Burning a full 20x max plan a day on code reviews - Need optimization help
retained text
We run about 6-10 claude code chats in parallel and our code review process is not only time consuming, but it's unsustainably blowing through a full (or more) 20x max plan a day if I let it. Curious what people are doing for code reviews that's more efficient. How it works right now: \- every change is a branch + PR, nothing goes straight to main \- local checks run before every push, then GitHub CI runs typecheck, tests, build, dependency audit, secret scan \- PRs merge through a merge queue only when CI is green AND an "independent review" check passes \- the review is AI. two separate PCs pull PRs off a queue and review them on different Max accounts, so the chat that wrote the code never reviews its own work \- how hard a PR gets reviewed depends on which files it touches, and anything touching auth, payments, personal data, dependencies, security headers or prod config gets bumped up: \- heavy: Opus (high effort) and Codex both review, then Opus verifies each finding \- medium: one Opus finder + Opus verifier \- light: docs etc, no AI review \- a verified critical finding blocks the merge. a verified high blocks if it's security, money, data loss etc. everything else goes to the backlog \- when a chat fixes a blocking finding, a smaller "fix check" reviews just that fix \- every reviewer reads our big [CLAUDE.md](http://CLAUDE.md) rules doc first The problem: it's eating 1-2 full 20x Max plans a day just on reviews. Last 8 days: \- 130 PRs, 452 review passes, 463 fix checks, \~153 hours of review time \- heavy reviews were 61% of passes and 69% of the time \- 37% of passes were re-reviews of a PR that had already been reviewed \- 142 passes were on PRs touching 3 files or fewer What we're considering: \- one review per PR at the very end, no reviewing drafts, no re-review after a clean pass \- Sonnet finds bugs, Opus only verifies (keep Opus + Codex for the really sensitive core code) \- one session per PR with a one-page checklist instead of the whole rules doc \- skip tests, generated files and visual stuff and let CI cover those Is switching reviews from Opus to Sonnet a good option? Anyone found Sonnet good enough as the bug finder? And any other advice on how do you keep the review → fix → re-review loop from eating usage. Would love to hear what people's go-to review processes are.
document fetched 2026-10-14 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1x5afg9/burning_a_full_20x_max_plan_a_day_on_code_reviews