Best Humanizer Skills for Claude: Make AI Writing Sound Less Robotic

A humanizer Claude skill is a set of editing rules that tells the agent which habits make prose read as machine-written and how to rewrite them without changing what the text says. The best all-round choice in the directory is blader/humanizer, and a handful of smaller skills do narrower jobs well: a quick filter, a detect-only audit, a Python scorer or a voice profile.

This guide compares the English-language options from the humanizing AI text topic, then points to the language-specific ones. For each skill you get a "best for" line, what it needs, its licence and a trade-off. The descriptions come from each skill's own SKILL.md and repository. The editorial team did not run them on sample drafts, so nothing here is a claim about output quality.

What a humanizer skill can and cannot do

A skill is a text file the agent reads before it edits. It cannot see into a detector, and it cannot make an agent a better writer than the underlying model. What it can do is replace a vague request such as "make this sound natural" with specific, checkable rules: cut the throat-clearing opener, drop the emphasis crutch, vary sentence length, keep the claims.

That has two consequences. First, rewrites are only as good as the facts in your draft, so the better skills tell the agent to preserve names, numbers and citations. Second, these skills aim at clear prose. Several of them say so directly, and none reviewed here promises to beat an AI detector. Detectors misfire on human text and miss machine text, and disclosure rules at schools and workplaces apply whatever tool you use.

The directory also lists skills built to lower an academic paper's detection score or to strip watermarks and provenance metadata. Those are a different job with different risks, and this comparison leaves them out.

Best humanizer Claude skills compared

"Passed" means the static check reported no findings. It says nothing about how well a skill edits.

The original: blader/humanizer

Best for: one thorough pass over a long English draft, especially if you can paste a sample of your own writing.

blader/humanizer works from a ranked list of 26 patterns in five groups: staging, rhythm, inflation, formatting and leftovers. The list is drawn from Wikipedia's "Signs of AI writing" page, which WikiProject AI Cleanup maintains. The workflow has three steps: mark the tells by strength, draft a rewrite that keeps every supported claim, then audit the result for patterns that survived. If you supply a writing sample, the skill matches its sentence length, word choice, punctuation, openings and transitions.

Requirements: none beyond the agent. The repository is under the MIT licence and the skill was updated within the last few weeks.

Trade-off: it is the heaviest skill in this list, at roughly 8,000 tokens loaded each time it triggers. Its own text also warns that patterns are strongest in combination, and that a single appearance of a weaker pattern is not proof of anything. Expect it to over-edit short messages.

nousresearch/humanizer: a port with more rules

nousresearch/humanizer lives in the Hermes Agent repository under the MIT licence. Its SKILL.md says it is ported from Siqi Chen's humanizer at version 2.5.1, and it lists 34 patterns, adding five beyond the original's 29 at that version. It also adds an explicit step for putting opinion, varied rhythm and specificity back into the text.

This is a case of one skill appearing in two repositories, so be clear about which you want. If you use Claude Code or Codex and simply want the method, install the original. Choose the port if its extra patterns or its "add personality" step match how you write, or if you already run Hermes Agent.

Smaller skills for narrower jobs

hardikpandya/stop-slop

Best for: a fast filter that costs few tokens.

hardikpandya/stop-slop is MIT-licensed and loads around 650 tokens. The core file is backed by three reference files for banned phrases, structural clichés and before-and-after examples. Its rules cover throat-clearing openers, emphasis crutches, business jargon, adverbs, binary contrasts, passive voice and em dashes. It scores a draft on five dimensions, directness, rhythm, trust, authenticity and density, and asks for a revision below a set total.

Trade-off: the rules are blunt. A ban on all adverbs and all em dashes suits some writers and annoys others, and the repository has been quieter than the others here, with its last update in March.

petergyang/no-ai-slop

Best for: people who want to see what is wrong before anything is changed.

petergyang/no-ai-slop is MIT-licensed. It has an edit mode that lists each change, a detection mode that names patterns without touching the text, and a playful mode that writes slop on purpose. It checks more than 20 patterns, such as "It's not X, it's Y" contrasts and faux-insight setups, and it adds basics like leading with the point. An eval file in the repository gives you quality checks to run against it.

Trade-off: the pattern list is shorter than blader's, so it catches the loudest tells and misses subtler rhythm problems.

theclaymethod/unslop

Best for: writers who want to teach the agent their voice once and reuse it.

theclaymethod/unslop is MIT-licensed and routes through four sub-commands: teach builds a voice profile from your samples, cleanup flags tells as suggestions, rewrite diagnoses and rebuilds a draft, and mimic drafts in your established voice and then validates it. Detection runs in three layers, literal phrases, structural patterns and an arrangement score, with standalone Python scanners that need only Python 3.8.

Trade-off: four modes mean more to learn than a single rule list. If you only want a quick cleanup, the router adds overhead.

iflytek/zero-slop and oaustegard/declauding

iflytek/zero-slop audits and rewrites formulaic prose with a local Python scorer, and it offers inspect-only, rewrite and embedded-gate modes. It lives in iFlytek's skillhub repository, which ships a NOTICE file alongside the MIT licence.

oaustegard/declauding targets technical text rather than essays. It rewrites model-sounding prose into plain writing, then checks that every claim still holds, for pull request text, docs and commit messages. At about 5,000 tokens it is heavy for a narrow job.

guillaumemeyer/clean-user-facing-text

guillaumemeyer/clean-user-facing-text starts from an unusual angle: it audits text for invisible Unicode characters, then rewrites it while keeping facts, citations, code and required disclosures unchanged. It needs Python for the audit script. Use it when text is going to a system that chokes on odd characters, or when you want the stricter preservation rules.

Skills for other languages and for the agent's own replies

If you write in a language other than English, use a skill built for it. op7418/humanizer-zh covers Chinese articles and documents, beefiker/humanize-korean handles Korean translationese and rhythm, smixs/humanizer-ru adds a lint script and a detect-only mode for Russian, and coji/natural-japanese writes and edits Japanese business documents and can score how AI-like a text reads. All four are MIT-licensed and passed the automated check.

A different problem is the agent's chat replies. hexiecs/talk-normal installs always-on rules that cut filler, hedging and padded closings from what the agent says to you. It is not an editor for your drafts, and because it persists in your workspace configuration, read what it writes there.

How to choose

Match the skill to the job. For a long article or newsletter, try blader/humanizer, and add unslop's teach step if you want a reusable voice. For a lightweight guard on every draft, install stop-slop. For a review that changes nothing, use no-ai-slop in detection mode. For engineering text, use declauding.

Whatever you pick, compare the output with your draft line by line. A humanizer should remove patterns, not facts. For installation steps, see the pages for Claude Code and Codex, and for the basics of how skills load, read what Claude skills are. If your writing goes to social platforms, the LinkedIn and social media skills comparison covers drafting skills, and the top skills list shows what else is widely used.