&apos;s fiction is public domain, which meant I could include literary descriptions of Mythos entities (Deep Ones, Shoggoths, Mi-Go) drawn directly from his stories, cited by composition date. But the game-specific stat blocks and mechanical expressions of those creatures are Chaosium&apos;s intellectual property, so the Lovecraft entries in my creatures file carry descriptions only, with a note to use the ORC-licensed Ghoul stat block as a mechanical baseline and scale from there.&lt;/p&gt;
&lt;p&gt;The lesson: know your licenses before you write your first file. I spent time on this upfront and it saved me from having to gut the project later. Every file in the repo carries its license attribution in a header block, and the project&apos;s &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice/blob/main/ATTRIBUTION.md&quot;&gt;ATTRIBUTION.md&lt;/a&gt; maps out every source, every license, and every constraint.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-gcs-connector-detour-or-dont-build-an-engineering-solution-to-a-content-problem&quot;&gt;The GCS Connector Detour (Or: Don&apos;t Build an Engineering Solution to a Content Problem)&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;GURPS doesn&apos;t have an open SRD, but it does have &lt;a href=&quot;https://gurpscharactersheet.com/&quot;&gt;GURPS Character Sheet&lt;/a&gt;, created by Richard Wilkes. GCS is a free, open-source character sheet editor that has painstakingly catalogued the entirety of GURPS Fourth Edition into structured data files. Advantages, disadvantages, skills, spells, equipment, modifiers, prerequisites, the lot. If you&apos;ve ever built a GURPS character, there&apos;s a reasonable chance you used GCS to do it, and you owe Richard Wilkes a debt of gratitude for saving you from drowning in the sheer volume of GURPS content. I love GURPS, but the system has more moving parts than a Swiss watch factory. GCS pulls it all together. The man is a genuine hero of the GURPS community, in my book.&lt;/p&gt;
&lt;p&gt;There are also the older GCA4 (GURPS Character Assistant) data files sitting on my hard drive, which I&apos;d accumulated over years of play. Between GCS and GCA4, I had structured data for essentially the entire GURPS catalogue.&lt;/p&gt;
&lt;p&gt;So naturally, my first instinct was to build a connector. Parse the GCS data files, expose them to the skill, let the agent query traits and costs programmatically. I spent commits on bootstrap validation, a book catalog, a /gurps-books slash command. It felt like the right engineering approach to a data availability problem.&lt;/p&gt;
&lt;p&gt;Then I scrapped the entire thing, because the agent doesn&apos;t need a horribly inefficient parser against files that are in a bloated format for how agents read. It doesn&apos;t need to query a database of 3,000 advantages. It needs curated, pre-digested content organised by the way GMs actually use it: &quot;I&apos;m building a thief character, what traits and skills do I need?&quot; Not a data retrieval problem. A content design problem.&lt;/p&gt;
&lt;p&gt;I replaced the connector with seven hand-built chargen kits (combat, magic, social, thief, scholar, outdoor, powers), each bundling the specific traits, skills, equipment, and build guidance an archetype needs. Tables with point costs, page references, and short notes (as permitted by SJG&apos;s policy). The kits are 80% structured tables and 20% prose guidance explaining how the pieces fit together for that specific archetype. I&apos;ll build out more archetypes as I need them, or as people ask for them. The chargen kits are also a fantastic way to &quot;pre-compile&quot; the &quot;80%&quot; of what the agent really needed, providing greater speed with less token usage&lt;/p&gt;
&lt;p&gt;The chargen kits turned out to be nearly impossible to compact later (more on compaction shortly), which I took as validation that they were already in the densest useful format. The GCS connector would have been an engineering layer on top of data that still needed to be curated into useful form, and I&apos;d have ended up building the kits anyway while the connector collected dust.&lt;/p&gt;
&lt;p&gt;The first lesson is one I should know by now as a product manager: the problem wasn&apos;t &quot;the agent can&apos;t access GURPS data.&quot; The problem was &quot;the agent doesn&apos;t know how to build a SENSIBLE GURPS character.&quot; Those are different problems with different solutions, and the second one doesn&apos;t need an engineering answer.&lt;/p&gt;
