Evaluate and improve AI prompts

Test prompts against realistic inputs, inspect failure modes, and keep the revisions that improve the work.

A prompt is not finished when it produces one good answer. Test it against realistic inputs and define observable success criteria.

Try incomplete context, competing constraints, sensitive information, and different audiences to expose failure modes before reuse.

Keep a small evaluation set, fix the highest-impact ambiguity first, and record the strongest version for the next workflow.