Prompt Compressor
Long prompts cost tokens on every single call. This rewrites yours for density - merging repeated instructions, cutting filler, tightening wording - while keeping every constraint, example and format rule that actually changes the output. You get the token savings, a diff of what changed, and a clear flag when it has compressed as far as it safely can.
1. Paste the prompt you want to shrink
2. How hard should it squeeze?
Your token savings appear here
You will see the shorter prompt, exactly how many tokens it saves, a word-by-word diff of what changed, and an honest flag if compressing further would start costing you meaning.
Compression, not truncation
Cutting a prompt in half is easy if you are allowed to delete requirements. The hard part - and the only version worth using - is saying the same thing in fewer tokens. So the instruction behind this tool names what must survive: word and character limits, output format and schema, worked examples and few-shot demonstrations, the role and tone, the audience, every edge case, every “do not do X” rule, and every number, proper noun and placeholder exactly as written. What goes is politeness, hedging, commentary about the prompt itself, the same instruction stated three different ways, and adjectives that do not change a thing about the answer.
Because that promise is easy to make and hard to verify, nothing is hidden: the result is shown as a word-by-word diff against your original, with the model's own list of what it cut and what it deliberately kept, plus a risk rating. If it had to squeeze so far that a requirement might be weakened, it says so instead of quietly reporting a bigger percentage.
Savings are measured with the same rough estimate used across this site's Markdown converters - about four characters per token, which is close enough for an honest comparison without shipping a multi-hundred-kilobyte tokenizer to your browser. Treat it as a directionally-correct free prompt token counter, not as billing-grade accuracy: your provider's exact count will differ by a few percent.
Häufig gestellte Fragen
How do I shorten an AI prompt without losing meaning?
Remove the things the model never acted on in the first place - pleasantries, framing, justifications, repeated instructions - and convert wordy prose into labelled, imperative lines. What must stay is anything that changes the output: length limits, output format, examples, tone, audience, edge cases and exclusions. This tool does exactly that rewrite and then shows you a diff so you can confirm nothing load-bearing went missing.
How much can I expect to reduce prompt tokens by?
It depends entirely on how the prompt was written. Conversational prompts written in full paragraphs commonly drop 40-60% at the Balanced level, while a prompt that is already a terse bulleted spec may only give up 10-15%. Light typically saves 15-30%, Balanced 35-55% and Aggressive 55-75%. If a prompt is already dense, the tool tells you there is nothing meaningful left to save rather than inventing a number.
What is the difference between the three compression levels?
They are genuinely different instructions, not presets on a slider. Light keeps your structure, sentence style and voice and only strips filler and repetition. Balanced rewrites for density: overlapping instructions get merged, prose becomes terse imperative lines, and sections can be collapsed or reordered. Aggressive goes telegraphic - labels instead of sentences, connective words dropped - and is told to stop and flag it rather than cut a real requirement to hit a target.
Is this a free prompt token counter too?
Yes. The box shows an estimated token count, word count and character count live as you type, before you compress anything, and the result compares the two versions side by side. The estimate uses roughly four characters per token, which is the same convention used across this site - accurate enough to compare two versions of the same prompt honestly, but not a substitute for your provider's exact tokenizer.
Can I use it to compress a prompt for the Claude API or a system prompt?
Yes - that is the case it pays off most. A system prompt or agent instruction block is re-sent on every call, so as a token optimizer for GPT prompts or a way to compress a prompt for the Claude API, the saving multiplies by your request volume. Compress it once, read the diff carefully, test the shorter version on a few real inputs, then ship it. For prompts that use caching, check the compressed version still sits inside your cached prefix.
Is my prompt stored anywhere?
No. Your prompt is sent once to a free AI model to be rewritten and is not saved on this site afterwards. The diff, the token comparison and the level switching all happen in your browser, so nothing extra leaves the page.