Paste a draft or upload a document. Mechanical patterns get fixed. Everything that needs your judgement gets flagged, with the reason.
Your text is processed and returned. Nothing is stored, logged, or used for anything else. Uploaded files are deleted before the response is sent. Limit 5,000 words per run.
It is rarely one word. It is the accumulation: a paragraph that opens with "It is important to note that", ends on a metaphor, and puts three abstract nouns in a row on the way through. Any one of those is ordinary. Together they are a fingerprint.
Roughly four groups. Stock vocabulary — the words that appear far more often in generated text than in human writing: comprehensive, crucial, showcase, delve, leverage, seamless. Filler — phrases that occupy space without carrying meaning, like "in order to" or "at the end of the day". Manufactured drama — the colon reveal, the self-answered question, the closing aphorism that converts a point into a quotable line. And structural evenness: sentences all about the same length, every claim stated with identical confidence, no first person anywhere in two thousand words.
A filler phrase can be deleted safely because removing it cannot change what you meant. A closing metaphor cannot: deleting it leaves the paragraph without an ending, and only you know which of your own sentences should carry that weight instead. Three abstract nouns in a row might be padding, or might be three real things. So the tool separates the two. It fixes what is mechanical and hands back what needs a decision, with the reason attached, so you learn the pattern rather than outsourcing it.
Strip every flagged instance and the writing flattens into something equally artificial — a scrubbed average with no voice in it. When a paragraph carries six of these patterns, fixing three usually restores a human cadence. Keep the hedge that reflects real doubt. Keep the aside that goes slightly off-topic. Those are the parts that sound like a person.