AgentSkillsCN

humanizer

对帖子草稿进行三重人性化处理:先检测 AI 模式并进行改写,随后进行最终校验(包括 Coffee 测试、模板测试、滚动测试以及 AI 检测)。 触发条件:“humanize”、“make it sound human”、“ralph wiggum”、“AI check”。

SKILL.md
--- frontmatter
name: humanizer
description: |
  3-pass humanization for post drafts. Detects AI patterns, rewrites,
  then runs final checks (Coffee, Template, Scroll, AI Detector tests).
  Triggers: "humanize", "make it sound human", "ralph wiggum", "AI check".
user-invocable: true

Humanizer

Every post draft MUST pass through this skill before delivery. This is a self-correction loop that rewrites AI-sounding copy into natural human writing.

When to Use

  • ALWAYS before delivering a final draft (Phase 6 of content-pipeline)
  • When a draft "sounds AI" or "too polished"
  • When you suspect a draft is formulaic

Pass 1: AI Pattern Detection

Read the draft and flag every instance of these patterns:

Kill on sight (ANY of these = draft FAILS this pass):

  • Em dashes (any occurrence)
  • Emoji (any occurrence)
  • "Here's the thing" / "Here's why" / "Let that sink in" / "Full stop."
  • "Let me break this down" / "Let's dive in" / "Here's what I mean"
  • Starting 3+ sentences with the same word
  • "That is" instead of "That's" / "do not" instead of "don't" / missing contractions
  • "It's not X. It's Y." / "Not only...but also..." (negative parallelisms)
  • Imperative commands: "Stop doing X. Start doing Y."
  • AI vocabulary: Additionally, Furthermore, landscape, tapestry, testament, pivotal, showcase, delve, foster, robust, comprehensive, seamless, cutting-edge, harness, realm, paradigm, holistic, synergy, leverage, utilize
  • "Serves as", "stands as", "functions as" when "is" works
  • "On another level", "marks a shift", "changes everything" (inflated significance)
  • "Innovation, inspiration, and insights" style forced triplets
  • Synonym cycling (same concept, 3+ different words)
  • "Could potentially", "it's worth noting that" (over-hedging)
  • "From startups to enterprises" (false ranges)
  • Any organization name

Structure fail (ANY = draft is too AI-structured):

  • All numbered points the same length
  • Every paragraph 2-3 sentences (no variation)
  • Blog outline format
  • Thesis, explanation, numbered list, question pattern
  • Every numbered point same grammatical start
  • Opening line longer than 10 words
  • Post ends with a question

Missing elements (ANY missing = add before proceeding):

  • At least one sentence under 5 words
  • At least one mid-post question (NOT at the end)
  • At least one specific number from source code
  • At least 5 contractions
  • At least one vulnerable/honest moment
  • At least one parenthetical aside (using parentheses)
  • "I" or "we" appearing naturally

Pass 2: Rewrite

Fix every flagged issue:

  1. Keep the same core message and structure.
  2. Make it sound like someone typing a post, not drafting an essay.
  3. Word swaps: leverage->use, architect->build, utilize->use, facilitate->help, implement->build/add/set up, comprehensive->cut, robust->cut, seamless->cut, harness->use, delve->cut, foster->build or cut, pivotal->cut, showcase->show
  4. Fix negative parallelisms: "It's not X. It's Y." becomes a single merged sentence.
  5. Fix em dashes: replace ALL with commas, periods, or parentheses. Zero tolerance.
  6. Fix vague claims: "on another level" -> specific metric. If you can't find a number, cut the claim.
  7. Fix synonym cycling: if you used "retrieval" once, use "retrieval" again. Don't switch words.
  8. Add pattern breaks: a parenthetical aside, a one-word sentence, a sentence that interrupts the flow.
  9. If numbered points exist, vary their lengths and opening words.
  10. Opening line: would you actually stop scrolling? If not, make it shorter and more blunt. Under 10 words.
  11. Closing line: must be a statement, not a question. Must land with weight.
  12. Remove all organization names.

Pass 3: Final Check

Read the rewritten draft and apply four tests:

Coffee Test: Does this sound like someone telling you about it over coffee, or like a LinkedIn thought leadership post? If the latter, rewrite.

Template Test: Could you swap the topic (trading bots -> databases -> deployment) and the post still works with minimal changes? If yes, it's too generic. Add specific details that ONLY apply to this topic.

Scroll Test: Read ONLY the first line. Would you stop scrolling? If not, make it shorter, more blunt, more specific. Must be under 10 words.

AI Detector Test: Read as a skeptic looking for AI tells. Check:

  • Any em dashes remaining?
  • Any "It's not X. It's Y." constructions?
  • Any AI vocabulary from the kill list?
  • Any vague claims without numbers?
  • Uniform sentence/paragraph lengths?
  • Synonym cycling?
  • If you find even ONE AI tell, fix it.

Could you tell this was written by AI? If yes, do another pass.

Self-Scoring (Internal Only)

After the final draft, score internally:

  • Human-sounding: 1-10 (target: 8+)
  • Specificity: 1-10 (target: 7+)
  • Hook strength: 1-10 (target: 8+)

If any score below target, do one more rewrite pass focused on that area. Do not share scores.

Output

Return only the final rewritten draft. No commentary, no "here's the improved version", no before/after. Just the clean post text.