ClaudeCode reviewLevel 2: Context

How Do You Rank Claude's Code Review Findings?

Ask for a fixed scale, and a concrete failure for anything at the top of it. Anthropic's Code Review service uses three levels, which work well in a prompt too: Important is a bug to fix before merging, Nit is minor and doesn't block, and Pre-existing is a real bug that this change didn't introduce.

Prompt to copy
Review [the diff between main and this branch / pull request 1234].

Label every finding with one level:
- IMPORTANT: a bug that should be fixed before merging. Give the input that triggers it and the wrong result.
- NIT: minor, worth fixing, doesn't block. Report at most [5].
- PRE-EXISTING: a real bug this change didn't introduce.

In this repo, IMPORTANT means [incorrect logic, unscoped database queries, personal data in logs, migrations that aren't backward compatible]. Style and naming are NIT at most.

Sort worst first and cite file and line. Base each claim on code you read, not on what a name suggests. If you're unsure about one, say so. If nothing is IMPORTANT, say that, and don't promote a nit.
  1. 1Importantfix before merging
  2. 2Nitminor, doesn't block
  3. 3Pre-existingreal, but not from this change
The three severity levels in Anthropic's Code Review service. They work as a scale in your own prompt too.

Try it now

  1. 1Run /code-review low for a first pass with only the most confident findings.
  2. 2Run /code-review high for more coverage. Later runs reuse the level you typed, even in a later session, and a notice such as Reusing high effort, the level you typed last time tells you when that happens.
  3. 3Paste the prompt above when you want your own definition of Important for this review.
  4. 4You'll know the ranking works if you can stop reading after the first group and still know what blocks the merge.

Use it when

Any review likely to produce more than a handful of comments, especially when you have to decide quickly what blocks the merge. Also when a script or CI job has to act on the result.

Skip it when

A one or two line change, where a ranking adds nothing. And don't let the label make the decision for you. Code Review itself never approves or blocks a pull request, and its docs say a review with no findings doesn't mean your code is error-free or secure.

Why it works

An unranked list treats a data-loss bug and a naming preference as equals. Pre-existing is the level people forget: without it, old bugs show up as blockers on a change that didn't cause them. A scale only means something if it fits your repo, and the docs note that Code Review's default calibration targets production code, so a docs or prototype repo may want a much narrower Important. Effort moves the balance too. At low, /code-review reports the findings it's most confident in, so you see fewer false positives. From medium through max, it broadens coverage.

Example

Suppose a review of a checkout change returns one Important finding, a null cart that crashes the total, with an empty cart as the input. It also returns one Pre-existing rounding bug in a neighboring function, and two nits. You fix the first before merging, file a ticket for the second, and decide on the nits. On GitHub, the service lays findings out the same way in the Details of its **Claude Code Review** check run. The docs' sample row reads Important, src/auth/session.ts:142, and a one-line description of the issue.

Going further

You can gate a merge on the service's result, even though its check run never concludes as a failure. The last line of the check's Details text holds the counts as JSON: <!-- bughunter-severity: {"nit":1,"normal":2,"pre_existing":0} -->. The normal key is the Important count. Read it with gh api repos/OWNER/REPO/check-runs/CHECK_RUN_ID, and fail your job when it's above zero. Treat a missing line as no verdict, not as zero findings, because a review that failed, timed out or was superseded doesn't have one. A title starting Code review not repeated means earlier Important findings are still open and there's no count to read.

Common mistakes

  • Using the labels without defining them for your repo, so style complaints land at the top.
  • Blocking a pull request over a Pre-existing finding. It's real, but it needs its own ticket.
  • Expecting /code-review to follow severity rules in REVIEW.md. It doesn't read that file. Local rules go in CLAUDE.md.
  • Replying to a wrong finding on the pull request and expecting it to clear. The docs say replying alone doesn't dismiss it, so resolve the thread.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course