メニュー
Humanize a document by removing AI-writing patterns via a multi-agent workflow (analyze → revise → adversarial review → loop-until-clean). Use when the user asks to "humanize", "de-AI", "remove AI tells", or "review for AI writing" on a file.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Detect AI writing patterns in text. Use when: (1) need to check if text sounds AI-generated, (2) reviewing content before publication, (3) analyzing writing for Wikipedia-style AI markers, (4) generating report-only analysis without changes.
Use when: (1) User asks to "review for AI writing" or "check for LLM signs" (2) User wants to "detect AI-generated text" or "humanize this document" (3) User wants to identify and fix signs of AI/LLM-generated writing (4) User mentions Wikipedia AI writing guidelines
Fix AI writing patterns in text. Use when: (1) ai-writing-analyzer found issues to fix, (2) need to humanize AI-generated content, (3) preparing content for publication, (4) iteratively refining text until clean.
Fix for Claude Code slash command frontmatter that fails to parse when `argument-hint:` is given square-bracketed placeholder values. Use when: (1) authoring `commands/<name>.md` files in a CC plugin and the frontmatter shows `argument-hint: [working-dir] [team-name]` or similar multi-token bracketed values; (2) a frontmatter-validation test (e.g. `yaml.safe_load(block)` over the SKILL/command frontmatter) raises `ScannerError` or `ConstructorError` complaining about flow sequences or "expected a mapping"; (3) the slash command silently doesn't appear in `/help` after install, or a plugin-marketplace test asserting the `description` field is present trips because the whole frontmatter parse blew up. The fix is to QUOTE the value as a single string. Covers the same trap for `commands/<name>.md` directly authored in CC format AND APM-managed marketplaces shipping commands via APM.
Organize GitHub repository issues with proper milestones, native relationships, and clean labeling structure. Use when grooming issues, setting up milestones, establishing dependencies, or cleaning up issue tracking in a GitHub repository.
Enable Claude Code to delegate specific tasks to Amazon Q CLI agents running in the background with full tool permissions. Use when bridging to Q's specialized integrations such as Quip, ticketing systems, or diagram generation that Claude doesn't have direct access to.
| name | ai-writing-humanizer |
| description | Humanize a document by removing AI-writing patterns via a multi-agent workflow (analyze → revise → adversarial review → loop-until-clean). Use when the user asks to "humanize", "de-AI", "remove AI tells", or "review for AI writing" on a file. |
| author | Raul |
| version | 1.0.0 |
| tags | ["writing","humanizer","ai-detection","editing"] |
| level | user |
This skill launches a multi-agent Workflow that humanizes a document: it analyzes AI-writing tells across three lenses, rewrites fidelity-first, then runs four adversarial reviewers (fidelity / residual-tells / quality-voice / gap) and loops until clean. It auto-fixes critical/high issues and surfaces medium/low (and anything the fidelity guard blocks) as prioritized tasks.
A full pass dispatches many agents and is token-intensive. Before running, confirm with the user: which file, and whether to proceed.
Resolve the target. Get the file path from the user. If they pasted text instead, write it
to a scratch file first (the engine works on a document). Set
ROOT="${CLAUDE_PLUGIN_ROOT}/skills/ai-writing-humanizer".
Pre-scan (deterministic). Run:
node "$ROOT/workflows/prescan.js" "<file>" "$ROOT/patterns/patterns.json" "$ROOT/config/default.config.json"
Capture stdout as mechanicalFindings (a JSON array).
Assemble args. Read <file> → text; read $ROOT/patterns/patterns.json → patterns;
read $ROOT/prompts/analysis-prompt.md, suggestion-prompt.md, verification-prompt.md →
prompts = { analysis, suggestion, verification }; read $ROOT/config/default.config.json →
config.
Run the engine:
Workflow({
scriptPath: "$ROOT/workflows/humanize.js",
args: { text, patterns, prompts, mechanicalFindings, config }
})
Write the revision. Write the returned revisedText to a sibling <name>.humanized.<ext>
(never overwrite the input). Show the user a diff of <file> → the revision.
Print the on-screen report (no file). Two sections, grouped by priority:
✔ Fixed — counts from fixedByPriority.⚠ Still flagged — the residual findings (span + why), grouped critical/high/medium/low.
Include the fidelity and quality verdicts.Create residual tasks. For each finding in residual, create ONE task-list entry, its
priority = the finding's priority. If residual is empty, create no tasks — say it's clean.
If the Workflow tool is unavailable in this harness, do not fail. Instead run the same phases inline: dispatch the 3 analyze lenses and 4 reviewers as subagents (Agent tool), do the revise step yourself, and apply the same loop-until-clean logic, then produce the same report + tasks.