Build clear, structured prompts for ChatGPT, Claude, Gemini, Midjourney and DALL-E. Fill in what you need, pick a framework, and the tool assembles a complete, professionally structured brief with a live score showing how strong it is. Free, no sign-up, and everything runs in your browser.
Most people type a single sentence into ChatGPT, get a generic answer, and conclude the model is not very good. Usually the model is fine and the prompt is the problem. A language model cannot read your mind about who you are writing for, how long the answer should be, what tone fits, or what you definitely do not want.
A structured prompt supplies that missing information. Instead of "write about SIP investments", you tell the model who to be, who is reading, how long, in what format, and what to avoid. The difference in output quality is not subtle — it is usually the difference between something you throw away and something you can actually edit and publish.
This tool builds that structure for you. You fill in what you know, and it assembles the prompt using a recognised framework, then scores it so you can see what is still missing.
The strength score measures how completely you have specified the request, not how good the eventual answer will be. A 100% score on a vague task still produces a vague answer — the task field is where most of the quality comes from.
These are established ways of organising a prompt. None is objectively best; they suit different situations.
| Framework | Stands for | Best for |
|---|---|---|
| CO-STAR | Context, Objective, Style, Tone, Audience, Response | Content writing and anything where tone and audience matter |
| RTF | Role, Task, Format | Quick everyday requests where you want a fast, clean answer |
| RACE | Role, Action, Context, Expectation | Task-focused work where the outcome matters more than the style |
| Chain of thought | Reason step by step, then answer | Analysis, maths, debugging and multi-step problems |
| Plain | No formal structure | When you want the fields written out simply without labels |
Image models need different information from text models. They respond to concrete visual description — subject, setting, style, lighting, camera framing — rather than instructions about tone or audience. A useful rule is to describe what a photograph of the scene would look like, not what you want the image to mean.
| Element | Why it matters | Example |
|---|---|---|
| Subject | The single most important field. Be specific about who or what, and what they are doing. | an elderly potter shaping a clay pot |
| Setting | Anchors the scene and fills the background sensibly. | a sunlit village workshop |
| Style | Decides whether you get a photo, a painting or an illustration. | documentary photography |
| Lighting | Changes the mood more than almost anything else. | golden hour |
| Framing and lens | Controls how close the subject is and how the background blurs. | medium shot, 85mm |
| Negative prompt | Removes common failures such as text and distorted hands. | text, watermark, extra fingers |
The tool formats the output differently for each target. Midjourney gets its parameter flags such as --ar and --no, Stable Diffusion gets a separate negative prompt block, and DALL-E gets plain descriptive sentences, since it responds better to natural language than to flags.
Your context and audience fields are probably empty. These two do more to sharpen an answer than any other setting, because they tell the model what situation it is writing into.
Set the length field explicitly. Models default to a medium-length answer when you do not specify, which is rarely what you actually wanted.
Set the tone field, and use the example field to paste a short sample of writing in the voice you want. An example is worth more than any tone label.
Put those items in the constraints field as direct instructions. Phrase them as what to do rather than only what to avoid — "explain jargon in plain words" works better than "avoid jargon".
Long, poetic descriptions confuse image models. Cut abstract or emotional wording and describe concrete visual detail instead. Add a negative prompt for the specific problems you keep seeing.
Add a constraint asking it to say when it is unsure rather than guessing, and enable the cite sources extra. No prompt eliminates this entirely, so always verify facts, figures and quotes yourself.
There is a recognisable rhythm to unedited AI output, and most readers now notice it even if they cannot name it. Sentences settle into the same medium length. Paragraphs open with the same handful of connector words. Conclusions tie everything up in a neat, symmetrical bow. None of this is wrong exactly — it is just relentlessly average, and averageness is precisely what a predictive model produces when it is not told to do otherwise.
The natural human tone control on this page exists to counter that. It does not ask the model to lie about being an AI or to hide anything — it asks for the specific, describable habits of good human writing: varied sentence length, a concrete detail instead of a vague claim, a clear position instead of a hedge in every sentence. These are legitimate writing craft instructions, the same ones an editor would give a junior writer.
