Stop your AI assistant from guessing the fix

The symptom moved

You paste a stack trace into an AI assistant. It proposes a fix that looks reasonable. You apply it, the error goes away, and a different one appears somewhere else, or the same bug comes back next week. The bug was never fixed. The symptom moved.

This is the most common way AI-assisted debugging fails, and it has a simple cause: the model was allowed to propose a fix before anyone knew the cause. A plausible fix is easy to generate. A proven cause is not.

Forbid the guess

The fix is to split debugging into stages, each with an exit condition, and not move on until the condition is met.

Frame the bug. Write one sentence that states the defect in a way that could be proved false. "Checkout is broken" cannot be. "Submitting the form with a saved card returns a 500" can.

Reproduce it. Get a command or a sequence of steps that fails on demand. Until the bug reproduces, every theory is untestable.

Find the cause. Identify the exact line, and the evidence that it is the cause. This is where the model is most tempted to guess, so ask it directly: what evidence would prove this hypothesis wrong? If it cannot say, it is guessing.

Fix and verify. Write a test that fails before the change and passes after it. A fix without that test is a hope.

Stop the class. Add one guard, such as a check, a type or a lint rule, that would have caught this kind of bug earlier.

Why stages beat one big prompt

One prompt that asks for "the cause and the fix" lets the model jump to the end. Stages force the slow part, proving the cause, to happen first, and give you a natural place to stop when a stage cannot be completed. Skipping a stage is sometimes right, but it should be a decision, not a default.

Putting it into practice

The Debugging Prompt System is five prompts, one per stage, each with its exit condition: triage, reproduction, root cause, fix and verification, and a postmortem that adds one guard. It includes the common ways AI debugging goes wrong, usage notes on when to skip a stage, and a full session that starts with the wrong first hypothesis.

When the bug is in production and the logs do not answer the question, see debugging production without redeploying.