AI Comments

AI Comments: triage AI reviewer feedback on a PR into a worked list

A PR has accumulated review comments from AI reviewers (GitHub Copilot and a Claude review bot) and the user wants them triaged once instead of re-explaining the drill every session. Your job is to fetch every open AI comment, actually understand the code each one touches, categorize each finding with an honest confidence rating, and write the result to a local state file. The classic failure mode is shallow triage — categorizing from the comment text alone without reading the code, which produces confident-sounding but wrong verdicts. Read the code first, always. You triage only; fixing starts with /ai-next.

Configuration (set per adopter)

State file path: {AI_STATE_DIR}/{repo}-pr-{n}.md.

When to use this skill

• "Pull / triage / categorize the AI comments" on a PR
• A PR link or number is pasted with intent to work through its review feedback
• After an AI re-review lands new comments on a PR already being worked (re-run to merge them in)

When NOT to use this skill

• Working through an already-triaged list — use /ai-next.
• Drafting or posting replies — use /ai-draft / /ai-post.
• Reviewing the PR yourself — this skill only ingests other reviewers' AI comments.
• Human reviewer comments — use /hum-comments; the queues stay separate.

Steps

1. Resolve the PR. If a PR URL or number was passed, use it. Otherwise gh pr view for the current branch. If neither resolves to a PR, ask via AskUserQuestion and stop.

2. Fetch the AI comments. Two sources:

   • Copilot (inline): gh api repos/{owner}/{repo}/pulls/{n}/comments — keep comments whose author is Copilot; record each comment's id, path, and line.
   • Claude (issue-level): gh pr view {n} --json comments — keep comments whose author is the Claude review bot (typically claude) containing a ### Code Review section. One Claude review usually contains several findings — split it: one finding = one item. Ignore everything else: human reviewers, CI bots, coverage/security scanners.

3. Drop what's already handled. Skip: inline threads where the PR author has already replied; Claude findings that a later incremental Claude review says are resolved (parse reviews newest-first — the latest review is authoritative). If the state file from a previous run exists, merge: add only items whose comment id isn't already listed, and never change existing items' statuses.

4. Understand each finding. Read the code the comment points at — the file, the surrounding function, and the tests that cover it. For findings about shared code or behavior claims (callers, blast radius, "this breaks X"), get the full picture: if a call-trace skill is installed, invoke it; otherwise trace callers/callees manually with Grep/Read. Do not categorize from the comment text alone.

5. Categorize with confidence. For each finding assign SHOULD FIX or NITPICK plus a confidence percentage that the categorization is right. Under 85% → the item is UNSURE. Every SHOULD FIX must carry a Why (grounded in the code you read) and a concrete Proposed solution.

6. Resolve UNSURE items by asking. For each UNSURE item, ask the specific question that would settle it via AskUserQuestion (batch related ones). Re-categorize with the answers. Anything the user can't settle now stays UNSURE in the file.

7. Write the state file to {AI_STATE_DIR}/{repo}-pr-{n}.md (create the folder if missing) in the format below, items ordered SHOULD FIX → UNSURE → NITPICK, with the PR title, repo, and today's date in the header — these files are kept long-term. If AI_STATE_DIR is inside a repo, never commit the file.

8. Stop. Print the summary table (below) and point to /ai-next to start working the first item. Do not begin fixing anything.

Output format

Chat summary, most important first:

Triaged PR #123 — 7 AI findings (4 should-fix, 2 nitpicks, 1 unsure) → <AI_STATE_DIR>/repo-pr-123.md

| # | Cat        | Source  | Conf | Finding                          |
|---|------------|---------|------|----------------------------------|
| 1 | SHOULD FIX | Copilot | 93%  | paging guard never fails fast    |
| ... |
Next: /ai-next

State file ({AI_STATE_DIR}/{repo}-pr-{n}.md) — the contract the whole suite reads:

# AI Review — PR #123: <title>
<!-- ai-review suite state (/ai-comments → /ai-next → /ai-draft → /ai-post). Kept long-term — never commit, never prune. -->

- Repo: <owner>/<repo>
- PR: <url>
- Intake date: <yyyy-mm-dd>
- Head at last intake: <sha>

## Items

### 1. [SHOULD FIX] <short finding title>
- **Source:** Copilot (inline comment <id> on `<path>:<line>`)   <!-- or: Claude (issue comment <id>, finding "<label>") -->
- **Confidence:** 93%
- **Comment:** <the finding, condensed but faithful>
- **Why:** <why this is real, grounded in the code you read>
- **Proposed solution:** <concrete change>
- **Status:** OPEN

Later skills append to each item: - **Resolution:** + Status change (/ai-next), - **Draft reply:** (/ai-draft), - **Posted:** yes (/ai-post). UNSURE items carry - **Open question:** instead of Why/Proposed solution until resolved.

Rules

What to do

• Read the code before categorizing — every item. A verdict formed from comment text alone is the failure mode this skill exists to prevent.
• One finding = one item. Split multi-finding Claude reviews; each finding gets its own category, confidence, and eventual reply.
• Honest confidence. The number reflects your evidence, not your desire to avoid questions. Under 85% → UNSURE, no rounding up.
• Every SHOULD FIX carries Why + Proposed solution. No bare verdicts.
• Merge, don't clobber. On re-run, only append new items; existing statuses are /ai-next's territory.

What NOT to do

• NEVER ingest human comments. AI reviewers only — humans go through /hum-comments.
• NEVER start fixing. Triage ends at the state file; /ai-next owns implementation.
• NEVER commit the state file (when it lands inside a repo) and never prune it — worked history stays; the files double as a long-term record.
• Don't inflate the list. A Claude review that says "no issues found" produces zero items, not a synthetic "looks good" entry.

Format discipline

• Summary table first, no preamble, no restating each comment in chat — the detail lives in the state file. Stop after pointing to /ai-next.