Skilltail
Use Ponytail logic for skills: stop at the first rung that holds.
- Is a skill needed, or would a prompt/rule do?
- Can existing skill-creator/plugin tools handle it?
- Can prose be deleted or moved to human docs?
- Can validation be run locally?
- Only then add the smallest durable instruction/resource.
Names:
- skill-compress / skill compression / 技能压缩
- skill-refresh / skill refresh / 技能迭代更新
- skill-seal / skill sealing / 技能安全封装
- skill-release / release readiness / 技能发布整备
Route:
- Compress/readability split: read references/skill-compress.md.
- Version/update from usage: read references/skill-refresh.md.
- Security audit: read references/skill-seal.md.
- Public release polish: read references/skill-release.md.
Run checks:
- python3 scripts/skilltail_check.py validate <skill-dir>
- python3 scripts/skilltail_check.py audit <path>
- python3 scripts/skilltail_check.py release <repo-root>
- python3 scripts/skilltail_check.py self-test
Always:
- Keep SKILL.md as short AI-operational text; put human explanation in README.md and state AI should not load it unless asked.
- Preserve only trigger text in frontmatter; body is procedure.
- Prefer references/scripts/assets over long body text.
- Validate at least once. If the user gives a count, run that many feasible passes.
- Ask before live external messages, production changes, credential writes, or non-sandbox file generation; otherwise use the local checks above or minimal sandbox artifacts.
- Leave the smallest runnable check that would catch breakage.
