pr-review

Files (1)

• SKILL.md

SKILL.md (first lines)

## 1. Pin the scope
- Get the fixed point (base branch, commit, or PR). If not given, ask.
- Run `git diff <base>...HEAD` and `git log <base>..HEAD --oneline`. Stop if the
  ref does not resolve or the diff is empty.
- Find the spec: issue referenced in commits, a path the user gave, or a doc in
  `docs/`/`specs/`. If none exists, say so and skip the Spec axis.

## 2. Set depth by risk
State the tier and why at the top of the report.
- **Low** (docs, config, tests, refactor already covered by tests): one
  structural read, Correctness and Standards axes only.
- **Medium** (ordinary feature or bugfix): every axis, on the changed code.
- **High** (data migration/deletion, auth/security, shared API/schema contract,
  concurrency, client delivery, or no test would catch a regression): every axis,
  plus callers/consumers of changed code, failure modes, and rollback. Suggest an
  independent second reviewer (fresh subagent or `ponytail-review`).

## 3. Review on separate axes (do not merge or rerank them)
**Spec**: requirements missing or partial; behavior not asked for (scope creep);
requirements implemented wrongly. Quote the spec line for each.

**Correctness**: logic, edge cases, error handling, concurrency if relevant.
Are the changes covered by tests that verify behavior, not implementation?

**Standards**: repo `AGENTS.md`/`CONTRIBUTING.md` and existing patterns. A
documented repo rule always wins; skip what linters already enforce.

**Safety and contract**: no secrets or client data in code/logs/fixtures,
external inputs validated, breaking changes to APIs/schemas flagged.

**ML (when applicable)**: data leakage between splits, seeds/config/data
version traceable (see `ml-repro`), metrics compared against a baseline,
`ml-over:` when a lower rung of the ML ladder would suffice.

## 4. Report
One line per finding, `file:line — problem — fix`, ordered by severity, grouped
by axis. Close with a verdict (ready to merge / needs changes), blockers listed
separately from suggestions. For large diffs, suggest running `ponytail-review`.

From G4rb4sh/MLE-setup, MIT.

More from G4rb4sh/MLE-setup