Early access opening soon

Ship prompts like you ship code.

Every prompt edit, automatically tested against your evals. PromptProof catches regressions before your users do - pass/fail diffs, score history, and merge-blocking alerts.

promptproof / checks  ·  PR #142 "tighten refund-policy prompt"
✗ 3 regressions detected 47 evals · 2.4s · support-agent v14 → v15
PROMPT DIFF
  You are a support agent for Driftlane. - Always offer a refund if the customer seems upset. + Offer refunds only when the order qualifies under policy §4. + If it does not qualify, offer store credit instead.
REGRESSIONS
angry-but-ineligiblenow refuses a valid store-credit offer 0.94→0.41
chargeback-threatrequired escalation step missing 0.91→0.38
partial-refund-eligiblemisreads §4 thresholds 0.89→0.44
44 passed · 3 failed · suite score 0.87 → 0.79 Block merge

BUILT FOR TEAMS WHO HAVE BROKEN PROD WITH A PROMPT EDIT

DriftlaneQuillworksParallax LabsHearthstackMooring AI
How it works

Three steps. Zero instrumentation.

No SDK to wire through your stack. PromptProof sits next to your prompts, not inside your runtime.

01

Connect your prompts

Point PromptProof at the repo, folder, or platform where your prompts live. We track every version from then on. promptproof init and you're done.

02

Attach your evals

Bring golden test cases, or let PromptProof draft a starter suite from your real traffic patterns. Each case states what good output looks like.

03

Every edit gets tested

Each prompt change runs the full suite automatically. Pass and it merges. Regress and it blocks, with a diff showing exactly which cases broke and why.

Features

Everything code review has.
Now for prompts.

🛡

Regression gates

Block merges when eval scores drop below your threshold. The "fixed one case, broke three" era ends here.

📊

Score history

Every prompt version scored and charted over time. See exactly when quality drifted and which edit caused it.

🧬

Side-by-side diffs

Prompt diff and output diff in one view. What changed in the words, and what changed in the behavior.

🚀

Git-native

Runs as a check on your pull requests. Your prompts already live in git - your prompt testing should too.

🔭

Model drift watch

Re-runs your suite on a schedule, so silent model updates that change behavior get caught even when you change nothing.

⚡

Starter evals from traffic

No eval suite yet? Generate one from anonymized samples of your real usage in minutes, then refine.

"We edited one line to fix a tone complaint. It quietly broke refund handling for three weeks. PromptProof would have caught it in 2.4 seconds."
Design-partner interview, AI platform team (12 engineers)
FAQ

Fair questions.

How is this different from Braintrust or PromptLayer?

Those are excellent observability and eval platforms that want to live inside your runtime. PromptProof is deliberately smaller: it is the CI gate for prompt changes. No SDK, no logging proxy, no instrumentation. If you can open a pull request, you can use it.

I don't have an eval suite. Is this useless to me?

The opposite. PromptProof drafts a starter suite from samples of real usage you provide, you approve or edit each case, and you have a working regression gate in under an hour.

Which models and frameworks does it support?

Anything with an API. Prompts are tested against the model and parameters you specify per environment, so your staging and production configs are tested separately.

What does it cost?

Free during early access. After launch: a free tier for solo builders, team pricing per seat. Waitlist members keep founder pricing for life.

Early access

Your prompts deserve a test suite.

Join the waitlist and stop reviewing prompt changes on vibes.