AI humanizer
Make AI-assisted writing clear and direct for your reader. Paste a draft to find the patterns that get in the way, each marked in your text with a fix that keeps to the facts you already wrote.
Your text with each tell marked
Make it clearer with your own AI
Copy a prompt that asks for clearer, more direct versions of the marked lines, using only facts already in your text. Paste it into ChatGPT, Claude or Gemini.
8 sentence patterns give most AI drafts away
A blind judge read 24 drafts written by Claude and counted 212 AI writing tells. 187 of them came from 8 patterns. Overused words such as "delve" and "testament" made up only 4, so a word list on its own misses most of what readers notice. How to build your own humanizer skill turns each pattern into a check and a rewrite rule for your agent.
| Tell | An example from the drafts | Tells counted |
|---|---|---|
| Forced triads | It is cheap, it is reversible, and it produces evidence rather than opinions. | 63 |
| Not X but Y | Four Days Is Not a Perk. It's a Strategy. | 27 |
| Bold labels on list items | Better forecasts. We're investing in improved wave and tide modeling. | 27 |
| One-line closers and fragments | $299 per month. That's it. | 26 |
| Comment tails ending in -ing | Chair utilization improved dramatically, allowing the practice to serve 15% more patients without extending hours. | 14 |
| Inflated significance | I haven't experienced any hot spots or blisters across all those miles, a testament to the thoughtful design. | 12 |
| Sales words | Fiddle-leaf figs are stunning houseplants that require periodic repotting. | 9 |
| Arguing with no one | It isn't perfect. The tongue is slightly thin and can shift if you don't lace carefully. | 9 |
The humanizer files each line it marks under one of these, and puts the rest (dashes, emoji, stock openers, chatbot leftovers) under other AI patterns. Counts are the mean of three judge runs, rounded.
Every fix keeps to the facts in your text
Ask a model to make a draft sound human and it adds what a human would have written: a story, a customer name, a reason, a number. In the same test, a plain "make it sound like a person wrote it" prompt added 54 claims the drafts never made.
So the fixes here never ask you for something new. A sales word gets a plainer swap or a cut. Inflated significance gets cut, or replaced with a fact the text already states. When a fix needs a fact you do not have, delete the phrase. The draft is usually better for it.
Rewrite the marked lines yourself or with your AI
Most drafts need a handful of edits, and doing them by hand takes a few minutes. Start with the lines marked as a fix, then read each line marked as your call and change it only if it fails the question.
To hand the job to a chat assistant, use Copy rewrite prompt. It asks the model to make each marked line clearer and more direct, to leave other lines alone, and to add no new names, numbers, dates or quotes. Paste the result back here to check it again.
Run the same checks in your agent with seo lint
This page runs the AI pattern rules from seo lint, the writing check in the SEO Skill CLI. Your agent can run it on a draft, fix each finding and check again until nothing is left. Install the CLI and run setup:
npm i -g seo && seo startThen ask your agent:
Check drafts/launch-post.md with seo lint and fix every error and warning. Use only facts already in the text. Run seo lint again until it reports 0 errors and 0 warnings.The seo lint docs cover presets for blog posts, landing pages and meta descriptions, and how to keep a flagged line on purpose. What AI slop is and how to remove it explains the research behind the patterns. You can also add the hosted MCP server to your agent and sign in with Google:
Run it in a terminal, then sign in with Google.
Or run it on your machine
Installs the CLI, signs you in and adds the skill to your agents. Site crawls and Bing data run this way, on your own machine. A crawl fetches many pages, and Bing blocks our hosted server.