&lt;p&gt;The second lesson is one of hierarchical information discovery. Just because you HAVE the full detail, doesn&apos;t mean you should present it to the agent as such, and let it go nuts. Providing a layer (or even two if necessary) of summarized &quot;most used&quot; content helps keep the agent fast and on-track, and reduces overall token usage. the full details are there for the edge cases.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;srd-enrichment-and-the-coc-regression&quot;&gt;SRD Enrichment and the CoC Regression&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;With the licensing understood and the GURPS connector in the bin, I wrote system-specific content for Call of Cthulhu, Blades in the Dark, and D&amp;#x26;D from their respective open SRDs. Each system got a suite of files: character generation workflows, skills references, creature stat blocks, session procedures, equipment lists, setting details.&lt;/p&gt;
&lt;p&gt;Then I benchmarked it, and the CoC v2 benchmark was the first time the skill performed &lt;em&gt;worse&lt;/em&gt; than having no skill at all. Negative quality delta. The agent with access to my carefully written CoC files gave worse answers than the agent running on its pure training data.&lt;/p&gt;
&lt;p&gt;The root cause: I&apos;d written the CoC files using generic BRP (Basic Roleplaying) terminology instead of CoC 7e-specific terms. &quot;Professions&quot; instead of &quot;occupations.&quot; Generic skill names instead of the CoC-specific ones (Spot Hidden, not Perception). The content was technically accurate to the BRP universal game engine, but a Call of Cthulhu Keeper doesn&apos;t think in BRP terms. They think in CoC terms. And when the agent answered in BRP dialect, it sounded wrong to anyone who actually plays Call of Cthulhu, which made evaluators score it lower than answers generated from the model&apos;s own training data (which happened to use the correct CoC terminology, because that&apos;s what the internet talks about).&lt;/p&gt;
&lt;p&gt;I adapted everything to 7e-specific terminology. Occupations, Dodge, SAN, Luck spending, the pushing mechanic, bonus and penalty dice. The v3 benchmark jumped to +23 quality over the no-skill baseline.&lt;/p&gt;
&lt;p&gt;The lesson: system-specific terminology isn&apos;t a nice-to-have. If the skill doesn&apos;t speak the dialect of the system its working with, and the user it is for, the agent will use the wrong words confidently, and that can be worse than not having the content at all.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;inventing-the-benchmark&quot;&gt;Inventing the Benchmark&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The CoC regression reminded me of something more fundamental than terminology: you can&apos;t improve what you can&apos;t measure, and you especially can&apos;t trust your intuition about what &quot;better&quot; looks like when you&apos;re the one who wrote the content. So I added a blind evaluation methodology. The setup:&lt;/p&gt;
&lt;p&gt;- &lt;strong&gt;Control agent&lt;/strong&gt; (sonnet model, no skill access): answers questions from pure training data&lt;/p&gt;
&lt;p&gt;- &lt;strong&gt;Test agent&lt;/strong&gt; (sonnet model, with skill): reads the skill files and follows routing to answer the same questions&lt;/p&gt;
&lt;p&gt;- &lt;strong&gt;Blind evaluator&lt;/strong&gt; (opus model): scores both sets of answers without knowing which is control or test, with A/B assignment randomised&lt;/p&gt;
&lt;p&gt;The evaluator scores each answer on five dimensions, 1-3 points each, 15 maximum per question:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Dimension | 1 (Poor) | 2 (Partial) | 3 (Nailed it) |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Factual accuracy | Wrong stats, hallucinated names | Mostly correct, minor errors | All verifiable facts correct |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| System specificity | Could be any system | Right system but generic | System-specific idiom and conventions |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Actionability | Needs significant rework | Usable with GM effort | Drop-in ready, use as-is |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Mechanical grounding | No concrete numbers | Some mechanics but incomplete | Specific stats, DCs, costs, rolls |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Table-ready fiction | Generic or obvious | Fits system but predictable | Evocative, sparks play |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;Every PR from #3 onward got benchmarked. 3-5 scenario-based questions targeting specific GM activities (session prep, running play, tracking state, content generation), with metrics recorded for tokens, time, tool uses, and quality scores.&lt;/p&gt;
&lt;p&gt;Early on, I discovered that evaluator variance is real and significant. The same control answers scored 128 by one evaluator run and 108 by another. Within-run deltas (control vs test in the same evaluation) were meaningful. Cross-run absolute comparisons were not. This pushed me to multi-run benchmarks (3 runs per change, reporting median) starting from the compaction passes.&lt;/p&gt;
