Curated, not collected
A golden set is hand-curated examples with known-correct outputs. It's the smallest, most carefully chosen subset of your traffic, used to validate every change. 50 cases is plenty for most prompts; for high-stakes systems, 200–500.
Construction
- Source — sample real traffic stratified by intent / category.
- Reference outputs — produced by your best human reviewer or current best model.
- Annotations — why this output is correct; what an alternative correct output would look like.
- Versioning — golden sets evolve; tag versions and run new prompts against the version they were built on.
Maintenance
- Add new cases when production surfaces a new failure mode.
- Retire cases when the underlying behavior is deprecated.
- Re-review reference outputs annually — what was correct last year may be wrong now.
- Audit for representativeness — does the set still match traffic distribution?