audit skill
Verify implementation matches change artifacts — dual-layer verification of code vs spec compliance and code vs real requirements. Use after implementation is complete.
- GitHub 스타
- 13
- 파일
- 1
- 마지막 커밋
- 2026년 6월 23일
- 라이선스
- MIT
- 허용된 도구
- —
audit skill은 무엇을 하나요?
audit은(는) agent skill입니다. Verify implementation matches change artifacts — dual-layer verification of code vs spec compliance and code vs real requirements. Use after implementation is complete.
audit skill 설치 방법
각 명령은 skill 폴더를 클라이언트의 skills 디렉터리에 복사합니다. 클라이언트를 재시작하거나 새 세션을 시작하여 로드하세요.플러그인에 audit skill 설치
/plugin marketplace add devnomad-byte/techneering /plugin install tn@techneering
Claude Code에 audit skill 설치
~/.claude/skills/audit/SKILL.md
$ npx degit devnomad-byte/techneering/skills/audit ~/.claude/skills/auditClaude Code (project)에 audit skill 설치
.claude/skills/audit/SKILL.md
$ npx degit devnomad-byte/techneering/skills/audit .claude/skills/auditCodex에 audit skill 설치
~/.agents/skills/audit/SKILL.md
$ npx degit devnomad-byte/techneering/skills/audit ~/.agents/skills/auditGemini CLI에 audit skill 설치
~/.gemini/skills/audit/SKILL.md
$ npx degit devnomad-byte/techneering/skills/audit ~/.gemini/skills/auditClaude.ai에 audit skill 설치
npx degit devnomad-byte/techneering/skills/audit audit && zip -r audit.zip audit # claude.ai → Customize → Skills → + → Create skill → Upload a skill → audit.zip
SKILL.md
GitHub에서 보기Verify Implementation
Dual-layer verification: does the code match the spec, and does the code solve the real problem?
Core Principle
NO COMPLETION WITHOUT VERIFICATION.
Every requirement traced. Every decision checked. Every gap documented.
Steps
Step 1: Select the Change
Same selection logic as tn:forge — auto-detect or ask user.
Always announce: "Verifying change: "
Step 2: Load All Artifacts
Read:
techneering/changes/<name>/proposal.md— for "Why" (Layer 2)techneering/changes/<name>/specs/**/*.md— for requirements (Layer 1)techneering/changes/<name>/design.md— for architecture decisions (Layer 1)techneering/changes/<name>/tasks.md— for task completion check
Step 3: Layer 1 — Code ↔ Spec Compliance
3a. Completeness Check
- Parse
tasks.md: count- [ ](incomplete) vs- [x](complete) - If incomplete tasks exist → CRITICAL
- Extract all requirements from delta specs (
### Requirement:headers) - For each requirement, search codebase for implementation evidence
- If requirements appear unimplemented → CRITICAL
3b. Correctness Check
- For each requirement from specs:
- Search codebase for implementation
- Assess if implementation matches requirement intent
- If divergence → WARNING
- For each scenario (
#### Scenario:):- Check if conditions are handled in code
- Check if tests exist covering the scenario
- If uncovered → WARNING
3c. Coherence Check
- If design.md exists:
- Extract key decisions
- Verify implementation follows those decisions
- If contradiction → WARNING
- Review new code for consistency with project patterns
- If significant deviations → SUGGESTION
Step 4: Layer 2 — Code ↔ Real Requirements
- Read the Why section of proposal.md
- Trace back from the code to the original problem statement
- Ask: Does this implementation actually solve the original problem?
- Check if the "What Changes" in proposal.md are all addressed
- If the implementation solves a different problem than stated → CRITICAL
Step 5: Generate Verification Report
## Verification Report: <change-name>
### Layer 1: Spec Compliance
| Dimension | Status |
|--------------|------------------|
| Completeness | X/Y tasks, N reqs|
| Correctness | M/N reqs covered |
| Coherence | Followed/Issues |
### Layer 2: Real Requirements Check
- Original problem: <from proposal.md Why>
- Implementation addresses it: YES/NO
- Gap: <if any>
### Issues
**CRITICAL** (Must fix):
- [issue with file:line reference]
**WARNING** (Should fix):
- [issue with recommendation]
**SUGGESTION** (Nice to fix):
- [issue]
### Final Assessment
- [PASS/WARN/FAIL with reasoning]
Red Flags
| Thought | Reality |
|---|---|
| "Tests passed, that counts as verification" | Tests verify code behavior, not whether requirements are met |
| "Spec and code roughly match is fine" | "Roughly" means divergence — must compare item by item |
| "Skip Layer 2 to save time" | Layer 2 catches "built but not what was wanted" problems |
| "Submit first even if verification fails" | Failed is failed — no "submit first" |
| "Give up after 3 failed rounds" | More than 3 rounds means a systemic problem needing human intervention |
Quick Reference
| Dimension | Key point |
|---|---|
| Dual-layer | Layer 1: code↔spec / Layer 2: code↔real requirements |
| Issue levels | CRITICAL / WARNING / SUGGESTION |
| Rollback logic | small deviation → fix / medium → redo / fundamental → back to draft |
| Loop limit | Max 3 rounds of audit → fix, then report to user |
| Path placeholder | Always use techneering/changes/<name>/ |
Rollback Logic
When verify finds deviations:
| Deviation Level | Action |
|---|---|
| Small (1-2 missing details) | Fix directly → re-run tn:gate → re-run tn:audit |
| Medium (wrong architecture direction) | Back to tn:forge, redo failed tasks → tn:gate → tn:audit |
| Fundamental (spec itself wrong) | Back to tn:draft to fix spec → tn:forge from scratch |
Loop limit: Max 3 rounds of verify → fix. After 3 rounds, pause and report to user.
Next Step
When done (following the user's choice in the tn:gate menu):
- User chose full verification (1) → after passing, invoke
tn:inspect - User chose standard flow (2) → after passing, invoke
tn:ship - User chose verify only (4) → finish after done, don't auto-flow
- Verification has WARNING → fix, then re-run
tn:gate→tn:audit - Large deviation → roll back to
tn:forge, redo failed tasks - Spec itself wrong → roll back to
tn:draft, fix spec docs - User requests stop → end the current flow
Guardrails
- Every issue must have specific file:line references and actionable recommendations
- No vague suggestions like "consider reviewing"
- When uncertain, prefer SUGGESTION over WARNING, WARNING over CRITICAL
- Graceful degradation: if only tasks.md exists, verify task completion only
- Always note which checks were skipped and why
- Unified path placeholder: always use
techneering/changes/<name>/
이 스킬의 파일
1audit skill 질문
Claude Code에 audit skill을 어떻게 설치하나요?
npx degit devnomad-byte/techneering/skills/audit ~/.claude/skills/audit를 실행합니다. Claude Code는 ~/.claude/skills(모든 프로젝트) 또는 .claude/skills(한 프로젝트)에서 skills를 로드하며, 설명이 작업과 일치할 때 skill을 사용합니다.
audit skill을 사용할 수 있는 클라이언트는 무엇인가요?
Agent Skills 형식을 읽는 모든 클라이언트(Claude Code, Codex, Gemini CLI, claude.ai(ZIP으로 업로드))가 사용할 수 있습니다. skill 내 스크립트는 클라이언트가 코드 실행을 허용하는 경우에만 실행됩니다.