Debugging With AI: Reproduce the Bug Before Requesting a Fix

If I ask “fix this bug” without a reproduction, the model may change code that was not broken. First I make the failure observable. After reproducing the bug, Use AI to Draft Tests, Then Challenge the Tests shows how to draft tests that would fail on the broken implementation.

For a small scoring error, I collect four items: Expected: a 50-second finish gives 0 bonus points. Actual: the UI shows -5 points. Reproduction: enter 50 in the supplied test page and click Calculate. Relevant code: the score function and the display line, not the whole project. Now my prompt can ask for diagnosis: From the supplied code and reproduction, identify the expression that permits a negative score. Propose the smallest fix. Add a test for 0, 50, and 80 seconds. State what you cannot verify from this snippet. I run the reproduction before and after the change. If the negative score persists, the diagnosis was incomplete.

Sometimes I ask for likely causes first, but I do not accept a cause without evidence. The model should point to the code path or suggest a test that distinguishes causes. “It is probably a race condition” is not useful if no timing evidence exists.

A fix that passes the failing case may break the normal case. Keep a tiny regression set: one failing input, one typical input, and one boundary input. Ask the model to change only the relevant code slice and explain why it should leave other behavior unchanged. Try it: Before prompting for a fix, write expected/actual/reproduction/relevant code. If you cannot reproduce the bug, ask AI to help design a reproduction step—not to rewrite the feature. Keep learning: Use AI to Draft Tests, Then Challenge the Tests, A Better Code Review Prompt Uses a Diff and a Risk Checklist, and A Tiny Evaluation Set for Your Favorite Coding Prompt.