Skip to content

Score your AI prompt, then fix what it forgot to say

A prompt is a piece of writing, and like any piece of writing it can be checked before it goes out. PromptCaret marks the parts a dependable prompt needs and yours left out, then hands back a version with them written in.

Free, unlimited, no account. The score runs in your browser as you type.

Author's copy1 of 7 parts present
write me a marketing email
  • Role / persona is missing
  • Context is missing
  • Constraints is missing
  • Output format is missing
  • Examples (few-shot) is missing
  • Specificity is missing

Already there: Task clarity

48/100Sparse

The seven parts are not our opinion. They are what turns up consistently in the published system prompts of the tools themselves: Cursor, v0, Devin and Claude.

Three pulls of the same job

The prompt from the top of the page, and then the same brief written twice more. Read down and watch the parts get added. Every number here is what the engine returns for the text beside it, the same engine that will read yours.

  1. 48Sparse

    What most people type. Six of the seven parts are missing or thin.

    write me a marketing email
  2. 71Partial

    A task and some context, but the model still picks the format and the voice.

    Write a marketing email for our new project management app aimed at startup founders. Keep it short and friendly.
  3. 100Complete

    All seven parts present. Nothing is left for the model to guess at.

    You are a senior B2B copywriter.
    
    Write a cold marketing email for Flowdesk, a project-management app.
    
    Context: the audience is founders at 5-20 person startups who currently run projects in spreadsheets. Our goal is to book a 15-minute demo.
    
    Constraints: maximum 120 words. Subject line under six words. One call to action. Do not use emojis, and do not invent customer names or statistics.
    
    Output format: return JSON with the keys subjectLine, body and cta.
    
    Example:
    Input: a time-tracking app for freelancers
    Output: { "subjectLine": "Your invoices, already written", "body": "...", "cta": "See a demo" }

Six doors, one rubric

Everything here reads a prompt the same way: the same seven parts, the same scorer. What changes is where you are when you arrive.

When you have a prompt to write today

When you would rather stop writing weak ones

Start with a prompt you already regret

The one from this week that came back wrong. Paste it on the desk and the marks appear as you type. No account, no card, nothing to cancel. Passes exist only to lift the daily cap on AI rewrites, they are one-time, and nothing auto-renews.