Amey-Thakur Prompt testing
A prompt in production is code with non-deterministic behaviour, which makes testing harder and more necessary. Without it, a model upgrade or a small edit silently changes behaviour for every user.
Method
- Define pass criteria per case. Exact match where the output is structured, and a rubric or a judge where it is prose. Undefined criteria mean the test proves nothing.
- Run each case several times. Output varies between runs, so a single pass may be luck, and the pass rate matters more than a single result (see agent-ensemble-voting).
- Include adversarial and boundary cases. Empty input, very long input, contradictory instructions, and injection attempts (see llm-guardrails).
- Test on every model you deploy. Behaviour differs between providers and versions enough that one model's results do not transfer.
- Re-run before every model upgrade. A provider updating a model changes your product, and this is the only warning you will get.
- Track cost and latency alongside accuracy. A prompt that is marginally better and twice the cost is a trade-off to make deliberately.
- Automate it in the pipeline. A test suite that must be run manually is run rarely and never at the moment it matters (see agent-eval-design).
Boundaries
Testing shows behaviour on the tested cases; the long tail remains unknown. Judge-based scoring inherits the judge's biases and needs periodic human calibration. Non-determinism means thresholds rather than guarantees, and flaky tests need statistical handling (see test-flakiness-budget).