ii23-skills

Each skill lives at skills/<name>/ with a SKILL.md and, where it needs them, a scripts/ folder. Scripts are stdlib-only Python so a skill installs by copying a folder.

Adding, renaming, or removing a skill means updating three places, or the repo starts lying about itself:

1. README.md — a problem-and-fix section under "Why these exist", plus a row in the Skills table. Every mention links to the skill's SKILL.md.
2. skills.sh.json — the skill's name in a grouping's skills array.
3. docs/<name>.md — the human-facing page. Follow .agents/writing-docs.md; a rename moves the file.

SKILL.md is the agent contract: what to run, in what order, and the traps to avoid while doing it. It is not the place for the reasoning behind a decision. That goes in .agents/adr/, and a rejected approach goes in .out-of-scope/ with the measurement that killed it — so nobody re-proposes it and no agent burns a session rediscovering it.

Numbers in any document are measured, never estimated. Every performance claim in this repo came from a real run on a real file; if you change a claim, re-run the measurement first and say what you ran it on.

Validate a skill after editing it:

python3 ~/.agents/skills/skill-creator/scripts/quick_validate.py skills/<name>

Frontmatter limits it enforces: name <= 64 chars, description <= 1024. The description is what makes a skill trigger, so it carries the trigger phrases (including Thai ones) and an explicit "do NOT use for ..." boundary.