Prompt Grader
Paste a prompt. Get back a critique that points at the actual weak words, and a stronger rewrite you can use straight away.
Curious how this was built? The build journal covers the shape, the wrong turns, the system prompt, what running it against the site's own 390-prompt library found, and what it cost.
Most prompts could be sharper. The difference between a vague prompt and a precise one is the difference between an AI that gives you a vanilla answer and one that gives you exactly what you asked for. The trouble is, when you write a prompt, you can't always see what's vague about it. Your own intentions are obvious to you. The AI cannot read them.
Paste a prompt below. The grader will quote the actual words back to you, point at what's vague or contradictory or missing, and offer a stronger rewrite. If your prompt is already strong, it'll say so honestly rather than invent problems.
Want to see how it works first? Try one of these:
You've used your ten grades for today.
Come back tomorrow, or use the "Or build a prompt..." button above to grade your prompt in ChatGPT, Claude, or Gemini directly.
Copy this to grade your prompt elsewhere
Paste this whole block into ChatGPT, Claude, or Gemini. They will return the same four-section critique. No daily limit when you do it this way; the AI is doing the work, not this site.
How this works (the model, the cost, the actual system prompt)
What's running under the hood
When you click "Grade it now", the page sends your prompt to a small Netlify Function that calls Anthropic's Claude Sonnet 4.6. The model reads your prompt, runs through a four-step critique (verdict, weaknesses, rewrite, what changed), and the result lands back on the page as structured markdown.
Why Sonnet, not Haiku
The site's chat widget and Persona Explorer use Claude Haiku, which is fast and cheap and absolutely fine for short conversational answers. Critique is a different job. Spotting a contradiction between two instructions, or recognising that a phrase is doing no work, requires the more capable model. Haiku will sometimes miss the subtleties.
The daily limit, and why it exists
Each grade costs roughly $0.005 AUD against my Anthropic API account. To keep the bill predictable, the function caps each visitor at 10 grades per day (counted by your IP address, reset at UTC midnight). If you hit the cap, the "Or build a prompt..." button still works without limit, because that route doesn't call this site at all. The system prompt and the four-section structure are the same either way.
Why a Netlify Function, not an Edge Function
An Edge Function would have run closer to you and been very slightly faster, but Netlify's edge proxy times non-streaming responses out at 30 seconds and the streaming budget on the standard wrapper is only 10. Sonnet at the token budget we use here finishes comfortably inside 30 seconds, so a normal serverless Lambda is the right shape. It also makes rate-limiting via Netlify's blob store cleaner.
The actual system prompt
This is the verbatim text the function sends to Sonnet. It's the same prompt you can copy into ChatGPT or Claude yourself if you want to run the grader without the daily limit:
What it doesn't store
The prompt you paste is sent to Anthropic for grading and to the function for rate-limiting. The function stores only your IP address and a count of grades used today, in a Netlify blob, expired daily. The text of your prompt is not logged on this site. Anthropic's data-handling policy applies for what they do with the message during grading; see anthropic.com/privacy.
Want to read how this was built?
The full build write-up, including the wrong turns and what surfaced when I ran the grader against the site's own prompts, is at How I Built the Prompt Grader.