A reproducible symptom, expected behavior, current diff, and saved baseline.
One confirmed cause, one focused repair, and a regression check that fails before the fix and passes after.
Demonstrate the failing check before the repair and the same check passing afterward; inspect the focused diff.
Planned: coming to the library.
When an AI-built app breaks, it is tempting to let the agent keep guessing, and each guess can bury the real cause or overwrite good work. This guides your coding agent through reproduction, diagnosis, a minimal repair and regression verification, in that order.
You start by saving a checkpoint of the work you want to keep and recording the current revision. You then reproduce the failure from a known starting state, collect evidence that separates causes, and ask for a diagnosis with a check that could disprove it before any files change.
The repair stays small and targets the confirmed cause. You repeat the original steps, check nearby behavior, and keep a regression check that tells the broken and fixed versions apart. The worked example is a task that appears to save but disappears after refresh, traced to a missing write.
You need a coding agent with access to your project files. Provider accounts, model weights, GPU compute and hosting are not included.
No reviews yet.
Sign in to rate this Vibe.
Go live on Vercel or Railway with the right settings, a checked live journey, and a known way back if the release goes wrong.
Inspect dependencies, run one known workflow, and preserve a recipe you can repeat.
Model tasks, apply a development migration, and verify fresh reads after restart.