Skip to content

SKILL.md

Latest commit

 

History

History
114 lines (90 loc) · 4.39 KB

SKILL.md

File metadata and controls

114 lines (90 loc) · 4.39 KB
name review
description Review the current branch's changes against intent. Returns a structured list of findings the caller can hand to a fix-applier, or an empty list when there's nothing to address. Use this for both self-review of your own work and reviewing someone else's PR.
license Apache-2.0
metadata
source audience
autonomous-agents

Review changes

Read the diff against the default branch and produce a structured list of findings. The caller decides what to do with them — push fixes via a subagent, post a review on GitHub, or both.

Process

  1. Resolve the default branch and diff:

    DEFAULT_BRANCH=$(gh repo view --json defaultBranchRef --jq .defaultBranchRef.name)
git diff "$DEFAULT_BRANCH"...HEAD

    Three-dot syntax shows just what this branch added, not unrelated changes that landed on the default branch since branching.

  2. For larger diffs, prefer spawning the explore sub-agent (via the task tool) to do the read pass. The explore agent is read-only and doesn't carry the implementation's "obvious in context" bias.

  3. If a test suite exists, run it to verify the changes don't break anything. Check package.json scripts, Makefile, pytest.ini, go.mod, or CI workflow files for the test command. If tests fail on code unrelated to the diff, note it but don't flag it. If tests fail on changed code, add a bug finding.

  4. Read every change against the PR's stated intent (PR body + linked issues via gh issue view). For each concern, classify it:

    • bug — logic doesn't match intent, broken edge case (null, empty, concurrency), incorrect assumption, factual error.
    • test-gap — new behavior added without test coverage on a critical path.
    • style — inconsistent with surrounding code (string quotes, brace style, naming).
    • scope — diff includes changes unrelated to the stated intent.
    • docs — stale comments or PR description doesn't match diff.

Output format

Emit a JSON block as the FINAL part of your reply, exactly:

{
  "findings": [
    {
      "kind": "bug" | "test-gap" | "style" | "scope" | "docs",
      "file": "<path/to/file>",
      "line": <number or null>,
      "summary": "<one sentence>",
      "suggested_fix": "<one or two sentences describing the fix>"
    }
  ]
}

If there's nothing to address, return { "findings": [] }.

Keep summary and suggested_fix terse — one or two sentences each. The fix-applier reading these has the file open already; don't restate context the diff already carries.

Confidence calibration

Assign a confidence level to each potential finding before including it. See references/confidence-calibration.md for the full table.

  • HIGH — you traced the code path and verified the issue. Report it.
  • MEDIUM — the pattern is suspicious and likely wrong. Report it, noting uncertainty in summary if relevant.
  • LOW — the code looks unusual but could be intentional. Do not report it.

Bug findings should almost always be HIGH. If you can't trace the code path to confirm the issue, it's speculation — downgrade or drop.

When to flag vs. not

  • A style finding only counts if the PR's neighbouring code uses a different convention. Don't flag general taste preferences.
  • A test-gap finding only counts when the missing test would have caught a real bug class — not "every new function needs a test."
  • A scope finding is high-confidence: the diff demonstrably includes a change unrelated to the issue. When in doubt, don't flag.
  • Don't include "everything looks good" notes in findings. Empty array is the correct positive signal.

Not a finding

Before reporting, check references/not-a-finding.md for patterns that look suspicious but are typically correct. Common examples:

  • Null checks on values that "should" be set (defensive, valid if function is public or called from multiple sites)
  • Empty catch blocks in error boundaries (framework convention)
  • Trivial getters/setters (not worth a test-gap finding)
  • Fixing a typo adjacent to changed code (not scope creep)
  • TODO comments describing known limitations (not stale docs)

Adapted from BYK/dotskills (Apache-2.0). The original was prose-only; this version produces structured output so a downstream fix-applier can act on findings mechanically.