humanizer
by blader · linked, not rehosted
33 named patterns of AI writing, with a before and an after for each, in one markdown file.
Add it to your project
Give your agent this prompt — it sets the repo up for you:
Install humanizer with `npx skills add blader/humanizer --global`, then read my last three blog posts as a voice sample and humanize CHANGELOG.md in file mode.
Open source (MIT) · all rights to blader.
What it's for & howRead it ↓Collapse ↑
What it is
A single SKILL.md with 33 numbered patterns of AI writing, each one carrying a before and
an after. The list is not invented: it comes from Wikipedia's "Signs of AI writing", the page
WikiProject AI Cleanup maintains from thousands of real cases on Wikipedia. That provenance
is the reason this beats the usual "make it sound human" prompt, which is a vibe.
When to reach for it
- Whenever an agent writes something a person will read: commit messages, PR bodies, docs, release notes, changelog entries.
- After a draft exists, never before. It edits prose, it does not decide what the prose says.
- When your writing has to sound like you, not like a generically de-slopped voice. Hand it a sample and the sample wins.
The part most people miss
Two things, both in the file, both skipped by nearly everyone:
Embedded mode. The skill names three ways it can be invoked, and the third one is
"Another task or agent is using this skill as one step of a larger job." Wired into your
CLAUDE.md, it stops being a thing you remember to use and becomes the last step of every
piece of writing your agent produces.
Voice calibration. "A sample outranks this skill's style rules, including the em dash rule in §14." Give it three paragraphs you actually wrote and it matches your habits instead of enforcing its own.
The hardest rule in the file is §14: "The final rewrite contains no em dashes (—) or en dashes (–)." Not a preference, a constraint, with a scan before the final answer.
Honest limits
- It ships no checker.
scripts/validate-package.pyvalidates the package, not your text. The model that wrote the rewrite is the same one grading it. If you want a number, count the tells with your own script. - The em dash ban is a style position, not a fact. Plenty of good writers use them. The skill knows this and lets a voice sample override it, but the default is absolute.
- Over-editing is a real risk. The file carries a "what NOT to flag" list for exactly this reason: polish is not proof of AI. Prose that was already human comes back flatter.