HumanAI — Text Humanization Engine
Rewrites AI-generated text to read as human-written across 9 languages, stripping machine fingerprints through a five-stage pipeline.
Pure text skill: no script execution, no credential access, no network exfiltration; README states explicit ethics rules ([VERIFY] flags, no legal/medical rewriting, declared changes). Deducted: SKILL.md instructs the LLM to fetch external content via GitHub raw URLs, making the data flow partially opaque; no explicit user-confirmation mechanism.
Instructions are largely self-consistent and KNOWN_LIMITATIONS proactively discloses LLM counting unreliability, nondeterminism and self-assessment bias — good failure-feedback awareness. Deducted: SKILL.md is truncated mid-file (Stage 2 incomplete), depends on shared/ files and raw-URL availability; no execution evidence, so static cap of 10 applies.
Scenarios (15 playbooks), 7 tone profiles, boundaries (short texts, legal/medical, poetry) and trigger conditions (pre-flight guard) are clearly described. Deducted: no Chinese support declared — 9 languages exclude Chinese; the pre-flight heuristic scoring precision is promised but unvalidated.
MIT LICENSE, CHANGELOG, bilingual README, architecture tree, CONTRIBUTING and KNOWN_LIMITATIONS give good layering. Deducted: version signals conflict — SKILL.md/README claim v4.0 while CHANGELOG stops at v3.0; PLAN.md is dated 2026-07-12 (future date); raw links reference 'Natural-skill' while the repo is HumanAI, and the update path depends on an unverified individual publisher.
README provides concrete input/output examples and the pipeline design is specific (specificity ladder, skip policy); the core task is conceptually achievable. Deducted: a prompt-only skill whose outcomes depend entirely on the underlying LLM; static review has no execution evidence and self-eval scores (e.g. 94/100) are self-assessment, so capped below 7.
A CI workflow (validate.yml), tests/benchmark directory and ZeroGPT external-validation design exist, offering auditable material. Deducted: no actual test content or run results visible in this review (zerogpt-results. not provided), and CI only checks file presence, not quality — capped at 5, scored 4.
- The skill supports 9 languages but not Chinese, and no Chinese output quality is declared; Chinese users must assess on their own.
- SKILL.md directs the LLM to fetch extended data via GitHub raw URLs; raw.githubusercontent.com may be unreachable from mainland-China networks, potentially breaking deep mode.
- Effectiveness depends entirely on the underlying LLM with no determinism guarantee; the author concedes in KNOWN_LIMITATIONS that self-eval scores are not independent verification.
- Version signals are inconsistent (v3.0 vs v4.0) and external links reference another repo name (Natural-skill); verify actual file availability before use.
- External validation relies on the ZeroGPT API (overseas service), requires a user-supplied key, and the service's judgment of output quality is itself contested.
What does this skill do, and when should you use it?
HUMAN-AI is a version 4.0, MIT-licensed skill that loads into any LLM as a system prompt. It targets text that 'smells like AI' and processes it through a pre-flight → cleanup → specificity → tone → rhythm → proofread pipeline: it estimates AI probability, detects the language, deletes boilerplate openers, burned words, fake transitions and conclusion regurgitation, upgrades vague claims into concrete details, assigns one of 7 tone profiles, and breaks the metronomic sentence rhythm. It covers English, Russian, Ukrainian, German, French, Spanish, Portuguese, Italian and Polish, each with dedicated AI-marker and burned-word lists. SKILL.md is self-contained and works standalone; the repo also ships 15 task scenarios, annotated examples, and optional ZeroGPT-based external validation.
