What makes a good prompt
A good prompt gives the model what a capable new colleague would need on day one: what to do, why, for whom, and what done looks like. The grader checks nine parts and weights each one by how much it tends to change the answer:
- Clarity, 25 points. The task, its scope and its finish line, said plainly. "Write a 120-word email" is clear; "do something about the email" is not.
- Context, 20 points. The background: the situation, what has happened so far, the facts the answer has to use. The model knows nothing about your work unless you tell it.
- Format, 12 points. The shape and length of the answer, such as five bullets, a table or an email with a subject line.
- Constraints, 10 points. Tone, words or claims to avoid, sources to use, things that must be included.
- Audience, 8 points. Who will read the answer. "For a busy CFO who skims" changes more than it seems to.
- Examples, 8 points. One sample of the output you want. A single example often does more than a paragraph of description.
- Success criteria, 7 points. What a good answer achieves, or how you will judge it.
- Role, 5 points. A point of view to answer from, such as an experienced recruiter or a careful editor.
- Steps, 5 points. An order of work, for jobs with more than one stage.
Instructions that pull against each other, like "be thorough" and "keep it to two lines", cost up to 10 points, because the model has to guess which one you meant.
Here is the same request written twice. When we ran both through the grader, the first scored 24 out of 100 and the second 97.
Before: 24 / 100
write a cold email about our expense tool
After: 97 / 100
You are an experienced B2B copywriter. Task: Write a cold email to finance leads at software companies with 50 to 200 staff, introducing our expense tool. Context: They close the books every month and lose days chasing receipts. Our tool reads receipts from email and matches them to card payments. Audience: busy CFOs who skim. Format: a subject line, then a body under 120 words, then a one-line PS. Tone: plain and direct. No hype, no buzzwords, no made-up numbers. Here is a subject line in the style I want: "Receipts, chased for you" What good looks like: a CFO could reply yes or no in one line.
The second version is longer, but every extra line answers a question the model would otherwise have to guess at. That is most of what prompt writing comes down to.
How the generator picks the settings
When you press the button, your idea goes to Jev, a decision model from TypeSafe. Jev does not write text. It answers a fixed set of questions about your idea and returns a probability for every possible answer:
- Which expert should answer, out of 20 roles, from copywriter and data analyst to teacher and recruiter.
- Which format fits, out of 10, from short paragraphs and bullets to a table, an email or JSON.
- Which tone (7 options), which audience (8) and how long the answer should be (4 lengths).
- What the job needs: whether an example would help, whether it should be worked through step by step, and whether key facts are missing so the AI should ask first.
- Whether it repeats, like a weekly report or a daily post.
The page then builds the prompt from fixed sentences, one per setting. No language model writes any of it, so the same settings always give the same prompt and nothing is made up about your work. The chip beside each setting shows how sure Jev was. Where it says unsure, check that setting first; change it and the prompt updates as you go.
Prompts for ChatGPT, Claude and Gemini
The same structure works across all of them, so the page serves as a ChatGPT prompt generator, a Claude one or a Gemini one without any changes. A role, a clear task, background, an audience, a format and a finish line are what every current chat model responds to. A few habits help everywhere:
- Put long material first. When you paste a document or transcript, put it above your instructions and ask your question at the end. Anthropic recommends this for Claude, and it does no harm with other models.
- Mark the parts. In a long prompt, label each section ("Context:", "Format:") or wrap pasted text in tags such as <transcript>, so the model does not mix up your instructions with your material.
- Save prompts you reuse. ChatGPT custom GPTs, Claude Projects and Gemini Gems all keep standing instructions, so the role, audience and tone lines only need writing once.
- Grade, then test. A high grade means nothing obvious is missing. Running the prompt on three real inputs shows whether it does the job.
Prompts for recurring work and AI agents
A prompt you run once can lean on you: if the answer is off, you ask again. A prompt that runs every Monday as an agent has nobody reading each run, so it has to carry more. On top of the parts above, a good agent prompt says:
- what starts each run, such as a time, a new email or a new row in a sheet;
- which inputs it gets and where they come from;
- what to do when something is missing or looks wrong;
- where the output goes, and in what exact shape;
- when to stop and ask a person instead of guessing.
So the generator doubles as a simple system prompt generator for AI agents: keep the role, audience, format and tone lines as the standing instructions, and use the task line for each run. For the full method, read how to write a prompt for a recurring AI agent, the five parts of a production prompt in AI agent prompt engineering, and how an agent prompt differs from a chat prompt.
When the generator thinks your idea repeats on a schedule, it says so under the prompt. Gravity runs that kind of recurring prompt as an agent, on the schedule you set, with nothing to build. It is in private alpha, and you can apply to the alpha.
What the tool cannot do
The generator picks from fixed lists, so an unusual role, format or audience needs a quick edit by hand after you copy the prompt. It cannot know facts about your business, your customers or your numbers. That is what the brackets are for, and a prompt with unfilled brackets will get a generic answer.
The grade measures how complete a prompt is, not whether it asks for the right thing. A tidy prompt can still aim at the wrong goal, and a short prompt can be right for a simple job. The score also comes from a model, so the same prompt can move by a point or two between runs. Use it to find what is missing, then judge the prompt by the answers it gets.
Questions
Is the prompt generator free?
Yes. The generator and the grader are both free to use, with no sign-up and no email needed. Each run takes about a second.
Which AI does it work with?
Any chat model you type into: ChatGPT, Claude, Gemini, Copilot, Llama, Mistral and the rest. The prompt is plain text built from parts every model follows, such as a role, a task, an audience and a format, so you can paste it anywhere. The grader works the same way on prompts written for any of them.
Does it store my prompt?
Only if you choose to. Your idea or prompt is sent to the model to pick the settings or score it, and it is not saved by Gravity, unless you press Share it or ask us to set it up as an agent; an idea shared that way stays private and is never published. We count how many times the tool runs each day, nothing more. Leave out names, passwords and other private details; the tool does not need them.
How is the grade calculated?
The grader checks nine parts and adds up points: clarity up to 25, context up to 20, format 12, constraints 10, audience 8, examples 8, success criteria 7, role 5 and steps 5, for 100 in total. Clarity and context are scored on a scale; the other seven earn their full points when the model is sure they are present and fewer when it is not. Instructions that conflict with each other take off up to 10 points.
Why does the generator not write the answer itself?
Because its job is the prompt, not the answer. The settings come from a decision model that chooses among fixed options, and the page builds the prompt from fixed sentences, so it never invents facts about you or your work. Fill in the bracketed parts with your own details and paste the prompt into the AI you already use.
How do I write a system prompt for an AI agent?
Start with the same parts as any good prompt: who the agent is, what job it does, who the output is for, and the format and tone. Then add what a one-off prompt can skip: what starts each run, which inputs it gets, what to do when something is missing, where the output goes, and when to stop and ask a person. A quick way is to generate a prompt here, keep the role, audience, format and tone lines as the system prompt, and use the task line for each run. Our guide to writing a prompt for a recurring agent covers the rest.