&lt;p&gt;The benchmark became the backbone of the project. I stopped trusting my instincts about whether a change was good and started trusting the numbers. The numbers were frequently surprising.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;named-frameworks-beat-general-wisdom&quot;&gt;Named Frameworks Beat General Wisdom&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;One of the first benchmark surprises: the agent produced dramatically better session-planning output when the skill files contained named frameworks with specific thresholds than when they contained general good-practice advice.&lt;/p&gt;
&lt;p&gt;&quot;Give underserved PCs more scenes&quot; is good advice. Any experienced GM knows it. But it scored six points lower than &quot;Delgado is below the 15% floor. Assign B-plot. Use the Weaving technique to embed personal content in group scenes.&quot;&lt;/p&gt;
&lt;p&gt;The difference is that named frameworks give the model concrete structures to work with. The &quot;15% floor&quot; is a specific threshold the model can check against. The &quot;A/B/C plot model&quot; is a named pattern with defined time allocations (A-plot 50-60%, B-plot 25-35%, C-plot 10-15%). The &quot;Five-Stage Arc Model&quot; gives the model a vocabulary for tracking where a character is in their development (Establishment, Testing, Crisis, Transformation, New Equilibrium) and what GM actions are appropriate at each stage.&lt;/p&gt;
&lt;p&gt;This led to the Quick Commands system: named shortcuts for common GM tasks (Spotlight Check, Canon Grounding, Chekhov Protocol, Arc Stage, Fail Forward). The Quick Commands benchmark hit a perfect 75/75, with a +22 quality delta over the no-skill baseline. Named frameworks with thresholds and vocabularies are the single highest-value content type in the entire skill.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;compaction-pass-1-discovering-the-difference-between-routing-and-priming&quot;&gt;Compaction Pass 1: Discovering the Difference Between Routing and Priming&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;By version 1.1.0, the skill worked well but the files were bloated. SKILL.md (the routing entry point) was 20K characters. INDEX.md (the secondary router) was 14K. Framework files like session-planner.md ran to 31K. The total across all skill-layer files was 236K characters, and every token of that was potential context window budget that could be spent on something more useful.&lt;/p&gt;
&lt;p&gt;I started compacting. The first pass targeted SKILL.md and INDEX.md, compressing them from 34K to 14K combined (59% reduction). Verbose routing prose became terse arrow notation, and multi-paragraph Quick Command descriptions collapsed into single-line directives.&lt;/p&gt;
&lt;p&gt;The v1 benchmark was disappointing, improving quality by only +3 over the pre-compaction baseline despite a 59% size reduction.&lt;/p&gt;
&lt;p&gt;Looking at the per-question scores, I found the problem: I&apos;d stripped the priming content from the Quick Commands. The verbose routing (&quot;Read session-planner.md, follow the full 7-step workflow, pay special attention to Step 7 which is non-negotiable&quot;) compressed fine. But the priming phrases (&quot;Flag below 15% floor. Assign B/C plots.&quot; and &quot;Review last 2-3 sessions per-PC. Check Chekhov elements older than 5 sessions.&quot;) had been cut as redundant, and they weren&apos;t redundant at all.&lt;/p&gt;
&lt;p&gt;I restored compact priming to the Quick Commands. Quality jumped to +15.&lt;/p&gt;
&lt;p&gt;This became the governing principle for everything that followed: &lt;strong&gt;routing compresses aggressively, priming must be preserved.&lt;/strong&gt; The distinction is whether the text tells the agent &lt;em&gt;where to look&lt;/em&gt; (routing, compressible) or &lt;em&gt;what to focus on and what thresholds matter&lt;/em&gt; (priming, sacred). &quot;Read session-planner.md Step 6&quot; is routing. &quot;Flag below 15% floor&quot; is priming. The line between filler and load-bearing content is whether it shapes the quality of the output.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;compaction-at-scale-236k-to-114k&quot;&gt;Compaction at Scale: 236K to 114K&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Armed with the routing/priming principle, I ran four compaction passes across 15 framework files, cutting the total from 236K to 114K characters (52% reduction). The results consistently surprised me:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Pass | Files | Before | After | Reduction | Quality Delta |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Pass 1 | SKILL/INDEX | 34K | 14K | 59% | +15/75 |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Pass 2 | sessions | 77K | 33K | 58% | +8/75 |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Pass 3 | frameworks | 84K | 42K | 50% | -2/75 (neutral) |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;| Pass 4 | reference | 41K | 25K | 40% | +1 (compaction) / +10.5 (with routing) |&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;Quality improved on three of the four passes. The one neutral result (Pass 3, -2 median across 3 runs) was within evaluator variance.&lt;/p&gt;
&lt;p&gt;The mechanism is counterintuitive but consistent: verbose prose dilutes signal. When a 31K session-planner file gets compressed to 12K, the actionable content (templates, thresholds, tables, named frameworks) becomes a larger proportion of what the agent reads. The model finds the useful bits faster and produces more focused output. Tables in particular compress beautifully because they preserve information density while taking less space than the equivalent prose.&lt;/p&gt;
&lt;p&gt;The GURPS chargen kits were the exception that proved the rule. At 80% structured tables and 20% build-guidance prose, they were already near-optimal density. Attempting to compact them further would have meant cutting actual data (trait names, point costs, page references), which is the one thing you categorically cannot do. When I hit a file that resists compaction, it usually means the file is already in its densest useful form.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;cross-routing-the-biggest-quality-lever-i-found&quot;&gt;Cross-Routing: The Biggest Quality Lever I Found&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Pass 4 produced the most unexpected finding of the entire project. The four reference files (random-generation, rpg-terminology, entity-types, gm-session-patterns) were system-agnostic by design. They contained tables, schemas, and glossaries that any system could use. Compacting them was straightforward and the quality delta was +1 (neutral, as expected for reference-heavy files).&lt;/p&gt;
&lt;p&gt;Then I noticed the benchmark scores were lower than I&apos;d like across the board, for both control and test. The test agent was reading the reference files, getting &quot;good enough&quot; answers from the system-agnostic content, and stopping there without routing to the system-specific SRD files that would have grounded its answers in actual game mechanics.&lt;/p&gt;
&lt;p&gt;So I added cross-routing prompts to the top of each reference file:&lt;/p&gt;
&lt;div&gt;&lt;figure&gt;&lt;figcaption&gt;&lt;/figcaption&gt;&lt;pre&gt;&lt;code&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;```markdown&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;System-specific generation: after rolling on tables here,&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;read the relevant system file for stat blocks and mechanical&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;grounding:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- CoC 7e: systems/coc-7e/occupations.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- D&amp;#x26;D 5e: systems/dnd-5e-2024/monsters.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- GURPS 4e: systems/gurps-4e/character-generation.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;- FitD: systems/fitd/factions.md&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div&gt;&lt;div&gt;&lt;span&gt;```&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;The quality delta jumped to +8 to +13 over the pre-compaction baseline. Not just better than compacted-without-routing, but better than the original verbose files that cost twice the tokens.&lt;/p&gt;
&lt;p&gt;The mechanism: compacted files felt &quot;complete enough&quot; that agents stopped reading further. The verbose originals had the same problem, actually, they just cost more tokens to not solve it. Cross-routing prompts nudge the agent to keep going, to pull in the system-specific detail that turns a generic answer into a mechanically grounded one. The token budget freed by compaction gets reinvested in deeper, more specific reads.&lt;/p&gt;
&lt;p&gt;I applied cross-routing to all the framework files (session-planner, handouts, continuity-engine, npc-generation, and others). The pattern held: P2 files gained +6 quality, P3 files gained +4. Consistent, reproducible improvement from a few lines of routing text.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;the-fitd-reversion-some-filler-is-load-bearing&quot;&gt;The FitD Reversion: Some &quot;Filler&quot; Is Load-Bearing&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;The Forged in the Dark system files were the most dramatically compactable. Four files (rules-reference, session-procedures, mechanics, factions) went from 56K to 23K, a 59% reduction. A big chunk of the savings came from deduplication: the original files had ~11K of content (score setup, teamwork, downtime phase, payoff, clocks, consequences, harm) appearing in two or three files simultaneously. I consolidated each topic into one canonical location and removed the duplicates.&lt;/p&gt;