At runtime it reads the input text and runs a pre-flight scan: detect the dominant language (asks the user if confidence is below 70%) and score AI probability from a marker table (scores below 20 with no other markers trigger a STOP and diagnostic-only output). Then five stages execute in order: Cleanup deletes throat-clearing openers, conclusion regurgitation, burned words, fake transitions, hedging, fake balance, empty intensifiers and rhetorical padding using per-language tables, and always replaces em-dashes; Specificity moves claims up a 5-rung ladder (pure abstraction to consequence-stated) using mechanism reveals and number injection, flags unverifiable numbers with [VERIFY], and never invents facts; Tone selects among expert/biz/human/social/landing/article/case profiles and calibrates them against a per-language cultural matrix; Rhythm enforces clause-based sentence length variety, opener rotation and asymmetric paragraph weight; Proofread runs language-specific residual scans and outputs the result with a [PIPELINE] header, quality score and changelog. Optional PowerShell/Bash scripts call the ZeroGPT API for independent detection scoring.
- A content marketer inherits an obviously AI-drafted blog post or SEO article and needs it to read as human-written and avoid AI-detector flags
- A localization team fixing translated copy (e.g. Russian-to-English or into German/French) so it sounds natively written via the translation-fix scenario
- An indie developer installs it as a skill in Claude Code or OpenCode to batch-clean landing pages, emails and press releases
- A B2B writer wants case studies with trustworthy numbers and a 'first attempt failed' honest narrative structure (case tone)
- A team wants diagnosis only: Audit mode scans a text and flags all AI patterns without rewriting
- A social media manager converts posts into short, fragment-heavy scroller style for LinkedIn/Twitter (social tone)
What are this skill's strengths and limitations?
- SKILL.md is fully self-contained and works with any LLM that accepts a system prompt — no platform lock-in
- Per-language depth: 9 languages each get their own AI markers, burned-word lists and cultural norms (e.g. de-Russification for Ukrainian, anti-Nominalstil for German), not machine-translated English rules
- Explicit ethics: no invented facts or statistics, mandatory [VERIFY] flags, and refusal to rewrite legal/medical/safety-critical text
- Three-layer quality validation: built-in quality score, a separate EVAL.md LLM evaluation, and optional ZeroGPT API external detection
- Documented joint-pipeline protocol with the same author's RankWise (SEO) and MindFluence (cognitive-bias marketing) skills, preserving heading structure and keyword density when humanizing SEO content
- No published, systematic detector-comparison data backs the 'beats AI detectors' claim; the repo offers only an optional ZeroGPT benchmark requiring your own API key
- External validation and integrity scripts depend on PowerShell or Bash, adding setup overhead outside a pure-prompt environment
- Rewrite quality depends heavily on the underlying model's ability to follow dense rule tables; weaker models may miss or over-apply deletions
- The specificity stage permits model-generated 'domain-typical' example numbers — even with [VERIFY] flags, users must fact-check each one
- Pre-flight scoring is explicitly unreliable on texts under 100 words and is skipped there
- The full shared/ assets (30+ word burned lists per language, 250-line tone profiles, cultural matrix) are not reproduced in the source shown; standalone users only get the embedded top-15 lists
How do you install this skill?
Three options: 1) OpenCode — copy the repo folder into project .opencode/skills/ or global ~/.config/opencode/skills/; 2) Claude Code — copy into ~/.claude/skills/; 3) Any LLM standalone — paste the full contents of SKILL.md as a system prompt; it contains the complete pipeline. Optionally add the shared/ files for richer per-language detail. Optional: set ZEROGPT_API_KEY to run the external detection benchmark scripts. The docs do not document Windows-specific setup steps or dependency versions.
How do you use this skill?
After installation, issue natural-language tasks. Full pipeline: "Rewrite this text to sound human. Language: ru." Scenario-specific: "Rewrite this as a landing page. DE." Single stage: "Just clean the AI patterns from this." Audit only: "Tell me what's wrong with this. Don't rewrite." Translation fix: "This was translated from Russian to English. Make it sound native." Output includes detected language, tone, declared pipeline stages, quality score and a changelog.
How does this skill compare with similar options?
The source names no external competitors, but defines two companion skills by the same author: RankWise (SEO content engine) and MindFluence (cognitive-bias marketing), which chain as RankWise → MindFluence → HumanAI. If you only need humanization, this skill alone suffices; for a full SEO or marketing content production chain, use them together.