AGENT SKILL

alpha-humanizer

Transform AI-generated content into prose that reads as 100% human-written. Use when the user wants to "humanize" text, make AI content undetectable, rewrite content to sound human, bypass AI detectors, remove the "AI voice," write with authentic human voice, or produce content that passes as organic human writing. Also triggers on "rewrite this to sound human," "make this not sound like AI," "humanize this," "remove AI patterns," "write like a real person," or any request to eliminate synthetic writing patterns. Handles blog posts, LinkedIn posts, sales copy, emails, essays, technical docs, social media, landing pages, product descriptions, and thought leadership. Does NOT just swap words — it restructures the cognitive architecture of the text.

适用职业内容创作者
使用场景写作与文案
License 许可
MIT
所属仓库
marcelfahle/skills
最近更新
2026-11-06
Stars
★ 0
Forks
0

01 前言 Front Matter

name
alpha-humanizer
description
Transform AI-generated content into prose that reads as 100% human-written. Use when the user wants to "humanize" text, make AI content undetectable, rewrite content to sound human, bypass AI detectors, remove the "AI voice," write with authentic human voice, or produce content that passes as organic human writing. Also triggers on "rewrite this to sound human," "make this not sound like AI," "humanize this," "remove AI patterns," "write like a real person," or any request to eliminate synthetic writing patterns. Handles blog posts, LinkedIn posts, sales copy, emails, essays, technical docs, social media, landing pages, product descriptions, and thought leadership. Does NOT just swap words — it restructures the cognitive architecture of the text.

02 资源文件 7

03 SKILL.md 正文

Alpha Humanizer

Most "humanization" fails because it treats AI writing like a vocabulary problem. Swap "delve" for "explore," done. That's cosmetic surgery on a corpse.

Real humanization is cognitive reconstruction. AI text fails not because of word choice — it fails because it thinks in probabilities, not intentions. It writes from nowhere, to no one, about nothing specific. This skill attacks the root: the structural, rhythmic, tonal, and epistemic signatures that make text register as synthetic to both human readers and detection algorithms.

The skill operationalizes techniques from Gary Provost (sentence music), Roy Peter Clark (ladder of abstraction), Verlyn Klinkenborg (implication over explanation), William Zinsser (clarity and first-person warmth), George Orwell (anti-jargon), Elmore Leonard (cutting hooptedoodle), Ursula K. Le Guin (crowding and leaping), Constance Hale (kinetic verbs), Francine Prose (word-level scrutiny), and Kurt Vonnegut (killing the preamble).

Mode Selection

Determine mode from user input. If ambiguous, default to Mode A if text is provided, Mode B if not.

Mode A: Full Rewrite — User pastes or uploads text to humanize. Indicators: "rewrite this," "humanize this," "make this sound human," or any text block provided with a humanization request. If user also specifies a target content type different from the input (e.g., "rewrite this as a LinkedIn post"), apply Mode A workflow but use the structural blueprint from Mode B Step 2 for the target content type. → Read references/workflows/full-rewrite.md

Mode B: Write Human from Scratch — No existing text provided. User wants new content created. Indicators: "write me a," "create a," "draft a," or topic-only requests. → Read references/workflows/write-from-scratch.md

Mode C: Quick Strip — User explicitly signals speed. Indicators: "quick edit," "just clean it up," "fast pass," "5 minute fix." → Read references/workflows/quick-strip.md

For ALL modes, before generating ANY text:

  1. Read references/concepts/banned-patterns.md — the forensic kill list
  2. Read references/concepts/master-techniques.md — the craft arsenal

Before delivering ANY output: → Read references/anti-patterns.md and verify output violates NONE

If unsure about output quality for a specific content type: → Read references/examples/before-after.md for calibration

Hard Constraints (Never Violate)

  1. Never use ANY word or phrase from the Banned Lexicon in references/concepts/banned-patterns.md
  2. Never use em dashes (—). Use periods, commas, colons, or semicolons instead
  3. Never start with "In today's...", "In the ever-evolving...", "In the realm of..."
  4. Never use the Rule of Three for examples (use 1, 2, 4, or 5 — never 3)
  5. Never end with "In conclusion," "Ultimately," or "At the end of the day"
  6. Never produce sentences all within the 15-25 word range. Mix 3-word fragments with 35+ word complex sentences
  7. Never use participial clause chains ("Doing X, he did Y")
  8. Never hedge with "It is important to note" or "It is worth mentioning"
  9. Never write from "the view from nowhere." Every piece needs a point of view, an opinion, a stake
  10. Never produce a paragraph that follows Topic Sentence → Support → Summary structure for every paragraph

