Questions and answers
The short version: it makes Claude’s prose read well to a person, it does not change AI-detector scores, and it’s free.
What is humanizing-writing?
It's a skill for Claude Code that guides how Claude writes prose: docs, READMEs, reports, emails, blog posts, pull request descriptions. It loads whenever Claude is about to write a paragraph or more for a human reader, so the draft comes out without the usual machine habits and there's no separate rewrite step. You can also hand it a finished draft and ask Claude to humanize it. It's free, open source under the MIT licence, and at version 2.1.0 as of October 2026.
Does it get text past AI detectors such as GPTZero or Pangram?
No, and we'd rather you read that here than find out later. Current detectors are trained classifiers. Xu et al. (2026) found that GPTZero and Pangram often pass text from base models as human and flag text from the instruction-tuned versions of the same models, which suggests they track the marks of instruction tuning itself. A style guide read by an instruction-tuned model doesn't change that tuning. Pangram also reports catching the output of nineteen humanizer tools more than 90% of the time, though that is the vendor's own figure. If you need a detector score, this is the wrong tool. It's built for the person who reads the text.
How is it different from blader/humanizer and other humanizer skills?
Two things, mainly. The widely used ones (blader/humanizer, avoid-ai-writing, stop-slop, no-ai-slop) are rewrite tools you run over a finished draft, and this one applies while Claude is writing. It also has a Scope section that says where the rules stop: API reference, legal text, fiction, marketing copy, someone else's writing, your own voice. And it's weaker than blader's in one clear way — blader reports a blind preference test of 16 out of 16, and ours is 18 pairs over two rounds, judged by Claude models.
How do I install it?
Run npx skills add AshwinSathian/humanize-writing-skill in a terminal. Or, inside Claude Code, run /plugin marketplace add AshwinSathian/humanize-writing-skill and then /plugin install humanizing-writing@humanize-writing-skill. You can also clone the repository and symlink it into ~/.claude/skills/humanizing-writing. None of the three needs an account or a review.
Does it work with Codex, Cursor, or models other than Claude?
We don't know. The file is a standard SKILL.md, so an agent that reads that format can load it. But it was written against the habits of Claude models, and every test in the repository ran on Claude Haiku, Sonnet, or Opus. Nothing here says how it behaves on another vendor's model.
Does it ban "delve", em dashes, or other words?
No. Word lists go stale. Wikipedia's list of AI vocabulary for mid-2025 onward is four words long, and the same page records that "delve" fell sharply in 2025. The rules here are about claims, sentence shape, and endings. An em dash is fine where a comma would hide the break, and a precise term such as "idempotent" always stays.
What do AI writing tells look like in 2026?
They've moved. In July 2026 The Economist compared 55,940 sentences from ChatGPT, Claude, Gemini, and Grok with human journalism and novels. It found long words, long sentences joined with "and", thin punctuation, "not X but Y", and lists of three. Only Claude used more em dashes than human writers. Reinhart et al. (PNAS, 2025) measured the trailing "-ing" clause at 2 to 5 times the human rate.
Will it rewrite my own writing or flatten my voice?
It shouldn't. When Claude edits text you wrote, the skill tells it to fix what you asked for and leave your hedges, fragments, and rhythm alone. When Claude drafts in your voice from a sample or a voice profile, your habits outrank the rules — if you use dashes and rhetorical questions, the draft does too, at about your rate. The one rule that never yields is the one against making things up.
How was it tested, and how far should I trust the result?
Twice, both times blind. Fresh Claude instances wrote three pieces with no skill, with version 1.1.1, and with the current rules, and model judges compared shuffled, unlabelled pairs. With Sonnet and Opus writing, each of two judges preferred the skill's text to the no-skill text in 5 of 6 pairs. With Haiku writing, three judges each preferred it in 3 of 3. Those are small samples and the judges are Claude models. The rules were also revised after earlier drafts did badly on the same three tasks, so the result is partly fitted to them. The pairs, the keys, and the losses are in the repository.
Does it invent facts to make writing sound specific?
It's told not to, in the one rule that has no exceptions, and it mostly holds. The rule was earned the hard way: version 1.x of the skill's own examples added a database connection pool and a Retry-After header that the originals never mentioned, and an early 2.0.0 draft produced an invented incident from "last quarter". After the rule was widened, no judge flagged an invented fact in a 2.0.0 passage from Sonnet or Opus. It does slip, though. In a later test on four new tasks, 2 of 9 passages written with the skill said something about the reader's team that nobody had supplied, such as "We currently write postmortems for the incidents that hurt, and skip the ones that were over quickly". A reworded rule brought that to 1 of 9, short of the bar set before the test, so the rule was left as it is. Check any claim about your own team or system before you publish.
Does it work for languages other than English?
Partly. The rules on claims, structure, and endings carry over. The English word examples don't, and every test so far was on English text.
How do I switch it off for one request?
Tell Claude. An explicit instruction from you overrides the skill's defaults. Ask for buzzwords or a mandated template and you'll get them, and they won't be changed back afterwards.
Was this site written with the skill?
Yes. Every page here was drafted by Claude with the skill loaded and my voice profile applied, and each figure was checked against the files in the repository. The profile has known misses. Drafts come out with shorter sentences and fewer dashes than I write, so where a page reads a little too tidy, that's why.