How to Debug AI-Generated Code — 3 Production Failure Patterns

#webdev #javascript #react #ai

I built a full inventory module with Cursor. Demo looked clean, client approved it. Then it started creating duplicate records on every save.

Root cause: the AI correctly wrote the upsert call, but never persisted the record UUID back to state. Every save treated the form as a new record.

After debugging multiple of these in production ERP and CRM work, I've found AI-generated code breaks in three predictable patterns — and each has a different debugging approach.

Pattern 1: Structural violations

AI puts logic in the wrong place. Functions in the parent component that should be in a helper file. Imports that violate module boundaries. Naming conventions that don't match the rest of the codebase.

Why it happens: if your steering file is long or the conversation context is large, the model deprioritizes your structural rules.

Fix: Re-state critical rules directly in the prompt — "write these functions in /helpers/moduleName.js, not in the parent component." Redundant, but it works.

Pattern 2: Demo-only functionality

The code works on the first run. It fails on the second interaction, with edge case inputs, or after a cancel-and-retry sequence.

AI optimizes for the happy path: one user, clean data, first interaction only. It doesn't model what the component state looks like before the operation or what happens when a user does something twice.

Fix: Run every user flow at least twice before accepting the code. Most AI state bugs are invisible on run 1.

Pattern 3: Silent state bugs

A file upload component with a mandatory comment dialog. First cancel: correctly blocks the upload. Second cancel: upload proceeds anyway.

The AI had set a boolean flag to block on cancel — but never reset it between attempts. Second attempt read the stale value.

Fix: Trace backwards from the symptom, not forwards from the trigger. Search for useState(false), useRef(), and any ID/UUID variables — that's where missing resets live.

The pre-delivery checklist

Before accepting any AI-generated module: run the flow twice, check API calls in the network tab (create vs update?), verify all boolean flags reset between interactions, and audit file placement before reading a single line of logic.

Takes 15 minutes. Catches the bugs that will take hours to debug after the fact.

Full breakdown with detailed checklist originally published at StackNova.