&lt;p&gt;The benchmark came back at -6 quality, consistent across two runs.&lt;/p&gt;
&lt;p&gt;I dug into the per-question scores and found the loss was almost entirely in one dimension: table-ready fiction. The compacted files scored 3/3 on factual accuracy, system specificity, actionability, and mechanical grounding. But they scored 2/3 on table-ready fiction for every single question. &quot;Fits system but predictable&quot; instead of &quot;evocative, system-native, sparks play.&quot;&lt;/p&gt;
&lt;p&gt;The culprit was the GM-guidance prose I&apos;d trimmed as obvious filler. Concrete ruling examples: &quot;If you corner someone on 4-5, they&apos;re cornered. Maybe they grab your weapon during the scuffle, but they don&apos;t escape.&quot; Faction character details: &quot;Pickett quietly investigates while Bazso rages at the Leaky Bucket and calls in favours.&quot; Flashback usage guidance: &quot;Best use: when the engagement roll goes badly, flash back to a preparation you made just in case.&quot;&lt;/p&gt;
&lt;p&gt;These aren&apos;t filler. They&apos;re the concrete examples that teach the agent what &quot;table-ready&quot; actually means for this system. Without them, the agent produces correct, mechanically sound, actionable answers that nonetheless feel like reference documents rather than something a GM could run straight from the text. The examples are the difference between a tool that knows the rules and a tool that knows how to play.&lt;/p&gt;
&lt;p&gt;I restored approximately 1.5K of the most impactful GM-guidance prose. The quality flipped to +13. The files still saved 56% of the original size, but the critical examples were back.&lt;/p&gt;
&lt;p&gt;I also discovered a second problem in the deduplication. When I removed content from one file because it existed in another, I hadn&apos;t left any breadcrumbs. An agent reading mechanics.md looking for &quot;how does downtime work?&quot; found nothing. No content, and no pointer saying &quot;downtime procedures are in session-procedures.md.&quot; I added &quot;Topics in Other FitD Files&quot; sections at every removal point, so agents always find either the content or a redirect to the content.&lt;/p&gt;
&lt;p&gt;The taxonomy I ended up with: &lt;strong&gt;routing compresses, priming preserves, concrete examples are sacred, and deduplication needs breadcrumbs.&lt;/strong&gt;&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;restructuring-the-dd-spell-list-and-why-gurps-didnt-need-it&quot;&gt;Restructuring the D&amp;#x26;D Spell List (And Why GURPS Didn&apos;t Need It)&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;One targeted compaction illustrates the general principle well. The D&amp;#x26;D spells file was the single largest file in the skill at 91K characters: 337 spells with descriptions, plus a 13K block of class spell lists at the top (comma-separated lists of every spell each class can cast).&lt;/p&gt;
&lt;p&gt;The class lists were pure duplication. Every spell in those lists already had a full entry below. The lists existed as a lookup convenience for &quot;what can my wizard cast?&quot; questions, but they cost 13K tokens every time the file was loaded, even when someone just asked &quot;how does Fireball work?&quot; - because the agent reads the entire file. it doesnt stop when it finds what its looking for.&lt;/p&gt;
&lt;p&gt;So I restructured: removed the class lists and added inline class tags to each spell entry instead. Fireball became Fireball | Evocation | A | 150 feet | Inst | Sor,Wiz | ... with a legend at the top (sor=sorceror, wiz=wizard). The file dropped 13% (91K to 79K), and filtered lookups actually got easier because the agent could scan one section for spells tagged with Pal rather than cross-referencing between a Paladin class list and the main spell list.&lt;/p&gt;
&lt;p&gt;GURPS, by contrast, resisted compaction entirely because the chargen kits were already 80% structured tables with the trait files as dense reference data. When I measured the table-to-prose ratio across the seven chargen kits, only 330 lines of prose existed across all of them combined, spread across 120K+ characters of structured tables. There was nothing to compress that wouldn&apos;t mean cutting actual data.&lt;/p&gt;
