AI Prompt Formatter
Write the parts of your prompt once, then get it in the structure your model actually responds to. Your words are never rewritten. The token cost of the structure is shown as you go.
- Runs in your browser
- Nothing uploaded
- No signup
- Free forever
Your prompt
Who the model should act as. Skip it if the task speaks for itself.
Background the model cannot infer. Systems, constraints, prior decisions.
The one thing you want done. Be specific about the verb.
Rules and boundaries. One per line.
Input and output pairs. The single highest-leverage thing you can add.
What you want back. Models pick for themselves when you don't say.
Formatted for
XML tags. Anthropic's own prompting guidance recommends them, and Claude is trained to treat tag boundaries as structure rather than as text to imitate.
Fill in the task and the formatted prompt appears here.Before you send it
- Missing. There is no task. Everything else is scaffolding around it.
- Check. No output format. The model will choose one, and it will choose differently next time.
Your words are never rewritten. This arranges what you typed into the structure the target model responds to, and tells you what it costs. Nothing is uploaded.
How to structure a prompt properly
Most weak prompts are not badly written. They are a single paragraph doing six jobs at once, which leaves the model to work out where the context ends and the instruction begins.
- 1
Split what you have into parts
Put the background in Context, the thing you want done in Task, the rules in Constraints. Only Task is required. A prompt with just a task comes out unchanged, because wrapping one sentence in headings helps nobody.
- 2
Say what you want back
The Output format field is the one most people skip and the one that changes the answer most. Without it the model picks a format, and picks a different one next time.
- 3
Pick the model you're sending it to
Claude gets XML tags. GPT and Gemini get markdown headings. Plain text gets uppercase labels for API calls and logs. The same fields render three different ways.
- 4
Read the warnings, then copy
The checks below the output flag a missing output format, conflicting length instructions, and filler that costs tokens without changing the answer. None of them rewrite anything.
Why the structure differs per model
Anthropic's prompting guidance for Claude recommends XML tags, and Claude is trained to read tag boundaries as structure rather than as text to copy. A prompt wrapped in <context> and <task> gives it an unambiguous split between what is background and what is the instruction.
OpenAI's guidance leans on clear section delimiters instead, and markdown headings are what the ChatGPT interface already renders. Google's guidance for Gemini emphasises explicit sections without pushing a particular markup, so markdown works there too.
This matters more as prompts get longer. In a short prompt the model can infer the boundaries. In a prompt carrying three paragraphs of context, a list of constraints and an example, the boundaries are the thing that stops your constraints being read as part of your background.
What structure costs
Tags and headings are tokens, and you pay for them on every call. The counter above shows the total and how much of it came from the scaffolding rather than your words, using the same byte-pair encoder OpenAI bills on.
For a one-line question the overhead is most of the prompt, which is why a task on its own is returned unformatted here. For a long prompt run repeatedly in an application, the overhead is usually a rounding error against the clarity it buys. The number is shown so it is your decision rather than the tool's.
What this tool will not do
It will not rewrite your words, invent constraints you did not write, or add an example you did not provide. Every formatter that improves your prompt for you is either sending it to a model, which means uploading it, or padding it with boilerplate you then ship without reading.
This runs entirely in your browser. Nothing is uploaded, and there is nothing to sign up for. What comes out is what you typed, arranged.
Frequently asked questions
Does it rewrite my prompt?
No. It arranges the text you typed into sections and never generates prose. The only thing it changes is whitespace: trailing spaces go, and runs of blank lines collapse. Indentation inside your examples is left alone, because that is usually code.
Why XML for Claude and markdown for GPT?
Anthropic's own prompting documentation recommends XML tags for Claude, which is trained to treat them as structural. OpenAI's guidance emphasises clear delimiters, and markdown is already what the ChatGPT interface renders. Same content, different convention.
Is my prompt sent anywhere?
No. The formatting is plain string manipulation in your browser. The token counter loads a vocabulary file once, then counts locally with no network at all.
Why does the output sometimes have no headings?
If the only field you filled is Task, the prompt comes back as you wrote it. Wrapping a single sentence in section headings makes it longer without making it clearer.
What does 'from structure' mean in the token count?
The difference between the formatted prompt and the same text with no tags or headings. It is the price of the scaffolding, shown so you can decide whether a given prompt is worth it.
Are the token counts exact?
They are exact for GPT models, computed from OpenAI's published o200k_base vocabulary. Claude and Gemini tokenize differently and neither vendor publishes an offline tokenizer, so treat the figure as a close guide for those. The AI Token Counter has the per-model breakdown.
Do the checks use AI?
No. They are fixed rules: a missing output format, more than one length instruction, filler phrases, an unusually long unbroken line. Deterministic, instant, and wrong sometimes. Ignore any that do not apply.
Do I need an account?
No account, no email, no signup. It is a static page.
The prompt was the easy half.
A good prompt earns a good answer, once. Then the tab closes and the answer goes with it, and the next time you need it you write the prompt again. Savelore saves the responses worth keeping, across all nine supported platforms, so the work you put into getting a good answer keeps paying out.