Check your AI prompt before you send it
Paste a prompt you actually use. The marks move as you type, and nothing leaves the browser until you press a button that says it will. Scoring is free, unlimited and needs no account.
Your prompt
Example prompt. What most people type. Six of the seven parts are missing or thin. Edit it and the marks move, or clear it and paste your own.
Scoring runs in your browser: instant, offline, unlimited, no account. The AI rewrite runs on the server against the prompt corpus: one a day without an account, 3 a day signed in, or unlimited with a pass.
Structure check
SparseMost parts are missing, so the model has to guess at the rest.
This counts which parts of a prompt are present. It can’t read whether what you wrote inside them is any good. That’s what the rewrite is for.
Marked in the margin
Open with a persona, e.g. "You are a senior backend engineer."† Cursor, Devin and Claude prompts all open by establishing an explicit role.
Front-load all context the model needs: inputs, audience, goal, and constraints, so it can answer in one shot.† Front-loading context removes the multi-turn 'context setting' you want to avoid.
State limits and do/don't rules, e.g. "Keep it under 200 words. Do not invent APIs."† v0 and Claude prompts heavily use explicit do/don't rules to stay on-scope.
Specify the exact shape, e.g. "Return JSON: { title, steps[] }" or "Use a markdown table.".† Production tools define a strict output contract so responses are parseable.
Add one input→output example to anchor tone, format, and edge handling.† Few-shot examples are the cheapest reliable accuracy boost.
Name the actual nouns: the tool, the audience, the numbers, the file you mean.† Ambiguity forces the model to guess, and guesses cause re-prompting.
Clean copy
SCAFFOLD[ROLE] You are a senior <domain> expert. <-- set the persona [TASK] write me a marketing email [CONTEXT] - Audience: <who this is for> - Inputs: <data/files/values the model needs> - Goal: <what success looks like> [CONSTRAINTS] - <hard limit, e.g. length / tech / scope> - Do not <thing to avoid> [OUTPUT FORMAT] Return the answer as <exact shape, e.g. a markdown table / JSON schema>. [EXAMPLE] Input: <sample in> Output: <sample out>
This is a scaffold: your own words dropped into the parts they were missing, with a blank where each one goes. writes those parts instead of labelling them, against patterns from the tool you picked above, and leaves a blank only where it would otherwise have to invent one of your facts.
Publishing gives this prompt its own public page you can send to someone. Your name and email are never shown on it.
The teardown
One real prompt pulled apart: what scored badly, and what fixed it. No schedule, no filler, and nothing else in your inbox.
No sharing, no selling, and an unsubscribe link on anything we do send.
Nothing filed yet. Analyse a prompt and choose “Save to library” to keep it here.
What it scores
Every prompt is measured against seven dimensions. The percentage is how much each one moves the overall score.
- Role / persona12%
- Who the model should be. Naming a persona sets the vocabulary, the depth, and the default assumptions for everything that follows.
- Task clarity22%
- What you actually want done, stated as a clear instruction rather than a topic. The heaviest-weighted dimension.
- Context20%
- The facts the model needs but cannot infer: the audience, the inputs, what already exists, and what counts as success.
- Constraints14%
- The boundaries: length, tone, what to avoid, what must be included. Constraints are what keep an answer on scope.
- Output format16%
- The exact shape the answer should take. Specifying JSON, a table, or a word count removes the most common reason a response gets thrown away.
- Examples (few-shot)8%
- One input-output pair, which anchors tone, format, and edge handling far more efficiently than describing them.
- Specificity8%
- Concrete nouns and numbers instead of vague qualifiers. "Under 120 words" is scoreable; "keep it short" is not.
The Academy takes one dimension per lesson and ends each with a sandbox. The templates skip the reading and hand you all seven already in place. And the daily drill gives one brief a day, scored against everyone else who wrote to it.
Common questions
What does PromptCaret cost?
Scoring is free and unlimited, with no account required. AI rewrites are metered: one a day with no account, 3 a day once you sign in. To lift that cap you buy a one-time pass: Day pass, Week pass or Month pass. Passes expire on their own; nothing auto-renews and there is no subscription to cancel. Pricing is shown in rupees inside India and dollars elsewhere. See passes.
How is the score calculated?
By a transparent heuristic, not a model judgement. Each of the seven dimensions is scored 0–100 by pattern checks and combined as a weighted mean, using the weights listed above. Because it inspects structure rather than meaning, a high score means a prompt covers the things strong prompts usually cover, not that its content is correct.
Does my prompt leave the browser?
The score that moves as you type does not. That runs entirely in your browser with no network request, which is why it is instant and unlimited. Pressing Analyse does send it: the prompt and its score are saved so you get a history, and so we can see which parts people leave out most often. The AI rewrite goes further still, to the Anthropic API. The privacy policy spells out all three.
What is the difference between the suggested rewrite and the AI rewrite?
The suggested rewrite is a template assembled from your own text plus the dimensions you are missing, instant and offline. The AI rewrite sends the prompt to a model along with matching patterns retrieved from the corpus, and returns a version written for the specific agent you selected.
Which AI tools does it target?
12 of them, including Cursor, Claude, ChatGPT, Gemini, v0, Devin, Windsurf, Replit Agent and Perplexity. Choosing a target changes which corpus patterns are retrieved, so the rewrite matches that tool's conventions.