Make AI text read like a person wrote it.

How it works

Locked details

Before anything is rewritten, Plainspoken finds the parts that must not change: numbers, dates, names, links, citations, code, quoted text, and [bracketed placeholders]. A draft that alters or drops one, or adds a new one, is rejected. Plainspoken never fills in a placeholder for you.

Two drafts, then checks

The writer model drafts each paragraph twice. A separate judge model lists the paragraph's claims and checks every draft against them, then scores the drafts and your original without knowing which is which. A draft wins only if it keeps every claim, adds none, and drops most of the tells. If none passes, the writer gets two repair rounds. If the meaning still does not hold, your paragraph is kept as written and marked, so a changed meaning never ships.

The tell score

The report is deterministic and runs in your browser. It counts patterns that guides on AI writing describe, such as stock vocabulary, contrast framing, and lists of three, plus rhythm checks like how much sentence length varies. The score is weighted matches per 100 words. It measures style habits. It cannot tell who wrote a text, and a lower score does not mean an AI detector will pass it: in the eval, one detector still flagged 41 of 42 rewrites. The patterns and their sources are in tells.json.

Privacy

Your text goes to the model provider for the rewrite and nowhere else. There is no database, and pasted text is never stored or logged. Your voice profile stays in this browser. A key you paste lives in this tab's memory and is sent only with your requests.

Limits

The free key takes 1,500 words per request, a few requests per network every ten minutes, and has a daily cap across everyone. With your own Anthropic or Groq key there are no caps beyond your provider's. The model is told to keep length within about 10%, so very padded text can come back shorter.