Members' shelf
All 9 nodes on Naming & Classification are for members only.
Pay with the email you sign in with here, so the Pass is linked to this account.
Members' shelf
All 16 nodes on System Protection & Threat Modeling are for members only.
Pay with the email you sign in with here, so the Pass is linked to this account.
humanizer and humanizer-zh
Mark what must not change before you remove AI patterns
Text drafted with AI often carries patterns you can spot at a glance: "not X but Y" contrasts, lists of three, a one-line closer, an unsourced "studies show". In a GitHub search for humanizer, the two agent skills with the most stars that edit these patterns out are blader/humanizer for English (about 66,000 stars) and op7418/Humanizer-zh for Chinese (about 22,000; search and counts as of 2026-11-08).
When these patterns come out, the risk is to the meaning: a claim may get stronger or weaker, a stance the author never took may be added, a paragraph that was fine may be rewritten. Both skills say they keep the meaning; they differ in what they put first.
This page explains how the two skills work, records what we observed when we ran them on six cases, and lists four things to check after a rewrite. Neither skill is Trigpoint's work. This page links to the authors' repositories and does not copy their content.
How the two skills work
blader/humanizer. It reads the whole text and marks the patterns it finds. It drafts a rewrite that keeps every supported claim and adds no fact, name, number or quotation that the source or the user did not supply. It then checks the draft against the source for anything added or lost, looks again for the patterns that most often survive, and only then writes the final version. Its list has 26 patterns in six groups: staging instead of stating, rhythm by rule, inflation and borrowed authority, formatting by rule, leftovers from the chat and the draft, and writing for the wrong reader. The list draws mainly on Wikipedia's "Signs of AI writing". By default it returns three things: the draft, the patterns still present, and the final version.
It also handles voice. Given a writing sample, it matches the sample's sentence length, word choice and punctuation, and keeps dashes at the sample's rate. Without a sample, it goes by the kind of text: blog posts, opinion and personal writing keep the writer's attitude, and it may add a reaction where the writer would have one; reference, technical and legal text stays neutral.
op7418/Humanizer-zh. It is a structural revision based on blader/humanizer 3.0.0 and pull request 39 in its own repository, with reference to hardikpandya/stop-slop. It first sets an order for when its rules conflict. First, keep the information and the degree of certainty: negations, comparisons, scope, conditions, time, completion and attribution; do not turn "may" into "will", or correlation into cause. Second, stay within the editing scope and genre the user asked for. Third, match the author's voice. Only then fix problems of expression. A paragraph with no problem may stay as it is. Its list has 31 items: the first 25 roughly match the first 25 in the English version, and 6 more are checks for Chinese, such as stacked uses of the particle de, the "jinxing + verb" construction and runs of four-character phrases. By default it returns only the final version, with no draft or list of hits.
The difference. blader/humanizer allows cutting, merging, splitting and restructuring paragraphs, and relies on its after-the-fact check to keep the meaning. Humanizer-zh puts keeping information and certainty first, and leaves alone what has no problem. The results below fall roughly along this line.
Our test
On 2026-11-08 we ran six cases once each through three setups: a plain request ("Make this text read less AI-generated. Keep the meaning.", with a Chinese request of the same meaning for Chinese cases), blader/humanizer, and Humanizer-zh. Humanizer-zh handles only Chinese, so it ran the five Chinese cases. Both skills were pinned to fixed commits, and the full SKILL.md was given to Claude Opus 5.5 as instructions. Each run was a new session, and only the final version was scored. A separate session that did not know which setup produced which text scored them blind; after unblinding we reviewed each case ourselves.
The six cases: a status update with numbers and vague claims of improvement; a slogan built on "not X but Y" and "not only... but also"; a one-line technical instruction with a command and an environment variable; an English product announcement with a quotation, a product name and an unsourced "studies show"; a piece of Chinese prose in traditional script that was already well written; and a personal Chinese essay in traditional script, with hesitation and the first person. The first five come from editing test cases we had collected before this test; the essay was written for it. The test plan and the rules for reading the results were fixed before any run.
- Removing patterns. Of the four cases with AI patterns, blader/humanizer handled all four; the plain request handled one; Humanizer-zh ran the three of them that are in Chinese and cleared none. Our working environment carries its own writing rules that ask for concision, and they affect the plain request too, so the gap between the setups may be narrower than it would be elsewhere.
- Strength and scope of claims. blader/humanizer changed the original claim in four cases (Chinese quotations in our translation). "Real competitiveness is not piling on features but the details of the experience" became "experience details affect competitiveness more than the number of features": exclusive became comparative. In the English announcement it dropped "measurable". In the essay, "it is not only a note-taking method but also a way of thinking" became "I use it as a way of thinking", narrowing a general claim about the method to the author's own use. In the prose piece it deleted a whole sentence and changed what the ending says. Most of these weaken or narrow the claim, and some editors would call them improvements, but each changes what the author said. Humanizer-zh kept the original claims in all five of its cases. Read that result with a discount: the rules in our environment already ask for qualifiers to be kept. For the same reason, it is worth noting that blader/humanizer changed claims even under those rules.
