What does this skill do, and when should you use it?
Humanizer is a free, open-source agent skill that strips the machine smell from AI-assisted writing. It scans text against a catalog of 55 named AI patterns, reports a 0-100 AI-tell density score, and rewrites the piece in one of five voice profiles so it reads like a specific, opinionated person. The skill is pure Markdown with zero dependencies and no network calls. It ships with strict guardrails: it never fabricates facts, never rewrites quoted or code content, and explicitly protects non-native English and naturally low-variance human prose from false positives.
- Scans text for 55 AI patterns across six categories: content, language & style, communication, filler & hedging, emerging, and craft/forensic
- Computes a 0-100 AI-tell score from weighted signals (burstiness, vocabulary blacklist density, sentence-length variance) when --score is set
- Rewrites in five voices — casual, professional, technical, warm, blunt — injecting opinion, concrete detail, and sentence-length variation
- Edits Markdown files in place with minimal targeted changes via --file --mode edit
- Masks fenced code blocks and blockquotes (--ignore-code / --ignore-quotes) so they are neither flagged nor rewritten
- Supports iterative convergence (--iterate N), heavier rewrites (--aggressive), and content-type rule layering (--purpose essay|email|marketing|technical|general)
- A blogger preparing a post, README, or LinkedIn note for publication who wants it to read like a real human wrote it
- An editor auditing an existing document for AI tells before shipping, using detect mode for a prioritized, per-pattern report
- A technical writer cleaning API docs or READMEs with the technical voice, keeping precise jargon while cutting "delve" and "robust" boilerplate
- A content creator comparing N maximally-different opening hooks via --openings and picking the strongest
- A brand team dropping a humanizer-context.md into the project root with brand samples and banned phrases to steer every rewrite
- A power user baking the core rules into CLAUDE.md or a system prompt so the agent writes clean by default, not just on demand
- Laundering genuinely AI-generated academic or compliance text to evade institutional AI detection — the source frames the goal as writing better, not beating detectors
- Rewriting source code, config, or structured data (.js, ., .yaml, etc.) — edit mode explicitly refuses non-prose targets
- Scenarios that require an online detector or cloud API — this skill is fully local with no network calls by design
How do you install this skill?
- --mode edit modifies files in place with no built-in confirmation or rollback; back up files before use.
- The 0-100 score is a heuristic signal, not a verdict: the author's own cited research shows expert humans near coin-flip accuracy, with documented false-positive risk for non-native English and neurodivergent writers. Do not use it to gate or judge individuals' writing automatically.
- The Chinese appendix (ZH1-ZH15) is a provisional, unvalidated draft; Chinese-language effectiveness is not guaranteed, and ZH13/ZH14 are unverified hypotheses.
- Publisher identity is unverified; the pattern catalog will drift as alignment recipes change (the author discloses this), so monitor repository updates.
- This is a static review: nothing was executed; evals are manual and trigger accuracy varies by model.
- Local filesystem
Node.js (optional, only for the standalone metrics CLI)
Three routes depending on your agent environment:
Claude Code (plugin marketplace)
claude plugin marketplace add Aboudjem/10x
claude plugin install humanizer@10xAny of 70+ Agent Skills-compatible editors (Cursor, Codex, Copilot, Gemini CLI, etc.)
npx skills add Aboudjem/humanizer-skillManual install (curl, project-scoped)
mkdir -p .claude/skills/humanizer && curl -sL https://raw.githubusercontent.com/Aboudjem/humanizer-skill/main/skills/humanizer/SKILL.md -o .claude/skills/humanizer/SKILL.mdFor other editors swap the directory (.cursor/skills/, .codex/skills/, .gemini/skills/, etc.); use ~/.claude/skills/ for a global install. The optional metrics CLI needs Node.js; the repo also documents pre-commit hooks and CI exit codes.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- humanize this: [paste a blog draft that smells like ChatGPT] --voice casual
- make this sound less AI --file draft.md --score — show me the 0-100 score first, then rewrite
- does this read like ChatGPT? detect only, list every pattern hit with suggested fixes
- rewrite this README so it doesn't sound AI-generated, keep all numbers and code blocks untouched, technical voice
Inside Claude Code or another Agent Skills client, trigger it with natural language ("humanize this", "make this sound less AI", "does this read like ChatGPT") or the /humanizer command. The flow is: parse arguments, detect patterns, apply voice and rewrite craft, output per mode. Three modes: detect (report only, no rewrite), rewrite (default, full transform), edit (in-place minimal changes via the Edit tool). Key flags: --voice casual|professional|technical|warm|blunt, --file for file input, --score to prepend a [Score: NN/100] header, --iterate N to loop detect-rewrite-detect until convergence (max 3), --aggressive for heavier rewrites, --purpose to layer content-type rules, --ignore-code / --ignore-quotes to mask spans before scanning. An optional Node CLI provides score (per-signal breakdown) and compare --check-facts (fails if the rewrite lost a number, percentage, URL, date, version, or acronym), suitable for CI and pre-commit gates.
