REhume
Remove AI writing tells from your prose
Paste a draft, see which phrases read like machine defaults, and get a cleaner version back. Detection and deterministic fixes run in your browser. Only flagged paragraph windows go to AI.
Voice calibration (optional)
Complete the security check below if prompted before humanizing.
What this tool does
I built REhume for a narrow editing job: strip the tells that make a draft sound like a default chatbot even when the facts are fine. Use it as an editing pass. If a policy requires disclosure, disclose. The output is a second draft you still have to read.
The rules come from the MIT-licensed blader/humanizer skill (v2.9.1), which itself follows Wikipedia's public "Signs of AI writing" guide. There are 33 numbered patterns: significance inflation, filler openers, synonym cycling, em-dash habits, copula avoidance, chatbot leftovers, and the rest. REhume is an independent site. It is not affiliated with the skill's author or with Anthropic.
What happens to text you paste
First pass is local. The browser flags matching patterns and applies free deterministic fixes (punctuation clutter, some formula phrases) without sending the document anywhere. If a span still needs judgment, the page cuts a short paragraph window and posts that window to /api/rewrite on Cloudflare Workers. The worker re-runs detection and refuses windows with no valid hit, so the endpoint cannot be used as a general LLM proxy.
Rewrites are capped: 8,000 characters of input, ten windows, tightmax_tokens. The model is instructed not to invent names, dates, citations, or facts that were not in the source. Voice calibration is optional: paste two or three paragraphs of your own writing if you want the rewrite closer to how you already sound. UK English style is an opt-in spelling pass after the rewrite, not a separate model.
Limits, parking, and privacy
The rewrite step uses Cloudflare's free Workers AI quota (10,000 neurons per day on this account). When usage hits the park threshold, the Humanize button stops until 00:00 UTC. Articles, detection, and autofix keep working. I would rather the site stay readable than serve broken AI calls.
I do not store pasted text after a request finishes. IP addresses are used for rate limits. Turnstile stops bots from draining the budget. Details live on the privacy policy and terms. Do not paste secrets, patient data, or anything your organization forbids sending to a third-party processor.
Who it is for
Use it on drafts you already intend to publish as your own work: blog posts, internal docs, product copy you will fact-check. Do not use it to submit coursework where AI assistance is banned, to launder generated text for a detector, or to rewrite other people's writing as if it were yours. I wrote When not to run a humanizer for the cases where the honest move is to leave the draft alone or start over.
If you would rather edit by hand, the article A manual pass over AI drafts is the checklist I use when I am not running the tool at all.
Guides on this site
The library is original prose about specific writing habits, not scraped Wikipedia dumps. Start with how the tool works, then pick a pattern you keep seeing in your own drafts.
- Boldface that does no work: Chat models bold every noun so the answer looks structured. I unbold everything, then put back at most one anchor per section.
- Challenges and future prospects sections I did not ask for: Outline furniture at the end of a draft: Challenges, Future Prospects, In Conclusion. I only keep a closing section when I have a dated risk or a next step.
- Honestly, look, here's the thing: Conversational openers are chat cadence pasted into documents. I start with the subject of the sentence, not a throat-clear.
- Emoji bullets that never belonged in the spec: Rocket-ship headings and checkmark lists are chat UI, not documentation. How I strip emoji without losing the actual steps.
- Inline-header lists that restate the heading: Bold label, colon, then the same idea in weaker words. I convert those vertical lists to sentences unless the labels are real field names.
- Name-dropping and fake notability: Cited in the New York Times, BBC, Financial Times, and The Hindu is not a biography. I keep one source with context and delete the fame stack.
All of them: article index. The numbered rules: 33-pattern encyclopedia. Short answers: FAQ. I am tboy1337; email contact@rehume.com or use the contact page.
Common questions
Do you store pasted text? No. Browser detection stays local. Rewrite windows are not kept after the request. Details: privacy policy.
Will this beat a detector? I do not promise that and I do not try to. Read why detectors are unreliable and the FAQ.
Why is Humanize idle? Empty input, or the daily neuron budget is parked until 00:00 UTC. Articles keep working.