Build a golden test set for your AI workflow

A golden test set is a maintained collection of representative tasks with agreed expectations. It gives a team a stable reference when changing an AI workflow. The expected result may be an exact field, a permitted action or a human-reviewed quality standard, depending on the task.

Choose examples for coverage

Imagine an assistant classifying incoming support requests. A useful set covers routine questions, mixed requests, missing account information and messages that require a human. Ten nearly identical easy questions provide less insight than a smaller but varied collection.

Document why each case exists. That makes it possible to retire redundant cases without accidentally losing coverage of an important failure mode.

Give each case a record

• Input: the request and the sources the system may use.
• Expected behavior: answer, action, clarification or refusal.
• Evidence: what a reviewer checks to determine correctness.
• Risk: the consequence if this case fails.
• Origin: synthetic example or authorized source, with sensitive details removed.

Do not write expectations by copying the agent's latest answer. Have the process owner confirm what should happen independently.

Protect against test leakage

Keep a development set for iteration and a held-out set for release checks. If every failure is immediately added to the prompt, the score may improve without proving broader reliability. Add new representative cases over time and inspect whether old cases still reflect current policy.

Version policy-dependent expectations. A test can become wrong when the business rule changes.

Use the set in ordinary maintenance

Run it after meaningful changes to instructions, models, retrieval or tools. Compare individual regressions, not only the overall pass rate. A gain on easy questions does not compensate for a new unauthorized-send failure. The test set should support a release decision and tell the team where to investigate next.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.