Prove your prompts work. Ship changes without redeploying.

Catch

Assert what the answer must do

Point an evaluation at the prompt and assert what has to hold: length, required phrases, banned words, JSON validity, regex, an LLM rubric. You get a 0–1.0 score and a pass or fail per assertion, so “worse” becomes a number you can compare across runs. No paid tier required.

Undo (Pro)

Keep the version you replaced

Saving an edit records the version it replaced in template history. Restore any recorded version through the API — and the state you restored from is snapshotted too, so a rollback is itself reversible.

Watch (Pro)

Re-score published templates on a schedule

Opt-in production monitoring re-scores each published template on a schedule, with the same pinned judge on every run, against the 0.65 pass line. Below it, a scoped repair is queued for you to adopt — the live published row is never overwritten by the sweep.

11

Assertion types

Length limits, required phrases, JSON validity, regex, latency — plus LLM-rubric scores. A measurable pass rate, not a gut feeling.

5

Constraint rules

Deterministic checks for no bullets, no markdown, banned words, max length and required prefix. Partial scores show which rule broke.

0–1

Semantic drift

Cosine similarity between original and optimized embeddings flags when "improvement" changed what you were asking for.