Twenty-six patterns make the editing policy inspectable
Humanizer puts 26 prose patterns in one SKILL.md file. The strongest rules target staged openings, fake profundity, clipped closers, unraised objections, and the familiar construction that contrasts X with Y for drama. Other sections cover repetitive rhythm, inflated claims, vague sourcing, decorative formatting, chatbot residue, and replies that explain context the reader already has. You can read the whole editorial policy before letting an agent touch a document.
The skill has 3 output paths. Pasted text produces a draft, a critique, and a final version. File mode changes prose while leaving code, data, frontmatter, and link targets alone. Embedded use returns only the finished text for a larger workflow. A writing sample can override default choices about rhythm and punctuation, which is useful for voice matching and also makes results depend on the context supplied with the request.
The host model performs every editorial judgment
There is no rewrite engine in this 14-file repository. Humanizer asks the agent that loads the skill to find patterns, preserve claims, draft a revision, inspect that draft, and revise again. This lets the model rebuild an awkward paragraph instead of swapping individual words. It also means 2 runs can disagree about which sentence is filler, how much personality fits, or whether a qualification must stay.
That distinction sets the buying decision. Humanizer is easier to inspect than an opaque rewriting service, but it is less predictable than a linter with fixed rules. The agent model, surrounding conversation, writing sample, and permissions all shape the output. A team can edit SKILL.md to fit its house voice. It still needs a person who can notice when cleaner prose has become a different claim.
What happened when we ran it
Our sandbox installed commit 225a6f3 in 10 seconds, adding 35 packages and consuming 37 MB on disk. The build completed in 4 seconds. We used a fresh unprivileged Debian container with 3 CPUs, 8 GB of RAM, Python 3.12, and no secrets. Pip-audit reported 0 known vulnerabilities in the installed environment.
No test script or target was available, so our run skipped tests. The checkout contained 14 files, about 147 lines of source, and occupied 0.1 MB. We found 1 CI workflow, no Dockerfile, and no tests directory. Those figures show that packaging is small and quick. They say nothing about factual preservation, voice matching, or consistency across agents because the lab run did not measure those outcomes.
A 16-of-16 quality result came with a detector warning
Issue 229 describes a user-run blind comparison in which judges preferred Humanizer's rewrite to raw model output 16 times out of 16. The README cites that result. The study's author also reports that detection barely changed, including 100 percent AI scores from one commercial detector on 4 variants. Humanizer's README now says detector evasion is outside the project's goal.
The study has limits that matter. Its judges came from one model family, and infrastructure failures stopped the wider experiment after 77 of 144 planned pairwise trials. Still, its practical conclusion fits the product: evaluate whether the rewrite reads better and keeps the meaning. Do not use a Humanizer pass as evidence that a person wrote the result, or as proof that a detector will classify it differently.
Three v3.1.0 examples violate the preservation rule
Issue 306 checks the exact v3.1.0 commit from our lab and finds 3 before-and-after examples that add or change facts. One rewrite adds claims about camel meat and southern pasta use. Another changes more than 3,000 square feet into exactly 3,000 and replaces spaces with rooms. A third invents a count of 2 vendors and a reason their prices are missing. The issue was closed within minutes, with no comment shown in the issue thread.
These examples matter because models learn behavior from demonstrations as well as instructions. The skill explicitly forbids inventing names, numbers, dates, quotes, citations, and other details, then asks for a final claim check. Its own examples do not all meet that standard at commit 225a6f3. Review numbers, rankings, locations, qualifiers, and causal statements against the original. A smoother sentence is a regression when it changes what the writer knows.
September maintenance is active, while validation stays manual
GitHub showed 53,394 stars and 2 open issues and pull requests when fetched. Both open items were pull requests. Version 3.1.0 and the repository's last push landed on September 28, 2026, after a run of edits to examples, installation directions, and pattern definitions. The following day brought issue 306, so scrutiny continued after the release even though no later code push appeared in the repository facts.
The project is alive and unusually readable for an AI editing tool. Its evidence stops short of repeatable rewrite tests. Our 10-second install and 4-second build do not close that gap, and a star count cannot close it either. Use Humanizer for a second pass with version control and a visible diff. Choose Vale or another linter when the same sentence must trigger the same rule on every machine.