System Prompt Testing
System prompt testing verifies that a model's system-level instructions — its persona, rules, and constraints, set by the application rather than the user — are actually being followed under normal use and are resilient against a user's conversational attempts to override or ignore them.
A system prompt is only as good as the model's willingness to keep respecting it as a conversation gets longer and more adversarial. Testing usually splits into two tracks: functional testing (does the model actually behave like the persona/rules describe under normal, well-intentioned use — the right tone, the right scope, refusing out-of-scope requests correctly) and robustness testing (does that behavior survive a user pushing back, role-playing, or trying to negotiate the model out of its instructions over several turns).
A common blind spot is testing the system prompt only in isolation, with a single clean turn, when real degradation tends to show up several turns into a conversation, as accumulated context starts to compete with or dilute the original instructions. Multi-turn test scripts that gradually apply pressure are more representative than single-shot checks.
System prompt tests should be re-run on every model version bump too, since the same system prompt can be followed strictly by one model version and loosely by the next, even with no code change on the application side.
