How Many Test Cases a Prompt Evaluation Needs

A prompt can look better on a handful of examples by chance. Plan from the decision: what metric matters, what smallest improvement would justify a change, how variable are outcomes, and how much uncertainty can you tolerate? Statistical power, significance level, baseline performance and sampling design affect the needed number. NIST guidance for proportion tests makes these dependencies explicit; it does not give one universal count for prompt evaluation. For binary pass/fail outcomes, report numerator, denominator and an appropriate confidence interval; Wilson intervals are commonly recommended for proportions. To compare versions, run them on the same representative cases and analyze paired outcomes. Repeated variants of one source or turns from one conversation may be correlated. Repeated model generations can measure run-to-run variation, but do not replace diversity in user inputs. Define the metric, sampling frame, strata, grading rules and decision threshold before testing. Include ordinary and edge cases, and keep a separate holdout if tuning is extensive. Oversampling rare safety failures is useful for discovery, but the raw rate is not prevalence unless weighted to the target distribution. Report uncertainty and assumptions rather than claiming a fixed count proves improvement. Independence matters: ten paraphrases of one source are not equivalent to ten unrelated tasks. If cases cluster by product, language or workflow, report groups and respect the design in analysis. Keep test cases distinct from prompt-tuning examples. Cases may share documents, users or workflows, so count independent sampling units rather than every generated variation as new evidence. Keep a separate holdout when tuning repeatedly, and distinguish repeated model runs from additional task coverage.