Retro

Turn a demonstrated mistake or repeated obstacle into the smallest useful improvement to this repository's development workflow. A finding that existing checks already cover, or that needs no additional prevention, is a valid outcome.

Establish the evidence

Use the current task's conversation, diff, review feedback, and available failure output. For another session or PR, use the evidence the user identifies; state missing context rather than inventing a history or searching unrelated sessions. Routine successful edits do not need a retrospective.

Read applicable AGENTS.md guidance and the relevant package/build scripts, check configuration, hooks, and CI jobs before proposing a check. Distinguish a missing check from one that exists but is unwired, skipped, or unable to fail its caller. During an explicit repository-wide retrospective, absent automated lint, type, or test enforcement is itself worth reporting. Inspect configuration first; do not launch a full build or test suite just to inventory it.

Choose useful prevention

• For a mechanical issue, prefer a deterministic check over another prose rule: an existing linter option, a focused regression test, or a small extension to the repository's checker. Reuse the normal check commands and CI; use a commit hook only for checks that are fast and suitable for staged changes.
• For behavior bugs, cover the failing behavior or invariant. For syntax, imports, or module placement, prefer the existing parser/linter or architecture checker. Avoid broad text matches that confuse valid code with the defect.
• For a judgment call that cannot be checked reliably, clarify the relevant existing review guidance or playbook. Do not create a new standards document merely to restate existing instructions.
• For repeated navigation or information gaps, add a focused pointer or repair the existing tool/documentation path when the session demonstrates why it is needed.

Weigh impact and recurrence against maintenance cost and false positives. Do not invent a rule from a one-off preference, duplicate existing coverage, weaken checks to make a run pass, or expand into unrelated cleanup. Keep repository policy in committed repository files; do not modify global instructions or other repositories.

Complete within scope

A review-only or ideas-only request returns recommendations without edits. During authorized implementation, apply the smallest relevant prevention within the task's scope; report broader workflow or policy changes as follow-ups. Existing authorization continues to apply, and this skill adds no approval gate or publishing permission.

When adding a check, verify that it detects the demonstrated failure and accepts the corrected behavior, using the repository's affected-check guidance. Reuse valid verification evidence; do not repeat unrelated checks. Keep retrospective reasoning in the agent workflow rather than adding an AI call to every commit or a new scheduled job.

Report only worthwhile findings, with the evidence, the prevention applied or proposed, and any verification limits. If no additional prevention is justified, say so briefly.

Inspired by mattpocock/skills retro PR #1083.