← sources

My plan code review loop after 6 months with Claude Code

source: reddit (reddit:ClaudeCode)published 2026-10-21 05:46ingested 2026-10-22 02:42

retained text

I've run the same loop every day since April on a team of 4, mostly a TypeScript backend and a React app. Around 600 PRs so far. Writing down what stuck because my own notes were a mess.



What actually works



\- Write the goal in 2 lines before anything else. What and why. Half my bad sessions were me skipping this.

\- Plan to the file level before any code. "Change X in the auth service" is useless, "add an expiresAt check in src/auth/session.ts and update the 3 callers" is a plan.

\- Point at files, don't paste them. Paths and line numbers. It gets worse the more you dump in, even with the big context.

\- Review twice. Once yourself on the plan, once on the diff with a reviewer that didn't write it.



What hurts



\- Vague prompts. "Fix the login bug" gets you a rewrite of the whole login flow.

\- Letting it pick packages. It pulls in whatever it saw most in training, we ended up with 3 date libraries.

\- One giant session for a whole feature. New task, new session.

\- "It compiles" as the finish line. Our worst bug this year passed tsc, lint and the tests the agent wrote for it.



My setup



\- Plan: plan mode with Opus. I tried a couple of separate planning tools and kept coming back, it's good enough once you force file level steps.

\- Code: Sonnet for the edits. On 80% of tickets I can't tell the difference from Opus and it saves a lot of usage.

\- Review: this is where I changed the most. I started with Claude reviewing its own PR in the same session and it mostly agreed with itself. Then a month of Cursor Bugbot, good on obvious bugs but it missed anything that needed context from other files. Ended up with Coderabbit on every PR plus its CLI before I push. It catches the cross file stuff and the summary is what my lead actually reads. Downsides, it's chatty on small PRs until you tune it, and it's $24 a seat, which my company pays so



Hot take (not sure if its even hot take anymore)



The agent isn't the bottleneck anymore, the guardrails are. Almost all our gains came from planning and review, the coding part was the easy bit.



What does your loop look like? Especially curious if anyone found a planning setup that beats plain plan mode.

document fetched 2026-10-22 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1xbeymx/my_plan_code_review_loop_after_6_months_with