Natural adds a short, general instruction: vary your sentences, use contractions, skip the obvious AI clichés. It is enough for a quick social post or an internal email. Strong is considerably more detailed. It names around fifteen specific overused AI phrases and tells the model never to use them, explains the concept of sentence-length variation directly, asks for at least one concrete example or number, and instructs the model to take a position rather than hedge every line. For anything that will be published under your name or your publication's name — a blog post, a news article, a newsletter — Strong is worth the extra length it adds to the prompt.
The "optimised for" setting adds a further layer on top of the tone instructions. Choosing newsroom tells the model to write in inverted-pyramid style with short, mobile-friendly paragraphs and source attribution, the way wire copy is written. Choosing students asks for a patient, term-defining explanation built up logically, the way a good teacher would explain a concept for genuine understanding. Choosing business asks for persuasive writing anchored in concrete numbers and outcomes, with the reader's likely objections addressed rather than ignored. Combine a profile with the human tone control and a target SEO keyword, and the resulting brief resembles what a professional prompt engineer would actually hand to a writer — not a single instruction, but a short specification covering voice, structure, audience and search intent together.
An article that ranks well and an article that reads naturally are not in conflict, but they do need to be asked for together. The SEO requirements section this tool adds is deliberately narrow: work the keyword into the title or opening line, into the first hundred words, and into at least one subheading, then move on to related terms rather than repeating the exact phrase. That is enough to satisfy search intent without producing the stiff, over-repeated keyword density that reads badly and, these days, tends to rank worse rather than better.
The prompt is assembled by your browser using templates. There is no AI model running here and no server involved, so your ideas, client details and unpublished work stay on your own device. The prompt is only seen by an AI service once you paste it there yourself.
Turn the natural human tone control to Strong. It adds a detailed instruction set to your prompt covering sentence-length variation, contractions, a list of overused AI phrases to avoid, and a request for concrete detail over vague statements — the same craft notes a good editor would give a writer. Combine it with the example field and paste a short sample in the voice you want; a real example does more than any instruction on its own.
Natural is a short, general instruction suited to quick, low-stakes writing such as a social post or an internal email. Strong is a full, detailed rule set — it names specific banned phrases, explains sentence rhythm directly, and asks for a concrete example — and is worth using for anything that will be published under your name.
It adds a writing-profile instruction matched to your audience. Newsroom asks for inverted-pyramid, source-attributed news writing. Students asks for a patient, clearly sequenced explanation built for understanding. Business asks for persuasive writing anchored in concrete numbers and outcomes. It combines with the human tone and SEO keyword settings to produce one complete brief.
Yes, if the piece needs to rank in search. The SEO requirements section this adds is deliberately restrained — work the keyword into the title, the opening line, the first hundred words and one subheading, then move to related terms. This satisfies search intent without producing the repetitive, over-stuffed writing that reads badly and tends to rank worse.
No. It assembles your input into a proven structure using templates that run in your browser. That means it is instant, free, works offline once loaded, and never sends your ideas to a server. The AI comes in when you paste the finished prompt into ChatGPT, Claude, Gemini or an image tool.
Text prompts work with ChatGPT, Claude, Gemini, Copilot, Perplexity and any other chat model. Image prompts are formatted for Midjourney, DALL-E and Stable Diffusion, and there is a general option for other tools.
Context, Objective, Style, Tone, Audience, Response. It is a widely used structure that covers the six things a model usually needs to know, which is why it is the default here.
How completely you have specified your request. Each field carries a weight, with the task counting for the most. It measures completeness rather than quality, so a fully specified but vague task will still score high and produce a weak answer.
No. Only the task is required. Each additional field makes the output more predictable, so fill in what genuinely matters for your situation and leave the rest.
Those are Midjourney parameters. The first sets the aspect ratio and the second excludes unwanted elements. They are added automatically when Midjourney is the selected target, and left out for other tools that do not understand them.
Yes. The prompts you build are yours. Whether the AI output can be used commercially depends on the terms of the AI service you use it with, so check those separately.
No. Everything happens in your browser and nothing is stored or transmitted. Closing the page clears it.