Give it evidence, not vibes

The single biggest lever on the quality of an AI debugging session is the evidence you supply. A prompt like fix my login bug gives the model nothing but room to guess. A prompt with the exact error, the stack trace, the relevant code, what you expected, what actually happened, and what changed recently gives it something real to reason about.

A durable structure to reuse:

Symptom: what the user or test sees. Expected: what should happen. Actual: what happens instead. Environment: language, framework, versions, where it runs. Recent changes: the last deploy or commit that touched this. Evidence: the error, the stack trace, the code path.

Notice what is missing from that list: your theory. Lead with facts, not with your hunch, or the model will politely agree with your hunch and reason from it. If you must share a suspicion, label it clearly as a suspicion and ask the model to argue against it as well as for it.

Vibes-based prompts produce vibes-based answers that merely sound authoritative. Evidence-based prompts produce answers you can check. The discipline is the same one good bug reports have always demanded; the AI just rewards it more visibly, because everything you leave out gets filled in with a guess.