Free AI Prompt Generator for ChatGPT, Claude, Gemini and Midjourney
Describe what you want and take back prompts written the way your model expects to be spoken to. Ten platforms, eight subject areas, five tones, and a structured RCO mode for anything longer than a quick question. The prompts are assembled in this page from a template library, so nothing you type is sent anywhere and there is no sign-up.
Why most prompts get bland answers
The common failure is naming a topic and stopping there. "Write about remote work" gives the model no reader, no length, no format and no angle, so it fills those in with the safest option available — which is why the answer reads like it could have been written for anyone. It was.
Four things close most of that gap: who is answering, who is reading, what shape the answer should take, and how long it should be. None of them is hard to supply, and together they turn a generic reply into a usable one. The fields on this page exist to collect exactly those.
The audience field is the one people skip and the one that changes the most. The same question written for engineering managers and for first-year students produces different vocabulary, different assumptions about what needs explaining, and different examples.
Every model wants to be asked differently
Prompts are not portable in the way people assume. Claude works well when the task and the instructions are separated with XML tags, which is Anthropic's own documented guidance. Llama models were tuned on an instruction-and-response format and follow it more reliably than free prose. Perplexity is built around sourcing, so asking for citations and recent material changes what it actually does rather than just how it sounds.
Copilot answers from the files and context open around it, so a prompt that tells it to work from the workspace beats one that asks in the abstract. Grok responds to directness and to being told where the evidence is thin. ChatGPT benefits from being asked to reason before answering.
Choosing a platform here applies that model's convention to the prompt rather than relabelling the same text.
- Claude — XML tags marking task and instructions apart
- Perplexity — citations, recency, and where sources disagree
- Llama — the instruction/response format it was tuned on
RCO: Role, Context, Output format
RCO is a way of laying out a prompt so that nothing important is left implicit. Role sets the expertise the answer should come from. Context supplies the subject and the reader. Output format fixes the length, the structure and the conventions of the reply.
Switching on RCO Structure Mode adds a fully written prompt in that form as the first result. It is longer than the others and worth using for anything substantial — a piece of writing, a plan, an analysis. For a quick question the shorter prompts are usually enough.
Image prompts are a different craft
Image models read a description rather than an instruction. What moves the result is subject, lighting, lens, composition and style — and the syntax of the specific tool. Midjourney takes flags such as --ar 16:9 for aspect ratio and --no for exclusions. Stable Diffusion takes weighted terms and a separate negative prompt line. DALL-E takes plain prose and has no formal negative-prompt field at all, so exclusions have to be phrased inside the description.
Choosing an image platform switches the output to that syntax, and the negative prompt you enter is passed the way each tool expects rather than appended identically to all three.
Asking for several prompts is worth doing here more than anywhere else, because image models are sensitive to framing: the same subject shot wide, square and close-up gives three genuinely different pictures.
What this tool is not
It does not write the answer. It builds the prompt, and you paste that into whichever model you use. Nothing here calls an AI, which is why it returns instantly, costs nothing, and needs no account.
It also means what you type never leaves the tab. The template library lives in the page, so a prompt about unreleased work or a client project stays on your machine.
Frequently Asked Questions
What makes a prompt good?
Specificity, and saying what the output should look like. Most weak prompts fail because they name a topic and stop — "write about remote work" leaves the model to guess the audience, the length, the format and the angle, and it will guess blandly. Naming the role, the reader, the format and the length removes the guessing, and that is most of the difference between a usable answer and a generic one.
Do different AI models need different prompts?
Yes, more than people expect. Claude responds well to XML tags that mark out the task and the instructions separately. Perplexity is built around sourcing, so asking for citations changes what it does. Llama models were tuned on an instruction/response format. Copilot works from the files open around it. Picking your platform here applies that model's convention rather than handing you the same text with a different label on it.
What is RCO structure?
Role, Context, Output format — the three things a model needs before it can answer well. Role sets the expertise to answer from, Context supplies the subject and the audience, and Output format fixes the length and shape of the reply. Switching on RCO Structure Mode adds a fully laid out prompt in that form as the first result, which is worth using for anything longer than a quick question.
Should I say who the audience is?
It is the single highest-value optional field here. The same request written for "engineering managers" and for "first-year students" produces genuinely different answers — different vocabulary, different assumptions about what needs explaining, different examples. Leaving it empty tells the model to write for a general audience, which usually means writing for nobody in particular.
What does the tone setting change?
The register the model writes in, and how much it assumes you already know. Technical keeps domain terminology and skips the explanations; Casual writes the way a knowledgeable colleague would talk; Creative gives the model permission to break the usual structure. Tone is a real instruction to the model, not a label on the prompt.
How do image prompts differ from text prompts?
Image models read a description, not an instruction. They respond to subject, lighting, lens, composition and style, and they respond to the syntax of the specific tool — Midjourney takes flags like --ar 16:9, Stable Diffusion takes weighted terms and a separate negative prompt, DALL-E takes plain prose. Choosing an image platform switches the output to that syntax.
What is a negative prompt?
A list of what you do not want to see — blurry, extra fingers, watermark, text. It is passed the way each tool expects: appended after --no for Midjourney, as a separate Negative prompt line for Stable Diffusion, and as an exclusion note for DALL-E, which has no formal negative-prompt field.
Why generate several prompts at once?
Because the first phrasing is rarely the best one, and it is quicker to compare four angles on the same topic than to rewrite one prompt four times. Each result approaches the subject differently rather than restating it, so they are worth reading before you pick.
Are my prompts sent anywhere?
No. The prompts are assembled in your browser from a template library held in the page — there is no API call, no account, and nothing is stored. What you type goes no further than the tab, which matters if the topic is client work or anything unreleased.
Is this an AI writing tool?
No, and that distinction is worth keeping. This builds the prompt; you paste it into ChatGPT, Claude, Gemini, Midjourney or whichever tool you use, and that tool writes the answer. Nothing here calls a model, which is why it is instant and free.