AI Prompt Analyzer

How to read the score honestly

The 0-100 number is a diagnostic, not a grade. It measures how much guesswork your prompt leaves on the table — not whether the output will be any good. A prompt can score 92 and still produce something you hate, because the analyzer cannot know your taste or your brand.

What it does tell you is where the model is filling gaps for you. Rough bands:

• Below 40. The model is inventing the answer's shape. Expect generic output and rounds of "no, more like this".
• 40-65. The request is clear, the constraints are not. Right topic, wrong length or format.
• 65-85. Solid working prompt. Remaining points are format precision and examples.
• Above 85. Diminishing returns. Stop optimizing, start testing real output.

Chasing 100 is a trap. Past 85 you add instructions that cost tokens and squeeze the model into stiff, over-specified prose. A 78 you have run three times beats a 96 you have never sent.

The seven checks, and why each one matters

Each of the seven dimensions maps to a specific way prompts fail.

• Clarity. One unambiguous action? Prompts with two verbs joined by "and" get half-answers to both.
• Role. A perspective that changes vocabulary and depth. Missing roles cause flat, encyclopedia-style output.
• Context. Audience, product, prior attempts. Scores drop hardest here, and fixes pay most.
• Constraints. Length, exclusions, hard rules. Without them the model defaults to its own average length, which is never yours.
• Output format. Table, JSON, bullets. Unspecified means prose, every time.
• Specificity. Concrete nouns and numbers versus filler adjectives. "Improve engagement" scores badly; "raise reply rate on cold emails" scores well.
• Examples. One sample of the output beats three paragraphs describing it. Few-shot versus zero-shot covers when to bother.

Two low dimensions usually explain a bad score entirely. Fix those first.

What vague wording actually costs

Words like "engaging", "professional", "comprehensive" and "modern" feel like instructions but carry almost no information. The model resolves them toward the statistical middle of its training data. That is why so much AI writing sounds identical — everyone is asking for the same average.

The cost shows up three ways. You rewrite the output, burning a round trip. You get length you did not ask for, because "comprehensive" reads as "long". And you cannot iterate, because when the answer is wrong you cannot point at which instruction failed. Replace each vague word with something checkable:

• "engaging" → "opens with a specific number or a question"
• "professional" → "no contractions, no exclamation marks"
• "comprehensive" → "covers these four points, 150 words each"
• "modern" → "short sentences, no jargon"

The test: could two people read your instruction and disagree about whether the output followed it? If yes, it is not a constraint.

Fixing a low score, and looping back

In order — the first two recover most of the points.

1. Add context. Two or three sentences the model could not guess: who reads this, what it is for, what already failed.
2. Name the format. "Output: markdown table, three columns" beats a paragraph of description.
3. Add a role that changes something. If swapping it changes nothing, skip it.
4. Convert soft adjectives into rules. See above.
5. Add one example only if format matters more than content.

If the prompt is a mess rather than merely thin, do not patch it — feed it to the prompt optimizer to rebuild the structure, or the rewriter to tighten wording without changing intent. For structured output, the JSON prompt builder handles schema constraints better than prose.

Then loop. Analyse, apply the top two fixes only, re-analyse. Two at a time means you learn which change moved the number — change five and you learn nothing.

Watch the dimension breakdown rather than the total. A jump from 51 to 74 tells you less than seeing context go from 2/10 to 8/10, because the second reveals a habit. Most people have one chronic weakness that shows up in every prompt they write. Find yours and you stop needing the tool for it. Common prompting mistakes catalogues the usual suspects.

Stop when the score plateaus or crosses 80. Then test real output and keep the winner in the prompt library so you are not rewriting it next month.

A worked example

A prompt that scores in the low 30s:

Clarity is fine — it is an email. Everything else is missing: no role, no context, no length, no format, two adjectives that mean nothing. You get a polite announcement that could belong to any company on earth. The rewrite scores in the mid-80s:

Every added line closed a gap the model was filling on its own. For a specific model's quirks, the ChatGPT prompt generator and Claude prompt generator start you from the right shape, and how to write AI prompts explains each block.