Claude humanizer skill: what blader/humanizer does and how to install it

A skill is a folder with a SKILL.md file: a name, a description of when to use it, and instructions the model reads when the task matches. Claude's developer documentation describes skills as filesystem-based resources loaded in stages, metadata at startup and the full instructions only when triggered, so many skills can be installed without filling the context window (Agent Skills overview, Claude platform documentation, read 9 November 2026). The same documentation lists custom skills as available in Claude Code, through the API and in claude.ai settings, and warns that a skill from an unknown source should be audited before use because its instructions can direct the model to run code or call tools.

A humanizer fits this shape because the job is a fixed list of edits applied to whatever text is pasted. Writing the list once as a skill means the rules load on demand, stay the same between sessions and can be read by anyone before they trust the output.

The README describes one account of why AI text reads the way it does: a language model writes the most likely next words for the widest range of readers, where a person chooses for one reader and one subject. Every tell the skill looks for is a form of that default. The skill marks each tell it finds, strongest first, drafts a rewrite without treating the original structure as fixed, checks the draft against the patterns and the original claims, then writes the final version (blader/humanizer README, read 9 November 2026).

Two rules in the README matter more than the pattern count. The skill does not make things up: a name, number, date, quote or citation must come from the source or the writer, and if a sentence needs a missing detail the skill asks instead of inventing one. And the skill edits for human readers: the README states that getting past AI detectors is not a goal and that detectors still flag most of its output. No hosted humanizer on this site publishes an equivalent sentence.

The 26 patterns are grouped in six sections, numbered by strength and frequency. The first five justify an edit on a single sighting; patterns marked weak alone count only when several appear together.

The README gives four routes, read 9 November 2026. In Claude Code 2.1.142 or newer, two plugin commands add the marketplace and install the plugin, which answers to /humanizer:humanizer. In Codex, one npx skills add command with --agent codex. In Claude.ai and Claude Desktop, download the repository as a ZIP and upload it as a skill in Settings. For any other agent the Skills CLI supports, the same npx command with --agent '*', or copy SKILL.md into the agent's skill folder by hand.

Once installed, the skill answers to /humanizer followed by the pasted text, or to a plain request such as "humanize this text". Pointed at a file, the README says it changes only the prose and leaves code, data, front matter and link targets alone. When text is pasted, it shows its work: a first rewrite, a short critique of anything that still sounds artificial, and the final version.

Two cautions sit outside the README. Claude's documentation says to treat a skill like installed software and to audit every bundled file before use; this repository is one SKILL.md plus packaging, which makes that audit short. And a skill runs on whatever model your agent runs, so the same file gives different rewrites on different models and on repeated runs.

The skill and the hosted tools perform the same operation, a model rewriting text, and are sold on different promises. Hosted vendors describe the output in terms of detector readings; the skill's README refuses that frame and describes the output in terms of readers. On the fields this site uses, the skill has no request cap beyond the agent's context window, no monthly pool, no price and no vendor page to check, and it has no published figure for meaning kept either, which is also true of every hosted tool.

The published method rates meaning kept at 35 of 100 points, errors introduced at 20, detector pass rate at 30 and limits and access at 15, with two editors reading each output blind. A skill whose stated goal is readability for humans and whose rule list forbids invented facts is built for the first two of those four fields. Whether its output holds a qualifier or a date better than a hosted tool is a question for a run, and no run has been published yet. No detector score should be the only evidence in a misconduct case.

• Use a pasted prompt if you humanize once in a while and want nothing installed; the ChatGPT humanizer prompt guide gives three in full.
• Use the blader/humanizer skill if you already work in Claude Code, Codex or another agent, want the rules in a file you can read and edit, and accept that detector readings are not what it is for.
• Use a hosted tool if you want strength modes, a per-request cap you can plan around and a vendor to hold to a limit; the plan facts are on the ranking and the free list.

Other open-source options, including a Python pipeline and a second rule-based skill, are in the open-source AI humanizers guide. Students should check their institution's AI policy before using any of them; a skill does not make undisclosed AI use acceptable. The rest of this series is in the guides hub.