Yesterday I read the official skills repository — the format’s own authors showing how an instruction file should be written. Then I opened GitHub this morning and the #1 repository in the world, above every agent framework and model release, is a single skill written by one person. It draws diagrams. 27 kinds.

So I ran the audit: does an independent developer’s #1 repo actually follow the rules the format’s authors just published?

The answer is more interesting than yes or no. It breaks the most-quoted rule outright — and I’ll argue it’s right to, for a reason the guidance doesn’t cover. It out-executes the official set on the deepest rule, and I measured the exact ratio (three numbers tell the whole story). And it does something nobody else is doing: before it will draw anything for you, it stops and refuses — with a sentence explaining why that’s one of the most quietly professional things I’ve read in an instruction file this month.

Also inside: the checklist that tells the model not to draw at all, the 514-line linter enforcing the requirement everyone skips, the “fidelity ledger” on imports, and the grep result that should embarrass a few skill authors.

⏱️ Chapters:
00: 00 I Thought That Was the End of the Story
00: 27 One Person, One Folder, #1 on GitHub
00: 48 Three Findings
01: 40 Standards Note
01: 54 What It Actually Is
02: 43 One Skill, Three Different Agents
03: 05 The Audit: The Rule It Breaks
03: 23 144 Words in One Field
03: 47 Why It’s Right and the Guidance Is Incomplete
04: 12 27 Capabilities Means 27 Triggers
04: 59 The Rule It Nails — Three Numbers
05: 26 32KB Loads. 484KB On Demand. 1.5MB Never.
06: 07 One Part Loaded, Fifteen On Demand, Forty-Five Never
06: 22 A Big Skill Isn’t a Long File
06: 38 The Gate That Refuses to Draw
07: 23 The One-Line Heuristic
07: 43 It Explains Its Own Failure Mode
08: 16 Adding Friction, Once
08: 23 The Taste Gate
08: 36 “If Yes — Don’t Draw”
08: 58 The Remove Test
09: 26 One Accent, Two Elements Max
09: 50 I Ran Yesterday’s Grep Again
10: 09 Not One Raised Voice
10: 27 The Scripts and the Accessibility Surprise
11: 09 A Diagram Is a Blank Rectangle to a Screen Reader
11: 49 Behaving Like Maintained Software
11: 59 The Import Path and the Fidelity Ledger
13: 07 That’s a Temperament
13: 21 Who Pays — There Isn’t an Invoice
13: 55 Value Has Moved to Codified Expertise
14: 18 Honest Limits
15: 06 What the Two Videos Add Up To
15: 50 Teaching an Agent to Do Less
16: 10 What Should Your Agent Refuse to Do?

🔗 Sources — open the checklist section first:
– The repository (MIT): https://github.com/cathrynlavery/diagram-design
– The instruction file: https://github.com/cathrynlavery/diagram-design/blob/main/skills/diagram-design/SKILL.md
– The reference folder (37 files, loaded on demand): https://github.com/cathrynlavery/diagram-design/tree/main/skills/diagram-design/references
– The accessibility linter: https://github.com/cathrynlavery/diagram-design/blob/main/scripts/lint-skin.py
– The official spec this is measured against: https://github.com/anthropics/skills
– skill-creator (three-level loading + the “pushy” doctrine): https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md

📺 The saga, on this channel:
– Anthropic Just Published Official Claude Skills — I Read Every Single One
– The AI Skills Gold Rush: I Read All Three
– Google Published 61 Official Agent Skills — I Read Every One
– Agent Plugins: The Skills Format War Is Official
– Cloudflare Gave AI Agents a Real Computer — Inside the Repo

#ClaudeSkills #ClaudeCode #AIAgents #OpenSource #SignalCoders

👇 Subscribe — a new plain-English breakdown of AI engineering, open source tools, and local LLMs every single day.
YouTube: @SignalCoders
Twitter/X: @SignalCoders

Our Playlists:
https://www.youtube.com/playlist?list=PLL70bDS-hH7Y
https://www.youtube.com/playlist?list=PLTnqXOgLPogE
https://www.youtube.com/playlist?list=PLZVB6navai50
https://www.youtube.com/playlist?list=PLEj9VzjWmmho

⚠️ Disclaimer: This video is for educational purposes only. Signal Coders is not affiliated with, or sponsored by, the author of this repository, Anthropic, or any company mentioned. This channel READ the repository at source — the skill was NOT installed, run, or benchmarked, and output quality cannot be assessed from source alone. All measurements (538 lines, 144 description words, 37 reference files, zero screamed imperatives, file sizes) are a snapshot of the tree pulled on the recording date; the project is versioned and actively moving. The multi-agent support claim is the repository’s README statement, reported as theirs. Skills execute code — review anything you install. Nothing here is professional advice.