blader/humanizer review

blader/humanizer is a free, MIT-licensed agent skill with 65,200 GitHub stars on 9 November 2026 that rewrites AI text against 26 named patterns and, by its own README, does not aim to pass AI detectors. This blader humanizer review puts the open-source skill on the same fixed fields as every hosted tool in the full humanizer ranking under one published method: verdict, best for, main limitation, limits, modes and what it changes. Installation is covered in the Claude humanizer skill guide and is not repeated here.

What does the README claim?

The README, read 9 November 2026, makes three kinds of claim. The first is about method: the skill is built on Wikipedia's "Signs of AI writing" page, the list Wikipedia editors use to catch machine text, and checks for 26 patterns grouped into six families, from "not X but Y" constructions and one-line closers through forced triads, inflated significance, decorative bold and chatbot residue. The first five patterns justify an edit on a single sighting; patterns marked weak alone count only when several share a passage.

The second is about meaning: the skill marks every tell it finds, drafts a rewrite, checks the draft against the patterns and the original claims, and does not invent a name, number, date or quote. If a sentence needs a detail that is missing, the README says the skill asks rather than inventing one. That is the exact behaviour the method's meaning rubric rewards.

The third is about detectors, and it runs against the whole category: "Getting past AI detectors is not a goal, and detectors still flag most of its output." The README also reports a blind test in which judges preferred the skill's rewrite over the original AI text 16 times out of 16, linked to a repository issue, read 9 November 2026. That is a readability result from the maker's own test, not a detector result, and this page records it as such.

What does the research say about editing for readers versus detectors?

The published research explains why those two goals come apart. Krishna and colleagues showed that a paraphrase model dropped DetectGPT's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv, read 9 November 2026). Heavy paraphrase moves detector readings; it also rewrites sentences a human editor would have left alone. Weber-Wulff and colleagues tested 14 AI text detectors and found every one below 80 percent on their accuracy measure, with accuracy falling further on paraphrased text (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026).

A rule list that removes "let that sink in" and a forced triad changes far less of a text than a paraphrase model does, so a detector trained on statistical regularities may still see the original. The skill's maker accepts that trade. No detector score should be the only evidence in a misconduct case.

Why review a skill on a page about humanizers?

Because people searching for a humanizer find it: the repository's own topics include ai-humanizer and humanize-ai-text, read 9 November 2026, and it has more stars than most hosted tools have published reviews. It belongs in the same table with the same fields so a reader can see the trade, and the other open-source options are compared in the open-source humanizers guide.

What does blader/humanizer change in your text?

The README's before and after samples show the edits: a product announcement loses its rocket emoji, its "It's not just a feature" pivot and its "Let that sink in" closer, and keeps every fact about shared drafts and anchored comments. A travel post loses "completely stole my heart", "hidden gem" and the closing sparkle, and keeps the hotel, the tram, the tart and the castle queue. Personal writing keeps the writer's opinions; technical prose stays neutral. Pointed at a file, the skill changes only prose and leaves code, data, front matter and link targets alone, read 9 November 2026.

The site's protocol runs each tool on its default mode and, where a stronger documented mode exists, that mode as a second row. The skill has no modes, so a published run will show it once, run through one named agent and model recorded in the raw file, since the same instructions produce different rewrites on different models.