mirror of
https://github.com/shipshitdev/skills.git
synced 2026-09-19 06:04:15 +08:00
VincentShipsIt
2.2 KiB
2.2 KiB
resources/
First-hand reference material for working in this repo. Unlike .agents/memory/system/ (this repo's own operating standards), resources/ holds extracted upstream guidance — what Anthropic and OpenAI themselves publish — kept separate so the source of truth is unambiguous.
skill-authoring/
Official skill-authoring guidance, extracted from 13 first-hand vendor sources and read on 2026-06-12. Every rule links back to its source URL.
| File | What's in it |
|---|---|
| anthropic.md | Anthropic's rules — Claude Code / Claude API / claude.ai. Frontmatter, naming, descriptions, structure, progressive disclosure, invocation, tool permissions, references/scripts, testing, anti-patterns. |
| codex.md | OpenAI Codex's rules, with the divergences from Anthropic called out. |
| frontmatter-field-spec.md | Field-by-field reference table (base spec + Claude extensions) with the constraints that bite. |
| checklist.md | One pass/fail checklist to run against any SKILL.md before shipping, plus live known-issues. |
| sources.md | Annotated bibliography of the 13 primary sources. |
| repo-gap-analysis.md | How this repo's standards measure against the official guidance, and the gaps to close. |
How to use
- Writing or reviewing a skill: start with checklist.md, reach for frontmatter-field-spec.md on field questions.
- This repo's own conventions (platform-neutral writing, Contract blocks, validation) live in
.agents/memory/system/and take precedence for repo work;resources/is the upstream ground truth those conventions are built on.
Maintenance
The skill-authoring docs were generated from a research pass and are dated. The platforms move — fields get added, bugs get fixed. Re-verify the time-sensitive items in checklist.md before relying on them, and re-run the research when the date drifts more than a few months.