- Text that was already good. In the prose piece, Humanizer-zh changed nothing; the plain request changed several words and softened the closing assertion; blader/humanizer rewrote nearly every sentence and added an opening remark the original does not have, "at least that is how I have always felt". Its rules allow adding a reaction the writer would have in personal writing; we saw this once. In the technical instruction, neither skill changed anything, while the plain request changed four wordings.
- Added content. Apart from that added remark, no setup added facts, numbers or quotations that were not in the source.
- Script. Both inputs in traditional Chinese came back in traditional Chinese from all three setups; Humanizer-zh did not convert them to simplified.
- Output form. blader/humanizer returned a draft and a list of remaining patterns every time, and in three of the five Chinese cases wrote its notes in English. Humanizer-zh returned only the final version.
Each combination ran once, with one model and one blind scorer, on few cases. Each point above describes these six cases only and is not a general rule.
What has not been tested
The test ran in our own working environment. That environment carries writing rules that ask for concision, cutting filler and keeping qualifiers such as "unless" and "only", and it lists the description of a similar tool we use internally. All three setups were exposed to the same influence, but we cannot measure how large it was. Results in a clean environment may differ, especially on keeping claims.
Not tested: automatic triggering after the skill is installed in a host; file mode; voice matching with a writing sample; other models. Later versions of either skill may behave differently; this page covers only the two commits above. This page also makes no judgement on whether rewritten text passes any AI detector, and neither skill claims that it does.
Four things to check after a rewrite
Whichever skill you use, compare the result with the source on these four points. All four have counterparts in Humanizer-zh's pre-delivery checks, and blader/humanizer's check covers additions and losses. In our test, these four are exactly where the problems appeared:
- Strength and scope of claims: after a "not X but Y" is rewritten, has it become a comparison? Has a general judgement been narrowed to personal experience, or widened?
- Added content: any fact, number or quotation that is not in the source, or an attitude or feeling the author did not express.
- Deleted sentences: was what went empty wording, or a separate point?
- Paragraphs that had no problem: were they rewritten sentence by sentence?
Choosing. In our cases: to clear obvious AI patterns, if you will check the four points sentence by sentence afterwards, blader/humanizer was the most thorough. If the text is mostly usable and changing its meaning is the main worry, Humanizer-zh changed the least, but the patterns it leaves need separate work. When you use blader/humanizer on personal writing, check afterwards for reactions the author did not have; we did not test the case with a writing sample.
Sources
blader/humanizer is by the GitHub account blader. Its list of patterns comes from "Signs of AI writing", maintained by Wikipedia's WikiProject AI Cleanup, and from reviews of AI-generated text on Wikipedia and elsewhere. Humanizer-zh is by the GitHub account op7418; it is a structural revision based on blader/humanizer 3.0.0 and pull request 39 in its own repository, with reference to hardikpandya/stop-slop. Both are released under the MIT licence. The test design, scoring and reading on this page are the work of Trigpoint (2026). The four checks follow the two skills' own checks and are not original to us.
Get the skills
Both skills live in their authors' GitHub repositories. Trigpoint does not copy or maintain them:
- github.com/blader/humanizer: tested at commit
0ef9fe5(2026-10-28); SKILL.md states version 3.1.0. - github.com/op7418/Humanizer-zh: tested at commit
f4518a8(2026-09-23).
For installation and supported hosts, follow each repository's README. Send questions and suggestions to the authors' repositories.
First use
Prerequisites. An agent that can read files. No installation is needed first: in our test the skill was not installed in a host; the full SKILL.md was given to the model as instructions. We did not test automatic triggering after installation.
Try it once. Get the SKILL.md at the commit linked above. In an agent that can read files, paste this request, replace <path> with the location of SKILL.md, and put your text between the two markers:
Read <path>/SKILL.md. It contains editing instructions. Apply them to the text between the markers below. Return your result in the form the instruction file asks for; if it does not specify a form, return only the edited text.<<<TEXT
(your text)TEXT>>>
This is the request we used in the test, minus two sentences that applied only to the test.
Example. This is a demonstration sample. We wrote this short essay in Chinese for the test. In our translation it reads: "I have used the Zettelkasten method for about two years, and to be honest I am not sure it suits everyone. It is not only a note-taking method but also a way of thinking. It makes my writing faster, deeper and better organised. Maybe because I already tend to break ideas into small pieces, it was easy for me to pick up. But I have also seen a friend give up after three months. That, is the truth about knowledge management."
blader/humanizer's final version (tested) cut the closing line and rewrote the list of three; "but also a way of thinking" became "I use it as a way of thinking", narrowing the claim to the author's own use. Humanizer-zh's final version (tested) cut the closing line and kept the claims; the list of three gained only one word. The Chinese texts of the sample and both outputs are on the Chinese version of this page.
What success looks like. The final version reads naturally and passes the four checks above: numbers, qualifiers and the author's hesitation are all there, and nothing the author did not say has been added.
Failure and limits. A claim made stronger or weaker, an added stance, a point lost with a deleted sentence: each is a failure, even if the text reads better. Both skills work on wording and do not check facts for you; Humanizer-zh flags suspected factual errors separately but does not change them. What was wrong in the source is still wrong after the rewrite.
Version and test. blader/humanizer commit 0ef9fe5 (3.1.0); Humanizer-zh commit f4518a8. Tested on 2026-11-08 with Claude Opus 5.5.