Guide
Five AI workflows that help developers without replacing verification
The best coding prompts reduce uncertainty, expose assumptions and end in a testable next step.
By Cernalo · · 9 minute read
1. Turn an error into a diagnostic tree
Provide the exact error, the expected behavior and the smallest relevant code path. Ask for likely causes ordered by evidence you can collect, not a single confident fix.
Then test the cheapest discriminating check first. This prevents an assistant from turning a familiar error message into an unrelated rewrite.
2. Explain a narrow code path
Ask the assistant to trace inputs, branches, side effects and failure handling for one function. Request a short statement of anything it cannot infer from the visible code.
This is especially useful when reading unfamiliar code because it gives you a hypothesis to compare with tests and callers.
3. Compare designs against explicit constraints
Name the actual constraints—traffic shape, consistency, cost, operations or latency—before asking for a comparison. A design is not generally “best”; it is appropriate for a set of trade-offs.
Ask for failure modes and the evidence that would change the recommendation.
4. Build a meaningful test list
Share the behavior change and ask for tests that distinguish success from the important regressions. Reject suggestions that merely mirror implementation details.
Include boundary values, authorization behavior and failure recovery when they matter to the feature.
5. Translate a technical result
Use the assistant to turn a verified technical result into a release note, incident update or review summary. Supply the audience and facts, and instruct it not to add metrics or causes you have not confirmed.
Before sending, compare the final draft with the diff, logs or test output that support it.