Raw view: Pre-PR Self-Review Checklist
No invisible, zero-width, or bidi-control characters detected.
Review this diff as the most exacting reviewer on the team would, before the PR is opened. The author wants to find problems now, while they are cheap, so be direct. Praise is not useful here; precise findings are.
<diff>
{{DIFF}}
</diff>
<intent>
{{INTENT}}
</intent>
<test_output>
{{TEST_OUTPUT}}
</test_output>
<review_norms>
{{REVIEW_NORMS}}
</review_norms>
<intent> is what the change is supposed to accomplish, in the author's words. <review_norms> may hold the team's conventions (commit style, test requirements, forbidden patterns); treat it as the rubric when present. <test_output> may be empty; if so, note that tests were not shown rather than assuming they pass.
Review in passes, each with a different question, and only read the diff once per pass:
1. Does the diff do what <intent> says, and only that? List every hunk that is not required by the intent (drive-by refactors, formatting churn, unrelated fixes). These belong in separate PRs because they hide the real change from reviewers and from `git bisect`.
2. Where can it be wrong? For each changed function, name the inputs that would break it: empty, null, negative, very large, concurrent, unicode, timezone boundaries. Check whether the diff handles them or whether existing callers make them impossible.
3. What does the diff change that the author may not have intended? Public signatures, serialized formats, error types, log messages that dashboards parse, default values, ordering.
4. Is it tested at the level it is risky? A new branch with no test, a bug fix with no regression test, or a test that asserts on mocks instead of outcomes each count as a finding.
5. Will a reader in a year understand it? Flag names that lie, comments that restate code, and comments that are missing where the logic is non-obvious.
6. Predict the review. Write the three questions a reviewer is most likely to ask, so the author can answer them in the PR description instead of in a comment thread.
Cite every finding by file and line from the diff. Mark severity: `blocker` (wrong or unsafe), `should-fix` (correct but costly later), `nit` (style). Do not inflate nits into blockers. If a pass finds nothing, say "Nothing found" for that pass; do not pad.
## Output format
**Verdict** - one line: `Ready to open`, `Fix blockers first`, or `Split this PR`, with a one-clause reason.
**Findings** - a table with columns: severity, file:line, finding, suggested change. Sorted by severity. Keep "finding" to one sentence and "suggested change" to one concrete action.
**Scope creep** - bullet list of hunks to move to another PR, or "None".
**Test gaps** - bullet list of behaviors introduced or changed without a test, each as a proposed test name.
**Reviewer will ask** - the three questions from pass 6, each with a suggested one-sentence answer for the PR description.
**PR description draft** - 3-6 lines: what, why, how to verify. Written for the reviewer, in the team's style if <review_norms> shows one.