Stuff: Skill for Claude Code

Read the root AGENTS.md and github-workflow before changes. Use this procedure monthly, or after a recurring failure needs investigation. The weekly todo-triage pass owns backlog maintenance and module audits. code-review owns review of a current PR or completed branch.

Establish the review window

1. Inspect the checkout and fetch origin/main. Unless the user specifies a base or stack, use a clean branch from current main. Search open PRs for an unfinished review-lessons audit before creating another one. If one exists, continue it through the GitHub workflow.
2. Use the requested period, or start from the last complete audit's recorded end time. Find that time in its PR body or accessible run summary. If no complete record is available, review the previous three months. Fix an end time in UTC before collection. Record both boundaries and the source revision.
3. Use three months of context to identify recurring patterns, including lessons from the previous audit. Treat that context as an overlap, not as new evidence on every run. Compare prior audit changes with current source to identify which changes actually landed.

Collect and verify evidence

• Find PRs updated during the review window, including older, closed, and still-open PRs. Read inline threads, submitted reviews, conversation comments, and follow-up fixes. Include resolved threads and later corrections to the original feedback.
• Include completed agent-review reports available in the current task or linked to the reviewed PRs. Preserve findings fixed before publication when the review report or fix history records them. Record inaccessible or unrecorded reviews as a coverage limit. A PR-only collection cannot establish complete agent-review coverage.
• Use scoped TODOs.md files, including completed items, and dated MODULE_AUDIT.md reports to locate agent-found issues and fixes. Verify them against source and history. A backlog entry or stale audit is not proof of a current defect.
• Collect clarification questions about scope, affected callers, validation, and design rationale. Compare them with the description and context available at the time of review. If the earlier description is unavailable, record that limit. Do not infer an earlier omission from the current body.
• Paginate the PR list and every comment, review, and thread collection. Split searches that reach an API result cap. Filter individual events by the recorded window, allowing overlap at the start boundary. Include fixes without review comments from the period's merged PRs and commit history.
• Trace each candidate lesson to the final change and current source. Record the comment, fix, affected behavior, existing guidance, and any executable guard. Treat unmerged proposals and declined feedback as context, not established repository policy.
• Check recovered opening or pre-review revisions when commits were rewritten. A correction may be folded into the final PR's first commit without a separate review comment. Record unavailable earlier revisions as a coverage limit.
• Classify by content and outcome, not by author. Apply the same evidence standard to human and agent findings. Separate substantive findings and clarification questions from status updates, screenshots, suggested review focus, and replies without a new finding. Raw comment volume is not a defect count. Inspect later corrections before counting a finding as confirmed. Deduplicate reports of the same defect across review sources. Repeated mentions are not independent recurrences.
• Before classifying a missed rule, verify that the rule existed when the offending change was made. Use Git history for the relevant instruction and code. Distinguish a missing rule from a rule added after the defect.
• Mark missing pages, unavailable sources, and unverified outcomes explicitly. An incomplete collection cannot establish that no changes are needed. Do not advance the last complete audit's end time from a partial run.