メニュー
Audit an existing harness for rot, gaps, and improvement opportunities. Use monthly (health check, quick pass) or quarterly (full audit, thorough assessment), and after any agent incident, major refactor, or structural change.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/adobe/ai-repo-harness-guide --skill harness-inspectコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Write, edit, or restructure content in this repository harnesses guide. Use when adding, updating, or reviewing any of the nine numbered chapters (01–09) or the companion materials (`guide/README.md`, `guide/READING.md`, `guide/FAQ.md`).
Review a multi-chapter guide, series, or book from six specialist perspectives. Orchestrates Series Editor, Non-Technical Proxy, Practitioner Proxy, Consistency Auditor, Diagram Reviewer, and Claims Auditor. Use for single-chapter review, full-series audit, consistency checks, diagram audits, and claims audits.
Build or extend a repository harness. Understands the repo first, then determines what to build, migrate, or create based on what exists.
SOC 職業分類に基づく
SKILL.md を表示中
| name | harness-inspect |
| description | Audit an existing harness for rot, gaps, and improvement opportunities. Use monthly (health check, quick pass) or quarterly (full audit, thorough assessment), and after any agent incident, major refactor, or structural change. |
| metadata | {"author":"https://github.com/afranken","version":"1.0.0","visibility":"public"} |
Two modes. Choose based on available time and the trigger:
| Mode | Scope | When |
|---|---|---|
| Health check | Quick pass | Monthly; after any structural change, make target rename, or harness file edit |
| Full audit | Thorough assessment | Quarterly; after an agent incident, major refactor, or new AI tool adoption |
If the health check surfaces issues, escalate to a full audit.
Check for an argument before proceeding:
| Argument | Action |
|---|---|
health-check | Run the health check directly — skip to the Health check section below |
full-audit | Run the full audit directly — skip to the Full audit section below |
| (none) | Ask the user: "Which mode — health-check (quick pass, monthly) or full-audit (thorough assessment, quarterly)?" Do not assume a default. |
Load references/before-after-probe.md to record current agent behavior before making changes. Run again after fixes to measure impact.
Load references/health-check.md and run the prompt twice independently (without referencing your first pass on the second run). Compare pass/fail verdicts; resolve any disagreements. Report pass/fail for each of the ten questions and list action items only — no full analysis.
Load references/full-audit.md and run the prompt twice independently (without referencing your first pass on the second run). Compare verdicts; resolve disagreements. For repeat runs, fill in the inspection mode and paste the prior verdict table — the agent will produce a delta automatically.
| Finding type | Action |
|---|---|
| Rot | Fix immediately — stale guidance actively misleads agents |
| Critical gaps | Address before the next agent task in the affected area |
| Quick fixes | Address in the current session — they compound if deferred |
| Token optimizations | Address when convenient; reduces initialization cost over time |
| Larger improvements | Add to project backlog with the assessment date as context |
Before acting on any finding immediately (rot, quick fixes), write a specific plan — file, change, reason. If your tool supports a planning mode (e.g., Claude Code /model opusplan), use it for the audit run itself, then switch to execution mode for fixes.