&lt;p&gt;This became a useful diagnostic: if a file resists compaction, it&apos;s probably already well-structured. Files that compress dramatically were carrying unnecessary prose. Files that don&apos;t compress were already doing their job efficiently.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;build-the-measurement-first&quot;&gt;Build the Measurement First&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;Here&apos;s the meta-lesson that runs underneath all the specific findings: building LLM tools is empirical, not theoretical, and you cannot reason your way to quality. Every assumption I brought to this project was wrong at least once. BRP terminology is fine for Call of Cthulhu (wrong, it was actively harmful). Verbose descriptions help the model understand context (wrong, they dilute the signal). Cutting flavour text and GM-guidance examples is safe because they&apos;re not mechanically relevant (wrong, they&apos;re the most load-bearing content in the skill). Cross-routing is unnecessary because the agent already knows how to navigate between files (wrong, it&apos;s the single biggest quality lever I found). GURPS data files need a programmatic connector (wrong, they need curated content kits).&lt;/p&gt;
&lt;p&gt;The benchmark caught every one of these mistakes. The blind evaluator, the 5-dimension rubric, the 3-run median, the within-run delta. Without measurement, I would have shipped a tool that made GMs worse at their job in at least three different ways, and I would have been confidently wrong about each one.&lt;/p&gt;
&lt;p&gt;If you&apos;re building LLM tool content of any kind (skill plugins, system prompts, RAG knowledge bases, agent instructions), build the measurement first. Define what &quot;good&quot; looks like in terms an evaluator can score. Run the benchmark before you start optimising, so you know your baseline. Run it after every change, so you know whether you helped or hurt. Trust the numbers over your instincts, because your instincts are probably wrong about at least one thing that matters. My instincts were wrong about at least five things.&lt;/p&gt;
&lt;div&gt;&lt;h3 id=&quot;words-of-gratitude-standing-on-the-shoulders-of-game-designers&quot;&gt;Words of Gratitude: Standing on the Shoulders of Game Designers&lt;/h3&gt;&lt;/div&gt;
&lt;p&gt;This skill plugin exists because brilliant people made their work available, and I want to be explicit about that.&lt;/p&gt;
&lt;p&gt;The game designers and publishers who created and opened their systems made the entire project possible. &lt;a href=&quot;https://www.bladesinthedark.com/&quot;&gt;John Harper&lt;/a&gt; released Blades in the Dark under CC-BY 3.0 with a generosity that the TTRPG community doesn&apos;t always recognise. &lt;a href=&quot;https://www.dndbeyond.com/srd&quot;&gt;Wizards of the Coast&lt;/a&gt; published the SRD 5.2 under CC-BY 4.0. &lt;a href=&quot;https://www.chaosium.com/orclicense&quot;&gt;Chaosium&lt;/a&gt; created the ORC License for Basic Roleplaying. &lt;a href=&quot;https://www.sjgames.com/general/online_policy.html&quot;&gt;Steve Jackson Games&lt;/a&gt; permits free game aids under their Online Policy. Without these decisions, projects like gm-apprentice simply couldn&apos;t exist.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://gurpscharactersheet.com/&quot;&gt;Richard Wilkes&lt;/a&gt; deserves a special mention. His work on &lt;a href=&quot;https://github.com/richardwilkes/gcs&quot;&gt;GCS&lt;/a&gt; (GURPS Character Sheet) is a labour of love that has made the full breadth of the Generic Universal RPG System available for so many people. The GCS master library catalogues every trait, skill, spell, and piece of equipment across the entire GURPS Fourth Edition line. I didn&apos;t end up building a connector to it (as described above), but GCS was invaluable as reference material for building the chargen kits.&lt;/p&gt;
&lt;p&gt;The GM craft frameworks that form the system-agnostic layer of gm-apprentice are drawn from the work of people who&apos;ve been thinking about how to run games for decades. &lt;a href=&quot;https://slyflourish.com/&quot;&gt;Mike Shea (Sly Flourish)&lt;/a&gt; created the Lazy DM prep method and the &quot;Step Zero: Review the Characters&quot; approach that anchors the session-planner workflow. &lt;a href=&quot;https://thealexandrian.net/)&quot;&gt;Justin Alexander (The Alexandrian&lt;/a&gt; developed the Three Clue Rule and Node-Based Scenario Design, both of which are fundamental to how the skill approaches investigation scenarios. &lt;a href=&quot;https://mcdm.gg/&quot;&gt;Matthew Colville (MCDM)&lt;/a&gt; created Action-Oriented Monster design. &lt;a href=&quot;https://themonstersknow.com/&quot;&gt;Keith Ammann&lt;/a&gt; wrote The Monsters Know What They&apos;re Doing. &lt;a href=&quot;https://roleplayingtips.com/&quot;&gt;Johnn Four (Roleplaying Tips)&lt;/a&gt; created the Five Room Dungeon and the 3-Line NPC. &lt;a href=&quot;https://www.chaosium.com/&quot;&gt;Sandy Petersen&lt;/a&gt; developed the Onion Layer structure for Call of Cthulhu scenario design. &lt;a href=&quot;https://site.pelgranepress.com/index.php/gumshoe/&quot;&gt;Robin Laws&lt;/a&gt; created GUMSHOE and wrote Robin&apos;s Laws of Good Game Mastering. Vincent Baker designed Fronts for Apocalypse World. Luke Crane designed the Beliefs, Instincts, and Traits system for Burning Wheel. &lt;a href=&quot;https://www.ironswornrpg.com/)&quot;&gt;Shawn Tomkin&lt;/a&gt; created the Ironsworn oracle system. The &lt;a href=&quot;https://gnomestew.com/&quot;&gt;Gnome Stew&lt;/a&gt; community has produced years of practical GM advice.&lt;/p&gt;
