사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/alirezarezvani/claude-skills --skill skill-doctor시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
skill-doctor — grade the agent setup from real sessions
Privacy is the contract. Everything runs locally. Transcripts are condensed, secret-redacted, chmod-0600, and never uploaded — the only shareable artifact is the report the user chooses to share.
Run from the repo being graded. Every artifact goes to one fresh scratch dir, never into the repo:
RUN="$(mktemp -d "${TMPDIR:-/tmp}/skill-doctor-XXXXXXXX")"
python scripts/collect_sessions.py --out "$RUN" # 1 — harvest + redact1 — Collect. Scans Claude Code project-history JSONL and Codex rollouts, discovers repo skills (.claude/skills, .agents/skills, .codex/skills, plugin layouts), detects skill usage (Skill invocations, slash commands, SKILL.md paths), samples newest-first, and writes redacted transcripts. Read $RUN/inventory.json: if sessions_sampled is 0, tell the user there is nothing recent to score (suggest --days 90 or --repo) and stop. skills_found 0 is fine — the report becomes a case for creating skills.
2 — Score. python scripts/score_aggregator.py --inventory "$RUN/inventory.json" --emit-template > "$RUN/session_scores.json". Read each transcript in $RUN/transcripts/ and judge it against both rubrics — scorers/efficiency.md and scorers/code-quality.md. Fill the template with a label from the rubric's table and a 1–3 sentence reason citing transcript specifics. Never invent numeric scores — the aggregator derives them from labels. Use insufficient_evidence when a transcript shows no judgeable diff. Also write 1–5 top_findings: the most impactful cross-session patterns, concrete and specific.
3 — Draft edits. Follow references/skill_edit_governance.md (the filing bar: would a competent agent with the current instructions still fail this way?). For each suggestion that clears it, write the full improved SKILL.md to $RUN/proposed/<skill>/SKILL.md, produce diff -u <current> <proposed>, and record it in $RUN/suggestions.json citing the sampled session id(s) that motivated it. Zero suggestions is a valid success — say why per finding. Never modify the user's real skill files in this step.
4 — Aggregate (the gate). python scripts/score_aggregator.py --inventory "$RUN/inventory.json" --scores "$RUN/session_scores.json" --suggestions "$RUN/suggestions.json". It validates labels against the rubric tables, refuses scores for unsampled sessions, requires substantive reasons, rejects suggestions that cite no scored session, computes overall = 0.5·efficiency + 0.35·code_quality + 0.15·skill_coverage, and writes report.json. Exit 4 is a stop: fix what it names and re-run; never hand-edit report.json around it.
5 — Render + tell. python scripts/render_report.py --report "$RUN/report.json" → one self-contained report.html (no JS, no CDN, dark-mode + print-to-PDF). Then tell the user the grade and the top findings in text, link file://$RUN/report.html, and ask whether to apply the proposed diffs to their real skills — apply only on an explicit yes, skill by skill.
Hard rules
- Never upload transcripts, session files, or any excerpt. Local only.
- Labels only, from the rubric tables. The aggregator owns all arithmetic.
- Every suggestion traces to a scored session — or it is dropped. Generic best practice is not evidence.
- Zero suggestions is a success, not a failure to report around.
- Exit 4 from the aggregator is a stop, not an error to swallow or bypass.
- Never touch the user's real skill files without an explicit per-skill yes; proposed edits live under
$RUN/proposed/. - A proposed skill edit follows write-a-skill discipline — trigger phrase in the description, smallest change that expresses the rule, replace over append.
Scripts
| Script | Role | Exit codes |
|---|---|---|
scripts/collect_sessions.py | Harvest Claude Code + Codex sessions, redact secrets, sample, inventory | 0 · 3 bad input |
scripts/score_aggregator.py | Validate labels/reasons/suggestions, compute grade, emit report.json | 0 · 2 warnings · 3 bad input · 4 validation failure |
scripts/render_report.py | report.json → single self-contained report.html | 0 · 3 bad input |
All support --help, --output json, and --sample (no real history needed).
References and assets
- `scorers/efficiency.md` · `scorers/code-quality.md` — the two rubrics, preserved verbatim from upstream
- `references/transcript_scoring_canon.md` — why rubric-anchored LLM judging works and where it fails (7 sources)
- `references/session_mining_privacy.md` — the local-only contract, redaction pattern canon (7 sources)
- `references/skill_edit_governance.md` — the filing bar for proposing skill edits (7 sources)
- `assets/session_scores.example.json` · `assets/suggestions.example.json` · `assets/report.example.json` — the three handoff shapes
