AGENT SKILL

skill-improver

Audit and improve skills - preserve intent, fix frontmatter, convert tutorial prose to actionable rules, track usage.

License 许可
Apache-2.0
最近更新
2026-11-10
Stars
★ 0
Forks
0

01 前言 Front Matter

name
skill-improver
triggers
Skill improvement, audit skills, refactor skills, skill refresher, drift detection, auto-heal
changelog
docs/ciclos/cycle28-20260815.md
description
Audit and improve skills - preserve intent, fix frontmatter, convert tutorial prose to actionable rules, track usage.
token_budget
2600

02 资源文件 1

03 SKILL.md 正文

When to Use

Audit and improve skills — preserve author intent, fix frontmatter, convert tutorial prose to actionable rules, track usage.

Rules

  1. Preserve author intent, critical rules, activation semantics.
  2. Audit-only by default.
  3. Never delete: content → references/.
  4. Don't invent triggers — mark ambiguous.
  5. Skills only — no opencode.json, no AGENTS.md.

Decision Gates

Signal Action
Invalid frontmatter Fix metadata
Tutorial prose Convert to rules, bg → references/
Over budget (>3200 B — check-token-budget.ps1 -Json) Rules stay, examples → references/
Branching prose Decision table
Conflicting rules Report both — escalate

Anti-Patterns

Anti-Pattern Fix
Rewrite entire skill Audit → fix gaps → repeat
Add unintended triggers Mark ambiguous
Delete instead of archive references/
Major bump for trivial minor=content
Ignore Engram mem_search first
Over-compress rules Preserve rules, compress examples

Anti-Rationalization

Rationalization Red Flag Verification
"reescribir en vez de audit-fix preservando intención" Rewrite completo borra intent original Regla 1 Preserve intent + audit-fix gaps only
"borrar en vez de archivar" Contenido eliminado sin mover a references/ Regla 3 Never delete → references/ + changelog
"añadir triggers no solicitados" Triggers nuevos no pedidos en spec Regla 4 Don't invent triggers — mark ambiguous

Red Flags

  • Doing work without checking output format → STOP, re-read skill
  • Second occurrence of same rationalization → force RED zone

Audit

  1. Read all SKILL.md in target. 2. Audit: metadata · triggers · budget · actionability · gates. 3. mem_search(last 50) — flag 90d+ untouched. 4. Report grouped-by-skill with severity.

Regeneration

  1. Audit + Engram (10 sessions) → 2. Drift check → 3. Fix gaps → 4. Compress if >5% new → 5. Version + changelog.

Output

{"audited":["skill-a"],"issues":[{"skill":"skill-a","severity":"high"}],"drifted":["skill-b"]}

Verification

  • Output matches skill ## Output contract + file:line citaton
  • cross-ref-check.ps1 → SKILL.md OK

Refs

karpathy-loop · opencode-skill-creator · skill-testing · immune-system · dreaming · skill-registry · skill-graph

Reference

docs/skills/skill-improver/reference.md

04 同仓库更多 Skills