How good is your AI prompt?

What prompt scoring is

Prompt scoring measures how completely a prompt specifies what you want, before you send it — then names what is missing.

It checks for the things that measurably change an answer: a clear task, an audience, an output format, constraints, an example. It does not judge whether your idea is good. It judges whether a model has enough to act on, or whether it has to guess.

One prompt, scored at every step

Every score below is real. Paste any row into the checker above and you will get the same number — the scoring is deterministic, so it never varies.

Seven words became eighty-one, and the score went from 26 to 84. Nothing clever was added — each step answered a question the model would otherwise have answered for you, differently every time. Note where it stops: even with an example, this prompt does not reach 100, because it never asks the model to think before drafting. For a blog post that is the right call.

The eleven parameters

Weights reflect how much each one changes an answer in practice. Every check awards partial credit: mentioning a thing earns some of the points, doing it properly earns the rest.

Four patterns that reliably work

Beyond filling the eleven slots, a handful of techniques change results consistently enough to be worth learning by name.

Few-shot — show one example

Give one or two examples of the output you want before asking for a new one. It communicates tone, structure and level of detail faster than any description, and it is the technique people most often skip. Even a bad example marked "not like this" helps.

Chain of thought — ask for reasoning first

"Work through it step by step before answering." On multi-step problems — analysis, comparisons, anything with trade-offs — this reliably improves the conclusion. On a one-line creative task it adds noise, so it is not a universal upgrade.

Delimiters — fence your input

When a prompt contains material to work on — a draft, a transcript, data — wrap it in triple backticks or tags. Without a fence, models routinely confuse "the text I am giving you" with "the instructions I am following".

Role — but a specific one

"You are a senior tax accountant specialising in SaaS revenue recognition" shifts vocabulary and assumed knowledge. "You are a helpful assistant" does nothing at all: it is what the model already believes.

What actively hurts a prompt

Some habits do not merely fail to help — they make output worse.

• Contradicting yourself. "Be comprehensive but keep it under 200 words" forces the model to pick, and you cannot predict which it drops.
• Politeness padding. "I hope you can help me with something, if it is not too much trouble" costs input space and adds nothing. Be direct.
• Asking for citations. A model without browsing will invent plausible-looking sources. Ask it to name well-known works instead.
• Requesting charts or images from a text model. Ask for a markdown table.
• Page counts. A model has no page. Use word counts.
• Burying the ask. Three paragraphs of background before the actual request. Lead with the task, then give context.

Prompt frameworks, and when they help

A framework is a reusable slot structure — a checklist for the eleven parameters above, arranged for a particular kind of task. They matter because a blank page is the hardest place to start, not because the acronym is magic.

These six are a sample. Frompting's library holds 61 across eight domains, and the generator picks one for you based on what you have written.

When a framework does not help: when you already know exactly what you want. A framework is scaffolding for an under-specified idea; if your prompt already scores in the eighties, adding one adds ceremony rather than clarity.

Scoring image and video prompts

The checker is built for text prompts, and about half of it transfers.

Transfers well: specificity, exclusions, constraints, a clear task. An image prompt lives or dies on concrete nouns, and "no text, no watermark, no extra limbs" is exactly the sort of exclusion that saves a generation.

Does not apply: length targets and reasoning instructions. An image generator has no word count and does not deliberate. Expect those to read as gaps when they are not, and ignore them.

Visual prompts also have their own concerns the checker does not model — camera and lens language, lighting, aspect ratio, negative prompts, and style references. If you are working in images, treat the score as a partial signal rather than a verdict.

The five mistakes that cost the most points

1. No audience. The most common omission and among the most expensive. A model writing for everyone writes for no one.
2. Format left implied. You picture a table; you get six paragraphs. You picture bullets; you get an essay. Say it.
3. Vague quality words. "Make it engaging" is not an instruction, it is a hope. Replace it with something checkable: "open with a concrete example, no adjectives in the first sentence".
4. No exclusions. Saying what you do not want is faster than correcting it afterwards, and it is the cheapest way to stop invented statistics and filler.
5. No example. Skipped almost universally, worth 10 points, and usually the fastest improvement available to a prompt already scoring in the seventies.

How the scoring works

The checker is deterministic. No language model grades your prompt — it is pattern analysis in plain PHP, so the same prompt always returns the same score. That matters more than it sounds: ask an AI to rate a prompt twice and you will often get two different numbers, which makes it useless for telling whether an edit actually helped.

Nothing is stored. Your prompt is scored in memory and discarded — not logged, not saved, not used for training.

Common questions

Does a higher score mean a better answer?

It means fewer decisions are left to the model. A low-scoring prompt can still get a good answer by luck, when the model happens to guess your audience and format correctly. A high-scoring prompt gets a usable answer repeatedly.

Should every prompt score 100?

No. Some checks do not apply to some tasks — a tweet does not need a reasoning instruction, a quick factual question does not need a worked example. Anything at 70 or above gives a model what it needs. Chase the specific gaps that matter for your task rather than the number.

Why did my prompt lose points for something it has?

Each check looks for the thing done concretely. "Use a good format" mentions format without specifying one, so it earns partial credit rather than full. The card tells you which half is missing.

Does prompt length matter?

Only as a proxy. Twenty words is usually too few to carry a task plus context; past about 600 the instruction starts getting buried in background. What matters is what the words do, not how many there are — two 50-word prompts in our testing scored 43 and 100.

Do these rules apply to every AI model?

Broadly yes. Clarity, audience, format and constraints help every model. Newer reasoning models need the "think step by step" instruction less, because they do it internally, but nothing on this list makes output worse.

Glossary

• Prompt engineering — writing inputs so a model reliably produces what you need, rather than needing several attempts.
• Zero-shot — asking without examples. What most people do.
• Few-shot — including one or more examples of the desired output.
• Chain of thought — asking the model to reason before concluding.
• System prompt — standing instructions that apply to a whole conversation, separate from any single message.
• Temperature — how much randomness the model uses. Lower is more predictable, higher more varied.
• Hallucination — confident output that is not true. Prompts asking for citations or statistics invite it.
• Negative prompt — things to exclude. Standard in image generation; useful as plain exclusions in text.
• Token — the unit models read and bill in, roughly ¾ of a word.