// what it does and why
Paste a rough prompt, press one button, get the same request rebuilt as a proper brief. It runs entirely in your browser. No account, no API key, nothing sent to a server, and it keeps working with the wifi off.
Open it →You type something like write me a blog post about coffee and press Polish it. What comes back is that same sentence wrapped in the structure a model responds to, ready to copy into whichever assistant you use.
Most disappointing answers from an AI are a prompt problem, not a model problem, and it is usually the same three omissions: who the answer is for, what shape it should arrive in, and whether the model is allowed to ask instead of guessing. I was typing that scaffolding by hand several times a day.
The clarifying-questions line is the one that changed my results most. Left alone, a model invents the missing detail and commits to it. Given permission, it asks, and you find out in one sentence that it was about to write the wrong thing.
There is no AI in this tool. Nothing is being understood. It scans your text for ten verbs, write, code, design, plan, analyze, explain, summarize, create, build and draft, and the first one it finds picks the role. It also watches for words that suggest a list and words that suggest code, and swaps two lines accordingly. That is the entire intelligence in the page.
It matches on substrings, so it can be caught out. Ask it about planets and it finds plan, decides you want a strategist, and says so. That is a real flaw, and I left it there: the role is one sentence at the top of the output where you can see it and delete it, and a wrong guess you can spot beats a page that stalls trying to be clever.
The Context section does not come back written. It comes back as an instruction to you: one to three sentences on who this is for, where it will be used, and any background that changes the answer. Everything else the tool prints is generic, and generic is fine there, because good structure is the same for everybody. Context is the only part that is yours alone.
Inventing something plausible in that slot would have been the worst thing it could do, because a field that looks filled in stops being read. A visible gap gets filled.
One page, no build step, no server, no storage. The rewriting is plain text assembly, so the result appears the instant you click and no request goes anywhere. A copy button puts it on your clipboard and flashes to confirm.
The honest trade: the thing that could genuinely improve your prompt is a language model, and calling one means either my key paying for everybody's traffic, or a box asking you to paste yours into a web page you do not control. I was not willing to ship either.
It cannot turn a vague idea into a good one. Feed it four words and you get four words in a nicer frame. It gives your request the shape a model expects, and the judgement stays yours: fill in the context, cut any requirement that does not apply, then send it.