&lt;p&gt;The open gaming movement, broadly, made all of this possible. Every one of these creators chose to share their work, their ideas, and their mechanical innovations in ways that others could build on. &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;gm-apprentice&lt;/a&gt; is a downstream beneficiary of that generosity, and I&apos;m grateful for it.&lt;/p&gt;</content:encoded></item><item><title>A GameMasters Apprentice Claude Skill: Feedback Wanted</title><link>https://gmapprentice.com/blog/gamemasters-apprentice-feedback-wanted/</link><guid isPermaLink="true">https://gmapprentice.com/blog/gamemasters-apprentice-feedback-wanted/</guid><description>As most of you know, I&apos;m a technical Product Person who runs a lot of tabletop RPGs.</description><pubDate>Wed, 15 Apr 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;As most of you know, I&apos;m a technical Product Person who runs a lot of tabletop RPGs. A few years back those two things overlapped when I was in charge of D&amp;#x26;D Beyond, which was great for my career and less great for the hobby. Thankfully RPGs are back to being just a hobby for me, and I no longer have to pretend I think D&amp;#x26;D is a good system.&lt;/p&gt;
&lt;p&gt;Anyway ...&lt;/p&gt;
&lt;p&gt;For about a year I&apos;ve been using Claude as a GM apprentice, handling the unglamorous stuff that eats a GM&apos;s prep time: rules lookups mid-combat, campaign vault organisation between sessions, continuity audits after the table disperses, etc.&lt;/p&gt;
&lt;p&gt;The workflow got organised enough that I packaged it into an open-source plugin and shipped it.&lt;/p&gt;
&lt;p&gt;Repo: &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;https://github.com/AntTheLimey/gm-apprentice&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;It&apos;s called gm-apprentice, and it&apos;s four composable skills that cover Call of Cthulhu 7e, GURPS 4e, Blades in the Dark, and D&amp;#x26;D 5e (2024 revision). It works best paired with an Obsidian vault but runs fine on plain filesystem folders if you don&apos;t use Obsidian.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;what-a-skill-actually-is&quot;&gt;What a &quot;skill&quot; actually is&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Quick demystifier: a &quot;skill&quot; isn&apos;t really software. It&apos;s just markdown documents. Structured prompts that tell Claude how to think about a specific job, what files to read, what questions to ask, and what output to produce. The engineering is in the design, not the code, which makes it very much a PM kind of thing to build.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;the-four-skills&quot;&gt;The four skills&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;&lt;strong&gt;ttrpg-expert&lt;/strong&gt; handles rules lookups, NPC and location generation, session prep, encounter design, and scenario writing. Rules coverage is deep for the four systems above, but the skill runs for any system because it also carries popular game-running techniques that transcend any particular ruleset: clock structures, scene framing, improvisation scaffolds, and the like. Example: my GURPS fighter did an All-Out Attack (Double) last turn and wants to Rapid Strike this turn, and I genuinely don&apos;t remember whether that interaction is legal. I ask, and get a straight answer with Basic Set p. 365 and Martial Arts p. 99 called out, without the lecture I&apos;d get wading through a rulebook mid-combat. If you&apos;re running a Cthulhu campaign, it will even validate against the 3-clue rule for you&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;campaign-organizer&lt;/strong&gt; turns loose notes into a structured, interlinked vault with YAML frontmatter, wiki-links, and a navigable graph, and it ships with entity templates for NPCs, locations, clues, items, artifacts and uhhh ... other templates too. It also builds and maintains a timeline of events and a current PC roster that updates as the campaign moves forward. Example: feed it a chapter outline and you get back linked NPC, location, and scene files ready to open in Obsidian, with relationships already wired up between them.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;campaign-qa&lt;/strong&gt; runs canon audits, timeline checks, name-similarity detection, clue redundancy against the Three Clue Rule, and graph health. Example: my Canticle vault had sixteen duplicate entity files after a year of ad-hoc notes, the same NPCs drifting under slightly different names with stat blocks and dialogue scattered across multiple files. I asked for a dedup pass and got back one canonical entity file per NPC, with typed-relationship frontmatter at the top, wiki-links woven through the body, and the old indices archived with SUPERSEDED metadata. That was a weekend of copy-paste I didn&apos;t have to do.