Packs / Code review for Claude Code
Code review for Claude Code
Review commands you run with one word, a reviewer subagent that can only read, and standards that keep the AI honest about what it found. Part of Pro.
What you type
| Command | What it does |
|---|---|
/review-changes | A strict review of your uncommitted changes, ranked blocker / should fix / nit, ending in a one-line verdict. |
/review-branch main | Reviews everything on your branch that is not on main yet, like a pull request review, and checks the commit messages match the code. |
/security-pass | Follows untrusted input from where it enters to where it is used, category by category. |
/find-missing-tests | Finds behavior in your changes that no test covers and lists the cases to add. |
/respond-to-review | Works through review comments you received, decides whether each is right, and drafts replies for you to post. |
the code-reviewer subagent | An independent second opinion that can only read and search. It cannot edit files or run commands. |
We measured it
Reviews are easy to sell and hard to trust, so we tested this one. We planted bugs in a small change and asked Claude Code to review it, once with the /review-changes skill and once with a plain "review my uncommitted changes" prompt. We read the answers, not just the scores.
| Test | Model | Runs each | With the pack | Plain prompt |
|---|---|---|---|---|
| Obvious bugs: SQL injection, off-by-one, missing null check, hardcoded key | Haiku | 5 | 20 of 20 found | 20 of 20 found |
Subtle bugs: a changed boundary, a sort that now mutates its input, a shared default array, a missing await | Haiku | 8 | 32 of 32 found | 25 of 32 found |
| The same subtle bugs | Sonnet | 3 | 12 of 12 found | 12 of 12 found |
- It helps a weaker model on subtle bugs. Haiku with a plain prompt missed the changed boundary in 3 of 8 runs and the in-place sort in 4 of 8. With the skill it missed nothing.
- It makes no measurable difference for obvious bugs or with the stronger model. We are not going to tell you it makes a strong model smarter.
- The answers are consistent. All 16 answers from the skill had severity labels and a final verdict. None of the 16 plain answers did.
- No false alarms. We planted a harmless
Math.random()as bait. In every answer we read, the reviewer said it was fine. - The cost was about the same: roughly $0.04 per review on Haiku.
These are small samples from one test with one language, and our keyword scoring needed correcting once after we read the answers. It is evidence, not proof. The test, its data and every raw answer are in the pack so you can repeat it.
What it can't do
- An AI review is a second pair of eyes, not a replacement for a human one. Treat findings as leads to check.
- It reviews your uncommitted changes or a branch. It does not fetch pull requests from GitHub by itself. The optional GitHub server needs Docker, which we could not run, so we have not tested that part.
- It needs Claude Code's Bash tool, which it uses to run read-only
gitcommands. If you give/review-brancha branch that does not exist, or run it outside a git repository, Claude tells you what is wrong rather than failing silently. We tested both.
Not released yet. Pro isn't open, so there is nothing to buy today. Join the waitlist and we'll email you when it opens, at the founding price. A useful companion is the Safe setup pack, which stops the AI from pushing or deleting things.