Clean up AI-generated code.

The patterns we see in AI-generated codebases

AI assistants optimise for “something that compiles and runs the happy path.” What they rarely optimise for is the codebase that will still be readable in three months. In practice we see a common set of failure patterns across Cursor, Claude Code, Copilot, Lovable, Bolt, and Replit output.

Duplicated logic across files. The same data fetch, the same date formatter, the same form validator gets re-generated two, three, or four times because each prompt starts without awareness of what already exists. We consolidate into shared utilities and remove dead forks.

any instead of real types. TypeScript becomes a formality. Bugs that the compiler should have caught at build time ship to production. We retype API responses, Supabase query results, and Zod schemas into a single source of truth.

Inconsistent error handling. Some routes throw, others return { error }, others swallow silently. We establish a single pattern, wire it through the data layer, and surface failures to the user instead of hiding them.

No test coverage on revenue paths. Stripe webhooks, login, signup, core create/update flows. We add integration tests against the flows that actually matter; related deep-dives: Stripe webhook not firing and app works locally, not in production.

What our scoped engagement delivers

A refactor plan up front, agreed with you. Consolidated utilities. Typed data layer. A lint/format pass in CI so drift can’t re-enter. A small architecture document so the next developer (or the next Cursor session) doesn’t repeat the same mistakes. Pricing and timeline are on the pricing page.