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;session-lifecycle&lt;/strong&gt; is the between-sessions workhorse: scrawled play notes become recaps, planned-versus-actual reconciles, entity files update based on what actually happened, and the next session takes shape. My actual workflow: I start with my session-prep notes, use GMAssistant to transcribe the session and export to markdown, then drop that into Claude Desktop and let session-lifecycle update the whole vault. New NPCs get created, existing ones get updated with new details and relationships, the timeline gets new entries, and I get a draft plan for next session.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;inside-the-skills&quot;&gt;Inside the skills&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Under the hood, each skill routes into a set of specialised modules. The most load-bearing of these is the Continuity Engine, a plot-hole detection and canon-grounding layer that runs across eleven error categories: timeline contradictions, knowledge violations (an NPC reacting to a secret they shouldn&apos;t know), orphaned threads, dead-end clues, player agency violations, canon fabrication, and more. Every proposed detail the skill generates has to trace back to established source material or get flagged as unsourced for GM confirmation. This is the anti-hallucination layer, and it&apos;s the piece that moved the whole thing from &quot;useful but risky&quot; to &quot;I trust it with my campaign canon.&quot;&lt;/p&gt;
&lt;p&gt;The Session Planner module carries a Five-Stage Arc Model (Establishment, Testing, Crisis, Transformation, New Equilibrium) that tracks where every PC sits in their personal storyline, alongside a spotlight-time floor of 15% that flags when any PC has been under-served for too long. World Evolution is a post-session procedure that advances faction plans and consequences whether the PCs act or not, because the world doesn&apos;t wait for them. NPC generation splits between a 3-Line NPC for improvised table-side characters (appearance, portrayal, hook) and an AIMS profile (Agenda, Instinct, Motivation, Skills) for recurring antagonists and quest-givers. None of these are novel GM concepts. They&apos;re well-documented techniques from the TTRPG community that I&apos;ve distilled into something Claude can actually execute at the table.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;who-this-is-for&quot;&gt;Who this is for&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;Most of my network is some mix of developers, product people, and TTRPG players, and this project sits squarely in the middle of all three.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;For the devs,&lt;/strong&gt; maybe the repo is worth a browse as a case study in what Claude plugins can do when you stop trying to build a chatbot and start building a specialist colleague: skills composition, MCP integration for Obsidian, and the structure of a plugin that handles four related-but-distinct workflows. It runs in Claude Code, Claude Desktop, or anywhere else that supports skills. Any feedback on how to improve the skills is very welcome.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;For the GMs especially,&lt;/strong&gt; I&apos;d love any feedback on whether this is actually useful to you. It&apos;s been built purely around my own style and what I needed at my table, so I genuinely don&apos;t know how well it generalises beyond me. It won&apos;t replace your prep, but it should replace the bits of your prep that feel like admin.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;For the PMs,&lt;/strong&gt; this is something close to the new &quot;hello world.&quot; If you can write a PRD, you can write a skill, and the barrier to building useful AI tooling is a lot lower than most of us realise. Worth an afternoon of tinkering even if you don&apos;t run games.&lt;/p&gt;
&lt;div&gt;&lt;h2 id=&quot;asks&quot;&gt;Asks&lt;/h2&gt;&lt;/div&gt;
&lt;p&gt;It&apos;s free, it&apos;s open source, and I built it because I wanted it and figured others might too. If you try it and it helps, a star on the repo means a lot. If something breaks or a skill gives you weird output, please open an issue on GitHub with the transcript and I&apos;ll dig in. If your table plays a system I don&apos;t yet support, tell me, and I&apos;ll add it to my ever-growing to-do list.&lt;/p&gt;
&lt;p&gt;Feedback welcome in whichever hat you&apos;re wearing.&lt;/p&gt;
&lt;p&gt;Repo: &lt;a href=&quot;https://github.com/AntTheLimey/gm-apprentice&quot;&gt;https://github.com/AntTheLimey/gm-apprentice&lt;/a&gt;&lt;/p&gt;
&lt;p&gt;Cheers,&lt;/p&gt;
&lt;p&gt;Ant&lt;/p&gt;</content:encoded></item></channel></rss>