PlainSpeak and watermarks-remover
If you already use watermarks-remover, this page says how PlainSpeak relates to it. They overlap on finishing text and differ on most else, and they work well together.
Facts about watermarks-remover come from its own README and docs at v0.7.0, read on 1 November 2026. Facts about PlainSpeak come from this documentation and the measurements on the Protocol page. Neither project endorses the other.
Two different jobs
watermarks-remover is a free, MIT licensed toolkit for stripping provenance marks from content you own: invisible characters in text, metadata in files, and, through optional external tools, marks in images. It ships two agent skills for Claude Code, Codex, Cursor and other hosts. remove-ai-marks drives the whole toolkit through a small local service. clean-user-facing-text is text only and self contained: it scores prose for machine-like signals, strips invisible characters, and has the agent rewrite the prose once while keeping facts and the writer's voice. That second skill is the closest to what PlainSpeak does.
PlainSpeak is a writing tool. It shows which wording makes copy read as machine made, rewrites it in a voice you choose while keeping your facts, and checks what changed. It runs as a desktop app with your own model key.
If your goal is a clean file, watermarks-remover covers far more ground. For finishing prose, both aim at copy that reads as yours: clean-user-facing-text as a workflow your agent follows, PlainSpeak as an app that runs each check itself and shows you why.
Side by side
| watermarks-remover (v0.7.0) | PlainSpeak | |
|---|---|---|
| Form | Agent skill plus a local Python service; scripts and an HTTP API | Desktop app for Windows and Mac |
| Setup | Install the skill into your agent host with one Python command, or from the Claude Code plugin marketplace, then start the local service: Python 3.10 or later, no other packages; Docker optional; some file formats also need system tools such as qpdf and exiftool | Install the app and add a model key in Settings; no Python, Docker or command line |
| Which model rewrites | Your agent host's own model (Claude in Claude Code, GPT in Codex, and so on), or, through its rewrite hook, Ollama or any OpenAI-compatible service | Any model on OpenRouter, or Anthropic, OpenAI or any OpenAI-compatible service, chosen per run in the recipe, with measured notes on how models behave |
| Price | Free, MIT licence | Scanning free; rewriting $59 once, with 12 months of updates; you pay your model provider |
| Invisible characters in text | Removed: zero width characters, bidi controls, tag characters, variation selectors, noncharacters and more, keeping those that build emoji or spell a word | Removed, the same classes, keeping those that build emoji or spell a word; the count is shown under the Original |
| Look-alike letters (a Cyrillic letter in an English word) | Replaced in an aggressive mode | Named under the Original, never changed |
| Unusual spaces (thin, hair, ideographic) | Turned into ordinary spaces | Counted as spaces; left in the text |
| Statistical watermarks in the wording | A heavy rewrite, described in its own docs as best effort, with an optional loop that rewrites until its own test detector passes | Built to lower one without claiming it is gone, since no one can confirm a mark without its key. The app suggests rewriting with a different model family from the one that drafted, and the default model is open weight, so its maker has no sampler in the loop. Measured: the rewriting model's output scores below its own marked draft in 18 of 18 pieces in every measured cell, though where the app's guard keeps the draft the mark stays whole; it leaves it readable in many, more with some models than others; Repurpose, which turns a piece into a social post, lowers it further. Neither reliably removes it |
| Evidence published | A benchmark tool that reports clear rates, score changes, meaning drift and cost for runs you make yourself | Measured runs under a decision rule fixed in advance, with every figure published: three draft models, two rewriters, a length curve |
| File metadata (C2PA, EXIF, document properties) | Stripped across about twenty formats, from PNG and PDF to DOCX and MP4 | Not in scope |
| Images | Optional external scoring and removal tools | Not in scope |
| Why copy reads as machine made | clean-user-facing-text: a score from 0 to 1 that needs no model, built from sentence length variation, a weighted list of stock phrases and vocabulary diversity, with an audit mode that lists flagged spans | Highlights each phrase and pattern in the editor, grouped and explained, with a score; style measures are described, never scored |
| Your own writing | clean-user-facing-text: a voice sample you own guides the rewrite; the score is not calibrated to you | Calibrates the scan to pieces you wrote before AI tools and compares against them; a sample can also set the target for Repurpose |
| Voice and facts | clean-user-facing-text: written rules the agent follows (never invent a fact, flag a gap, never add a voice the source lacks), domain presets and a plain language mode; a score before and after as a gauge. Its benchmark measures meaning drift and whether numbers and links survive | Voices you can edit; checks the app runs itself on every rewrite: a facts check on numbers, dates, links and quotes, and a guard that keeps your original when a rewrite reads worse |
| A second opinion | Not in scope; its detectors check for test watermarks with a known configuration (above) | An optional audit by two models from families that wrote no version of the piece, since a model reads past its own family's habits; they quote any wording that still sounds generated. Like any reader, they cannot test for a watermark |
Measured on the same drafts
We ran watermarks-remover's rewrite (v0.7.0, its default paraphrase tactic, as shipped) and PlainSpeak's Rewrite with the same model, DeepSeek V4.1 Flash, on 54 drafts we had watermarked with our own key, and scored both the same way. watermarks-remover's paraphrase left about a third or less of the mark PlainSpeak's Rewrite left, by replacing nearly all the wording: it kept 4 to 9% of each draft's four-word runs, against 36 to 66% for the Rewrite. Set against the draft as a whole by a text embedding, its output reads 0.93 to 0.94 alike and the Rewrite's 0.98 to 0.99. That is a measure of similarity and not a check of meaning. Neither removed the mark reliably, and both kept every checkable fact. Each does what it is built for: one changes the words, the other keeps the writer's.
Where they agree
Both projects say the same thing about statistical watermarks, in their own words. A watermark of this kind lives in the choice of words, so only rewording moves it, and rewording changes the copy. watermarks-remover calls its rewrite best effort and says no tool can certify that a vendor's own check will fail. PlainSpeak's measurements say a rewrite's output lowers a test mark in every piece measured and removes it reliably in none. Neither claims more.
Both also advise rewriting with a model from a different family than the one that wrote the draft.
Using both
They fit together. Draft and finish the writing in PlainSpeak, then run the finished file through watermarks-remover if you also want its metadata stripped. PlainSpeak already removes invisible characters from what it exports, so running watermarks-remover's text cleaner afterwards should find little or nothing.