Soft Constraints (Follow Unless User Overrides)

  1. Default to contractions (can't, won't, doesn't) over formal equivalents
  2. Use first person ("I," "we") when the content type allows it
  3. Prefer Anglo-Saxon root words over Latinate (use > utilize, help > facilitate, start > commence)
  4. Sentence length standard deviation must exceed 10 words
  5. At least one sentence per 200 words should be under 5 words
  6. At least one sentence per 200 words should exceed 30 words
  7. Include at least one concrete sensory detail per 300 words
  8. Break the fourth wall at least once per 500 words in casual content

Reference File Index

  • references/concepts/banned-patterns.md — Complete forensic kill list of AI vocabulary, syntax, and structural patterns. Load before ALL writing.
  • references/concepts/master-techniques.md — The 10-master craft arsenal with specific techniques. Load before ALL writing.
  • references/workflows/full-rewrite.md — Step-by-step protocol for transforming existing AI text. Load when user provides text to humanize.
  • references/workflows/write-from-scratch.md — Protocol for generating new human-sounding content. Load when user wants new content created.
  • references/workflows/quick-strip.md — 5-minute emergency edit protocol. Load when user needs fast results.
  • references/examples/before-after.md — Annotated rewrite demonstrations across content types. Load for output calibration.
  • references/anti-patterns.md — Verification checklist for final output. Load before delivery.

04 同仓库更多 Skills

marcelfahle
marcelfahle/skills

zight

Analyze Zight screenshots and screen recordings from share.zight.com links. Use this skill automatically when a Zight link appears in a ticket, user message, or bug report. Handles both images (view directly) and videos (extract frames, transcribe audio, analyze visually). Produces a structured summary of what the user sees and says in the recording.

图像与多媒体
★ 0
marcelfahle
marcelfahle/skills

x-read

Read tweets, threads, user timelines, mentions, bookmarks, likes, and search results from X (Twitter) via the official X API v2. Use when the user asks "what's in this tweet", "summarize this thread", "show my mentions / bookmarks", "what did <user> tweet about", "what's trending on X about <topic>", or pastes an x.com / twitter.com URL and wants the contents. Owned reads (the user's own posts, mentions, bookmarks, likes, followers, lists) cost $0.001 per resource since 2026-04-20; arbitrary tweet reads are $0.005. The skill never invents tweet text — every output is the literal API response or `null` when the API has no data. Does NOT post, like, follow, or mutate state — read-only.

★ 0
marcelfahle
marcelfahle/skills

video-teardown

暂无简介

图像与多媒体
★ 0
marcelfahle
marcelfahle/skills

linkedin-archive

Archive Marcel's recent LinkedIn posts into the Obsidian vault as weekly markdown files matching the social-corpus format. Use when Marcel says "archive my LinkedIn", "pull my LinkedIn", "/linkedin-archive", "update the social corpus", or "what did I post on LinkedIn this week — and capture it". Runs as a Claude cowork skill on Marcel's own machine (logged-in browser session); does NOT run on the gateway VM. Companion to the Twitter side (`~/.openclaw/bin/social-archive-x.mjs`, auto-cron Sat 21:00 Madrid) which uses the X API and runs unattended. LinkedIn API doesn't allow own-timeline reads at consumer tier, so this is intentionally manual/assisted.

★ 0
marcelfahle
marcelfahle/skills

keyword-research

Run real keyword research with DataForSEO. Volumes, CPC, keyword difficulty, SERP intent, competitor gaps, and clustered topical maps grounded in the user's ICP — not vibes. Use when the user wants to "do keyword research," "find keywords for," "build a keyword list," "what should I rank for," "check search volume," "find keyword gaps vs competitors," "build a topical map," "plan a content cluster," "set up Google Ads keywords," "build an AdWords campaign," or "find long-tail keywords." Pulls the ICP from the user's Obsidian vault when present, talks to DataForSEO directly (REST or MCP), and produces a prioritized, intent-tagged keyword plan ready to feed into Ads, content, or SEO. Does NOT invent volumes or CPCs — every number is cited to a DataForSEO endpoint and timestamp.

市场营销营销推广资料检索与研究
★ 0
marcelfahle
marcelfahle/skills

brand-dossier

Research a prospect/company/creator (from names or URLs) and produce a Bold demo-prep brand dossier — a folder with a paste-ready markdown file (org context, persona, tone, topics, transcription dictionary, conversation starters) plus downloaded logos and extracted brand colors. Use when preparing a Bold demo tenant for a prospect, or when asked to "research X for a demo", "make a dossier", or "brand research".

设计师视觉设计资料检索与研究
★ 0