メニュー
Before opening a PR, audit the working diff for test + doc + config gaps — surface changes that shipped without matching tests, docs, or env updates — and remediate or flag them. Gate PR submission on critical gaps.
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
| name | pre-pr-audit |
| description | Before opening a PR, audit the working diff for test + doc + config gaps — surface changes that shipped without matching tests, docs, or env updates — and remediate or flag them. Gate PR submission on critical gaps. |
| audience | tenant |
| allowed_tools | ["file_read","file_write","bash_execute","lint_and_typecheck","test_coverage_for_diff","git_create_pr","deploy_preflight"] |
| metadata | {"category":"quality","complexity":"medium"} |
You run immediately before opening a PR, after the implementation is complete and locally verified. Your job is to catch the gap between what changed and what was tested + documented, then close the closable gaps and flag the rest.
This is different from code-review, which is a post-hoc quality check. pre-pr-audit is a pre-submission gate: it blocks PR creation if the diff ships code that isn't tested, exposes config no one documented, or introduces public API no docs mention.
main) — diff compared against this.Run these in order. Each emits findings; findings accumulate into the final report.
From git diff <base>...HEAD + uncommitted changes:
tests/, *.test.*, test_*.py.).env.example? Is it in CLAUDE.md if it's load-bearing?CLAUDE.md — stale status tables, missing mention of new subsystem, outdated architecture diagrams.README.md — Quick Start missing new commands, missing install steps for new deps..env.example — every new env var documented with a comment explaining purpose + fallback behavior.docs/ — any file whose subject overlaps the changed code should be checked for staleness.Use the test_coverage_for_diff tool — it runs coverage scoped to the
diff vs the base ref and returns structured {file, uncovered_lines}
findings without you having to parse pytest output.
test_coverage_for_diff(base_ref="main", repo_path=".")
→ {uncovered: [...], summary: {files_changed, changed_lines_uncovered}}
Any entry in uncovered is a 🔴 finding. If changed_lines_uncovered == 0 but no test files changed, downgrade to 🟡 (tests may have existed
already and just exercised the new code).
Use the lint_and_typecheck tool — it runs ruff + mypy (Python)
and/or eslint + tsc (TypeScript), auto-detects language, and returns
structured findings. Missing toolchains are skipped (not errored) so
a Python-only repo without pnpm won't trip this pass.
lint_and_typecheck(paths=["."], repo_path=".")
→ {findings: [{tool, severity, file, line, code, message}], skipped: [...]}
Every severity: error finding is a 🔴 blocker. Warnings are 🟡.
.env.example (with placeholder value).deploy_preflight(overlay_path="infra/k8s/<overlay>") — it surfaces un-pinned images, missing resource requests, missing privileged: false, and other Kyverno-policy blockers. verdict: "blocked" → 🔴.For 🔴 and 🟡 findings, close them in the same PR whenever the fix is local:
.env.example entry → add with a comment explaining purpose + fallback.coding skill's test patterns).For findings you can't close (e.g. an external doc that requires another team's sign-off), add a bullet to the PR description under ## Follow-ups so the reviewer has the full picture.
Return structured JSON:
{
"verdict": "ready" | "needs-remediation" | "blocked",
"surface_delta": {
"added": [{"kind": "endpoint", "symbol": "POST /v1/render/card", "tested": true, "documented": true}],
"changed": [...],
"removed": [...]
},
"findings": [
{"severity": "blocker|worth-doing|nice-to-have",
"category": "test|doc|config|surface",
"file": "path/to/file",
"description": "...",
"remediation": "action taken | follow-up noted | requires external review"}
],
"remediations_applied": <int>,
"follow_ups_logged": <int>,
"test_coverage_delta": {"files_changed": N, "files_with_new_tests": M}
}
verdict = "blocked" → do NOT call git_create_pr. Return the finding list to the caller.verdict = "needs-remediation" → close the closable gaps, re-run the audit, then proceed.verdict = "ready" → call git_create_pr(title=..., body=...) to create the PR. The tool enforces additional guards (protected branch refuse, conventional-commit warning, CODEOWNERS auto-reviewers, PR template fallback) — include the audit report in the body under ## Pre-PR Audit.coding skill + CI.code-review skill.Caller: "Audit tests and docs before I open the PR for the CEQ render pipeline."
Skill:
1. git diff main...HEAD
2. Grep new symbols, check test coverage
3. Check CLAUDE.md, README.md, .env.example for staleness
4. Emit findings, close 🔴/🟡, log 🟢 as follow-ups
5. Return verdict + report → caller proceeds or blocks
Governed campaign copy generation from Tulana SKU packs -- claims-disciplined email subject lines, body copy, and CTAs grounded only in campaign-safe claims, es-MX primary.
Production-ready code implementation following MADFAM coding standards with git worktree workflows, strict type checking, and test-driven development.
Professional email delivery via Resend API — transactional, marketing, and autonomous CRM outreach across 10 verified MADFAM ecosystem domains.
Model Context Protocol server development for building platform extensions and tool integrations.
Refresh the staging database from prod with PII masked. Dump prod → mask-and-copy into staging → flip ArgoCD staging app to reconcile → smoke-test the staging URL → emit a refresh report. HITL-gated at the mask step because the target database is overwritten in place. Per the PhyndCRM PP.5 staging spec.
End-to-end testing with Playwright patterns for office-ui, Phaser scene testing, and comprehensive test coverage.