How to Score a ChatGPT Prompt (What a Good One Actually Looks Like)

Everyone can tell you a prompt is "bad." Almost nobody can tell you which part of it is bad — which is the only information that helps you fix it. So here's a scoring system, and a free tool that runs it in ten seconds.

Why "good prompt" needs a number

Two people look at the same prompt and disagree about whether it's any good. That argument never resolves, because "good" isn't measurable — but its four components are. Once you score them separately, the fix becomes obvious: your 78 isn't a vague B, it's a 25/25/25/3 with no output format.

That's the whole value of scoring. Not the number — the diagnosis.

The four dimensions, and what each score means

Role (0-25)

Does the prompt tell the AI who to be?

• 0-5: No role at all. The model answers as a generic assistant.
• 12: A domain is mentioned but no role is assigned ("marketing tips").
• 18: A role is assigned but it's junior or vague ("act as a marketer").
• 25: A specific, senior role with a track record ("act as a growth marketer who has scaled DTC brands from zero to $1M").

Context (0-25)

Does the prompt supply your actual situation?

• 0-5: One line, no details. The model has nothing to tailor to.
• 10-15: Some detail, but nothing about your audience, budget, or current state.
• 25: Real specifics — numbers, audience, what you've already tried — plus [bracketed placeholders] for details that change between runs.

Constraints (0-25)

Does it say what to prioritize and what to rule out?

• 0-5: No boundaries. The model wanders toward the safest general answer.
• 12: A vague preference ("keep it simple").
• 25: Explicit priorities and exclusions — "prioritize speed to first dollar; rule out anything needing paid ads or an existing audience."

Format (0-25)

Does it specify the shape of the output?

• 0-5: Nothing. Expect prose you'll have to reorganize yourself.
• 12: A hint ("give me a list").
• 25: A precise spec — "a numbered 5-step plan; each step with one concrete action, the tool to use, and how I'll know it's done."

Scoring a real prompt

Take the most common prompt on the internet:

give me some marketing ideas

Role 4, context 1, constraints 2, format 6. Total: 13/100. Not because it's badly written — it's barely written at all. The model fills the vacuum with the average of everything it has read about marketing.

Now the same request, engineered:

Act as a growth marketer who has scaled DTC brands from $0 to $1M. My product is a $49/mo meal-planning app for busy parents; budget $500; 5 hours a week; Instagram hasn't worked. Recommend 5 channels ranked by speed to first sale. Rule out anything requiring paid ads or an existing audience. Format as a numbered list — for each channel give the first action to take this week, the realistic 30-day result, and one free tool.

Role 25, context 25, constraints 25, format 18. Total: 93/100. Same model, same question, completely different answer.

What to do with your score

Under 35 — you wrote a question. Add a role and an output format; that alone usually moves it past 60.

35-70 — one dimension is dragging you down. It's almost always constraints or format, because those feel optional and aren't. Find the lowest number and fix only that.

70-85 — you're close. The usual gap is context: you've described the task well but not your situation. Add two sentences about your audience and what you've already tried.

Over 85 — stop optimizing the prompt and go run it.

The trap: long is not the same as engineered

A 400-word prompt with no role, no constraints, and no format spec still scores badly. Length correlates with score only because longer prompts tend to contain more of the four parts. Padding a prompt with adjectives moves nothing. Adding one line — "Format as a numbered list; each item with an action, a tool, and a done-check" — moves it 15 points.

Score, fix, re-score

The loop that actually improves your prompting: paste a prompt into the Grader, read which dimension scored lowest, add that one thing, re-score. Three rounds and you'll stop needing the tool — you'll be writing 85s from scratch because you've internalized what's missing.

If you'd rather skip the manual round-trip, the Prompt Improver rewrites a weak prompt into all four dimensions automatically and tells you what it changed and why.