Define the task
Write the intended user, input range, prohibited behavior, and success criteria before editing the prompt. Without a target, prompt improvement becomes taste.
Build an evaluation set
Include ordinary cases, edge cases, incomplete inputs, and known failures. Keep it stable so versions can be compared.
Use multiple measures
Score correctness, source support, completeness, instruction following, format validity, and actionability. Some checks can be automated; subjective criteria need a clear rubric.
Record the environment
Save prompt version, model, model settings, tool access, test date, and representative failures. Retest after important provider changes.
Worked example
Evaluation case: Extract action items from meeting notes.
Checks: every item appears in notes; owner/date are not guessed; output has the required columns.
Failure cases: no action items, conflicting dates, missing owner.
Record: prompt version, model ID, settings, input, output and check results.How to check the result
Do not use only easy cases or the examples included in the prompt. Repeat when randomness matters. Decide what constitutes failure before reading the output, and keep a record of regressions.