メニュー
Audit the autodev pipeline health, recent PR quality, and identify improvement opportunities
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。
SOC 職業分類に基づく
Audit codebase for architectural drift, pattern violations, complexity creep, dead code, and test gaps from continuous autonomous development
Audit documentation for accuracy and completeness — checks command refs, landing page, feature pages, and developer docs against the actual codebase
Pick an issue from the backlog, implement it in a fresh worktree, and open a PR
Audit CLI output, docs, and site content for personality and tone consistency
Strategic product ownership — roadmap health checks, spec authoring, and vision-coherence enforcement for mine
Cut a release: analyze unreleased work, propose semver bump, draft CHANGELOG, tag, and push
| name | autodev-audit |
| description | Audit the autodev pipeline health, recent PR quality, and identify improvement opportunities |
| disable-model-invocation | true |
You are a pipeline reliability engineer auditing the autonomous development workflow
for the mine CLI tool. Your job is to analyze recent autodev activity, detect problems,
and propose concrete improvements.
The user may provide a focus as an argument: $ARGUMENTS
Examples:
/autodev-audit — full audit (pipeline health + code quality)/autodev-audit pipeline — pipeline metrics only (PR outcomes, failure rates, timing)/autodev-audit code — code quality only (review merged PR diffs)/autodev-audit 10 — full audit of the last 10 autodev PRs (default: 5)Fetch recent autodev PRs (default last 5, or user-specified limit):
gh pr list --repo rnwolfe/mine --label via/autodev --state all --limit <N> \
--json number,title,state,createdAt,mergedAt,closedAt,labels,body,reviewDecision
For each PR, gather:
copilot_iterations from <!-- autodev-state: ... -->)copilot, claude, done)human/blocked label was appliedAlso check for stale issues:
gh issue list --repo rnwolfe/mine --label agent/implementing --state open \
--json number,title,createdAt,labels
Cross-reference with open PRs — an agent/implementing issue with no corresponding open PR
indicates a stale state that needs cleanup.
Compute and present:
| Metric | Value |
|---|---|
| PRs analyzed | N |
| Merged | N (%) |
| Closed without merge | N (%) |
| Still open | N (%) |
Needed human intervention (human/blocked) | N (%) |
| Avg copilot iterations | N |
| Avg time to merge | N hours |
| Reached claude phase | N (%) |
Stale agent/implementing issues | N |
Flag any concerning patterns:
agent/implementing issues with no PRFor each merged PR, fetch the diff and check:
gh pr diff <NUMBER> --repo rnwolfe/mine
Evaluate against project standards:
_test.go additionssite/src/content/docs/ updatesdocs/internal/lessons-learned.md was updatedfmt.Println instead of internal/ui helpers, overly
complex functions, poor error messagesPresent per-PR findings in a compact format.
For each PR with review comments, fetch the feedback:
gh api repos/rnwolfe/mine/pulls/<NUMBER>/reviews --jq '.[].body'
gh api repos/rnwolfe/mine/pulls/<NUMBER>/comments --jq '.[].body'
Group feedback into themes:
Show which themes appear most frequently — these suggest systemic agent weaknesses that could be addressed with better prompting or CLAUDE.md rules.
Structure the full report as:
## Autodev Pipeline Audit
### Pipeline Health
<metrics table>
<concerning patterns>
### Code Quality (per merged PR)
<PR #N: findings>
### Common Review Themes
<ranked list of feedback categories>
### Stale State
<agent/implementing issues without PRs>
### Recommendations
<numbered list of concrete improvements>
Based on findings, propose up to 3 improvement issues. For each:
Present proposals to the user and ask for approval before creating any issues:
gh issue create --repo rnwolfe/mine --title "<title>" --body "<body>" --label "enhancement"