Claude Code to production
Claude Code produces the most structured output of the agents we see — and the most of it. The risk isn't that the code is bad; it's that nobody has verified it, that it abstracted things that didn't need abstracting, and that the tests it wrote pass for reasons that aren't the ones you'd want.
Cleaning up a codebase built with Claude Code is mostly a matter of review and pruning: removing abstractions built for a single implementation, auditing the large multi-file changes that were merged without a human reading them, replacing tests that mock the code under test with tests that exercise it, deleting stubs and duplicated utilities left across sessions, and tightening the permissive defaults the agent chose to make something work. Gen2Prod audits the repo for free, fixes what matters at a fixed price from $290, and leaves a CLAUDE.md that keeps the agent inside the conventions afterwards.
The list we find on most Claude Code projects, in the order it matters. Each one links to the service that fixes it.
A repository pattern over one table, a service layer that forwards every call, a plugin system with one plugin. We remove the indirection that isn't earning its keep, so a change touches one file instead of four.
Fixed by Refactor & cleanupClaude Code writes tests readily, and when the real dependency is awkward it mocks it — sometimes so thoroughly that the test proves nothing. We identify those and replace them with tests against real behaviour on the paths that matter.
Fixed by Refactor & cleanupA 40-file change that 'refactored auth' was merged because it looked right. We review the merged history for what changed behaviour without anyone deciding it should, and fix or document each one.
Fixed by Security hardeningA CORS wildcard, a disabled RLS policy, a `catch` that swallows the error — each added to get past a failure and never revisited. We find them and tighten them.
Fixed by Security hardening`// TODO: implement` in a route that returns 200, a feature flag for a feature that doesn't exist, three `formatDate` helpers. We delete what's dead and finish or remove what's half-done.
Fixed by Refactor & cleanupGenerated READMEs and comments drift as sessions accumulate. We rewrite the ones that matter and delete the ones that lie, so the next developer — or agent — isn't misled.
Fixed by Handoff & docsA cleanup isn’t a rebuild. Claude Code did real work, and most of it stays.
Passing tests are necessary and not sufficient. The question is what they'd catch. A test that mocks the database, mocks the auth check and then asserts the mocked value came back will pass through any bug you could introduce. In the free audit we read the tests as carefully as the code and tell you which ones are load-bearing.
Different. Lovable and Bolt generate a stack and get security wrong in predictable ways; Claude Code edits your stack and gets structure slightly too right — more layers, more abstraction, more code than the problem needs — while leaving the same permissive defaults when something blocked it. The fix for the first is hardening; the fix for the second is review and pruning.
We read the merged history, not just the current state, because the risk is in the sweeping changes that were accepted without a human reading them. We look for behaviour that changed without a decision, tests that can't fail, abstractions with one caller, and security settings loosened to get past an error. Everything we find is rated and reproduced in the report you keep.
Yes — it's part of every Claude Code engagement. We write down the conventions we established: the architecture, how errors and auth are handled, what a test must assert, what not to abstract. The agent follows a clear CLAUDE.md closely, so the codebase keeps its shape as you keep building.
Start with the free audit. In 48 hours you'll know what's broken, what each fix costs, and whether it's even worth doing — before you've spent a dollar.
Free audit, no card · Fixed price before we start · Your code stays yours