TypeScript refactor for AI apps — remove any types, ship a type-safe refactor of AI-generated code from Lovable, Bolt, v0 and Cursor

A fintech founder on Lovable described the loop exactly the way Medium's Nadia Okafor did in Vibe Coding in 2026: 'the filter worked, but the table stopped loading. I asked it to fix the table, and the filter disappeared.' Every week, the AI broke a working feature. Every week, the founder spent a day and dozens of credits getting it working again. We opened the repo and the reason was unambiguous: strict mode was off, tsconfig was barely configured, and 340 instances of `: any` sat in the payment flow alone. Every function signature was a suggestion.

The concrete cascade looked like this. Cursor was asked to add a new status ('refunded') to the order type. It added the literal to one interface in one file. It didn't update the discriminated union in the second file. It didn't update the switch statement in the third file, because the switch's default branch was `(status: any) => ...`, which happily swallowed the new value. The payment confirmation email went out for refunded orders because the condition `if (status !== 'cancelled')` was true. Three customers got refunded and also charged a reminder invoice. No compiler error at any stage. The regression wasn't found by tests because there were no tests on that path.

The refactor took three weeks. We flipped strict mode on at the tsconfig level, then fixed the resulting 1,200 errors domain by domain — API layer first, then forms, then components. Every PR stayed green in CI. We introduced Zod at every IO boundary (Stripe webhooks, Supabase reads, form inputs, env vars) and derived types from the schemas so the types could never drift from the runtime checks. We shipped a .cursorrules file that tells the AI to respect strict mode. Six weeks later the founder logged zero 'the AI broke it' incidents. That was the number before refactor: four per week.

The durability of the fix matters more than the one-time cleanup. Strict TypeScript gives the AI builder a feedback loop it previously lacked. Before: Cursor renames a field, the caller still compiles because everything's `any`, the break surfaces at runtime days later. After: Cursor renames a field, the compiler lights up immediately, the AI sees the error in its own terminal and fixes the caller in the same session. The regression loop doesn't stop because the founder got better at prompting; it stops because the compiler is now talking back. Several clients have reported that their AI builder simply started producing better code after the refactor — same tool, same model, new contracts. Strict types are, in effect, the cheapest prompt engineering there is.

This is also the fastest wedge against the specific pain pattern Medium's Vibe Coding in 2026 documented: 'the filter worked, but the table stopped loading.' Without strict types, there's no mechanism to catch that a shared type was used inconsistently across components. With strict types plus Zod at IO boundaries, a shape change in one place forces explicit handling everywhere the shape flows. The AI can't silently break working features because the compiler won't let it silently do anything.