What are this skill's strengths and limitations?
- All 55 patterns have documented trigger lists and 34 carry before/after pairs — explainable, not a black-box score
- Zero dependencies, zero network calls, pure Markdown; README explicitly promises no telemetry, no account, no API key
- Hard no-fabrication guardrail: the Concretizer pass only uses details implied or stated in the source, and flags or asks when a detail is missing
- Detailed false-positive protections: clusters over isolated tells, protection for non-native English (citing Liang et al.), neurodivergent writing styles, and quoted/code content
- Optional fact-check (compare --check-facts) fails a rewrite that silently lost numbers, URLs, dates, or acronyms — CI-friendly
- The 0-100 score is a heuristic signal blend (burstiness, vocabulary density, trigram repetition); reproducible but not semantic, and the repo itself warns self-scoring inflates in the same session
- Zero-tolerance em dash policy (P13) may feel heavy-handed for authors who deliberately use em dashes
- Chinese support is explicitly provisional (references/patterns.zh.md); the catalog is built for English prose, so non-English effectiveness is unverified
- Some rules are community-reported with no controlled study (e.g., semicolon clusters flagged as LOW-MEDIUM confidence)
- Rewrite quality still depends on the underlying LLM; the skill supplies rules and process, not the model
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Humanizer — AI Writing Humanizer & Detector Skill this page | 67 · Recommended | ★ 282 | 1mo ago | MIT |
| Shuorenhua: Chinese AI-Tone Cleanup Skill | 65 · Recommended | ★ 2.1k | 25d ago | MIT |
| HumanAI — Text Humanization Engine | 52 · Use with care | ★ 38 | 3mo ago | MIT |
| FLOW Blog Workflow | 52 · Use with care | ★ 2.4k | 11d ago | MIT |
| GEO Citation Optimizer | 52 · Use with care | ★ 3k | 11d ago | Apache-2.0 |
The README credits Wikipedia's Signs of AI writing (CC BY-SA) as the source of the core pattern catalog (P1-P30) and cites Liang et al. (arXiv:2304.02819) on detectors misfiring against non-native English writers. Unlike hosted AI detectors, it is an explainable rule catalog plus a local heuristic score, and it pairs detection with rewriting — detectors only tell you it reads AI; this tells you where, why, and fixes it. The README names no specific commercial competitor.
How did FollowSkills review this skill?
Evidence shows: pure Markdown, least-privilege tool list (Read/Write/Edit/Grep/Glob/AskUserQuestion), declared zero dependencies and no network calls, SECURITY.md discloses threat model (prompt injection) and data flow. Deducted for: --mode edit rewrites user files in place without mandatory explicit confirmation, backup, or rollback path; humanizer-context.md is auto-read from the project root. Not full marks.
Evidence shows: self-consistent flag/mode definitions, evals. covering trigger and no-trigger cases, node --test suites for the CLI. Deducted for: static cap of 10; the submitted SKILL.md text appears truncated at P50, and CI greps for 'Voice Profiles'/'Soul Injection' sections not visible in the provided SKILL.md, raising a self-consistency question; evals are manual-only with no automated trigger harness; rewrite quality and failure feedback depend on the model and were not executed.
Evidence shows: explicit trigger phrases, counter-examples (should_trigger=false), thorough false-positive guardrails (quotes, code, neurodivergent writers, non-native English), local execution with no network. Deducted for: Chinese support is a provisional appendix the author states is unvalidated by a native writer, with burstiness/perplexity explicitly flagged as non-portable to Chinese and ZH13/ZH14 unverified; reproducibility detail of the --score rubric is thin at the skill level.
Evidence shows: layered docs (standalone SKILL.md plus on-demand references), MIT LICENSE, CI-enforced presence of CHANGELOG/FAQ/SECURITY/CONTRIBUTING, cited provenance (Wikipedia, HC3, arXiv), honest known-limitations section. Deducted for: unverified publisher means maintenance responsibility and update path rest on an individual repo; minor inconsistency between CI checks and the provided SKILL.md content.
Evidence shows: 55 patterns with 34 before/after pairs, three full worked examples (technical/blog/LinkedIn), a compare --check-facts check that fails rewrites losing facts, and a no-fabrication principle. Deducted for: static cap of 7; actual rewrite quality and the 0-100 score's behavior were not executed, and no quantified precision/recall evidence for pattern detection exists.
Evidence shows: multiple auditable sources (Wikipedia, arXiv 2301.07597, arXiv 2304.02819, HackerNews threads), a per-pattern coverage mapping against the Wikipedia guide, an Honest limits section disclosing that the source guide itself is contested, and license compliance notes (HC3 CC-BY-SA, no reuse of unlicensed code). Deducted for: static cap of 5; all scientific claims (HC3 statistics, ACL 2026 study) are uncorroborated by independent reproduction, and score behavior is only indirectly pinned by CLI tests, not the skill's rewrite paths.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →