Product · Prompts

Production prompts, version-controlled and critiqued.

Author, version, and ship the prompts your AI runs on. Six-dimension critique on every revision. Two-version Compare with a model-graded diff. Run one prompt across Claude, GPT, and Gemini side by side. BYOK at every tier.

I.Six capabilities, one workbench

What ships with every workspace.

Free tier ships every AI instrument. No demo call, no credit card. The cost lever is the 250 monthly AI call cap. Upgrade for unlimited calls and extended version history; the capabilities themselves are uniform.

01Six-dimension critique

Every prompt scores against Clarity, Completeness, Structure, Technique Usage, Robustness, and Efficiency. A concrete scorecard, not hand-wavy AI review. Run it on demand from the editor.

02Two-version Compare

Model-graded structural diff between two revisions of the same prompt. Answers the question no other workbench answers: did the rewrite actually improve clarity, or just shuffle words.

03Playground · Compare models

Run one prompt across Claude, GPT, and Gemini side by side. Per-model output, latency, tokens, and cost. Optional LLM-as-Judge step picks the best per criterion. Save as a public Model Report.

04Version history with diff

Snapshot-based versioning. Diff, restore, branch, annotate. Git-for-prompts without pushing to git. Solo keeps full history; Team keeps it shared.

05Eval suites with judge models

Test cases, rubrics, and LLM-as-a-judge graders in the same workbench. Catches regressions before the next provider update breaks production.

06AI pair in the editor

Six instruments next to you: Brainstorm, Critique, Improve, Rewrite, Compare, Convert. The pair drafts; you ship. Workbench Model honored across every action. Convert turns a plain prompt into a complete Agent Skill bundle in one shot.

II.Author to ship

The day-in-the-life of a prompt.

Six stations. The same workbench drives all of them. You move through them at your own pace; the AI pair sits next to you the whole time.

  1. 01
    Author

    Open the workbench. Author the prompt with the AI pair next to you. Brainstorm if you're starting cold.

  2. 02
    Critique

    Score against six dimensions. Read the per-dimension feedback. Improve targets the weakest one first.

  3. 03
    Improve

    Apply targeted edits one by one, accept the full Rewrite, or use Compare to model-grade two revisions of the same prompt.

  4. 04
    Run

    Open Playground · Compare models. Stream the same prompt across Claude, GPT, and Gemini side by side. Save the Model Report for a shareable artifact.

  5. 05
    Assay

    Promote winning cases into an evaluation suite. LLM-as-Judge graders run on every save so a regression surfaces before it ships.

  6. 06
    Ship

    Pull via REST API or the TypeScript SDK. Version is pinned; deployments roll forward, not over.

III.Your keys, your bill

BYOK at every tier. Including free.

Your Anthropic, OpenAI, and Google API keys connect directly. Prompt Assay never proxies or marks up inference. The bill stays with your provider. There is no markup on a single token at any tier.

New accounts receive a 7-call platform-funded demo budget on first signup so you can run Critique, Improve, and Rewrite before connecting a key. Adding any BYOK key permanently retires the demo budget. See our trust page for the data-handling and provider-isolation details.

IV.Common questions

Five questions we get every week.

How is this different from prompt management in Notion or a Google Doc?
Notion has no critique, no diff, no eval, no Compare across models, and no version pinning for production. Prompt Assay scores every prompt against six dimensions, runs LLM-as-a-judge graders against test cases, and lets you compare two revisions side by side with a model-graded judgment. Your raw strings move to a workbench that treats them like code.
Does Prompt Assay see my prompts or my inference traffic?
Your prompts are stored in your workspace; your provider keys (Anthropic, OpenAI, Google) connect directly via BYOK. Prompt Assay never proxies inference. The bill stays with your provider. There is no markup on a single token at any tier, including Free.
Which models can I author and test prompts against?
Anthropic Claude (Opus 4.7, Sonnet 4.6, Haiku 4.5), OpenAI (GPT-5 family), and Google Gemini (2.5 family). Bring your own keys for each provider; pick the Workbench Model that drives the AI pair and the Playground models independently for any given run.
What happens to my prompts if I downgrade or cancel?
Your prompts stay in the workspace on every tier, including Free. Evaluation suites are open on every tier too; the Free-tier lever is the 250 monthly AI calls cap. Paid-only differences: extended version history beyond the 7-day Free view, and public API access. Nothing is deleted on downgrade.
Can I use Prompt Assay if my team is on Humanloop or LangSmith today?
Yes. See /alternatives/humanloop and /migrate/humanloop for the Humanloop sunset path (deadline-driven), or /alternatives/langsmith and /migrate/langsmith for the LangSmith pricing path. Both ship a concrete migration guide that gets you from export to a live workspace in a working afternoon.

Author your first prompt in five minutes.

Free to start. No credit card. Bring your own keys when you’re ready. Or run a few calls on the platform demo budget first.

Also authoring Agent Skills? See the Skills surface