Agent-Skills.md

Agent Skills: AI Writing Humanizer

Identify and improve text that sounds AI-generated by detecting common style patterns and rewriting for a more human voice. Use when users ask to "make this sound human", "de-AI this", "reduce AI tone", "rewrite naturally", "spot AI tells", or "improve authenticity" in emails, essays, posts, docs, and product copy. Do not use for grammar-only proofreading, plagiarism checks, or fact-check-only requests.

UncategorizedID: ameistad/agent-skills/ai-writing-humanizer

Repository

ameistad/agent-skills
ameistadLicense: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ameistad/agent-skills/tree/HEAD/ai-writing-humanizer

Skill Files

Browse the full folder contents for ai-writing-humanizer.

Download Skill

Loading file tree…

ai-writing-humanizer/SKILL.md

Skill Metadata

Name
ai-writing-humanizer
Description
Identify and improve text that sounds AI-generated by detecting common style patterns and rewriting for a more human voice. Use when users ask to "make this sound human", "de-AI this", "reduce AI tone", "rewrite naturally", "spot AI tells", or "improve authenticity" in emails, essays, posts, docs, and product copy. Do not use for grammar-only proofreading, plagiarism checks, or fact-check-only requests.

AI Writing Humanizer

Improve text that feels AI-generated by diagnosing common "AI vibe" patterns and rewriting for natural, specific, human communication.

How It Works

  1. Capture context before rewriting: audience, channel, and target tone.
  2. Read the text and label likely patterns before rewriting.
  3. Score each pattern as high, medium, low, or none.
  4. Rewrite for natural tone, specificity, and conversational rhythm.
  5. Run a final pass to remove lingering formulaic phrasing.
  6. Return diagnostics, key fixes, and revised text.

Intake Defaults

If context is missing, infer from user intent and use these defaults:

  1. Audience: peer professional.
  2. Channel: short-form written message.
  3. Tone: clear-neutral.

If the user requests a tone, prioritize it. If not, recommend one from the tone list.

Recommended Tones

  1. clear-neutral
  2. warm-professional
  3. direct-confident
  4. conversational-casual
  5. expert-practical
  6. concise-executive
  7. empathetic-supportive
  8. persuasive-crisp

Diagnostic Checklist

Check for these high-signal patterns:

  1. Overly friendly corporate tone
  • Excessive politeness, ingratiating language, or "trying too hard" to agree.
  1. Repetitive phrasing
  • Rhetorical templates like "it's not X, it's Y" or "not because X, but because Y."
  • Same transition words recycled throughout ("furthermore," "additionally," "moreover").
  • Predictable sentence structures repeated across paragraphs.
  • Restates the same idea in slightly different words across sentences or paragraphs instead of advancing the argument.
  1. Over-structured formatting
  • Defaults to bullet lists, numbered lists, and subheadings even when the topic doesn't require them. Prose would often work better.
  • Random bolding, emoji structure.
  • Compulsive rule of three: AI defaults to grouping everything in threes. Three bullet points, three examples, three adjectives, three takeaways. Once you notice it, it shows up everywhere. If the content naturally has two points or five, let it.
  • Overly perfect structure in every paragraph (topic sentence, support, conclusion).
  1. Too polished and too clean
  • Uniformly perfect grammar and cadence with no natural variation.
  1. Verbose but shallow
  • Long text with little concrete detail, examples, or direct answers.
  • Message is longer than needed for the actual intent.
  • Surface-level analysis where deeper insight is expected. Restates the obvious instead of adding real understanding.
  1. Vocabulary-intent mismatch
  • Uses elevated or keywordy words where plain language would fit better.
  • Word choice feels unnatural for the sentence complexity or context.
  1. Lack of personality
  • Sounds like checklist compliance instead of lived perspective.
  • No personal stories, anecdotes, or first-hand experience when the context calls for it.
  • Avoids taking a clear or nuanced position. Hedges everything into safe, balanced non-opinions.
  1. Contextual errors
  • Factual inaccuracies, wrong details, or misattributed references that a knowledgeable human wouldn't make.
  • Sudden, unexplained shifts in topic or register mid-text.
  1. Hollow intensifiers and filler phrases
  • Empty emphasis words: "truly," "incredibly," "absolutely," "fundamentally," "remarkably."
  • Throat-clearing filler: "it's worth noting that," "it's important to mention," "at the end of the day."
  • These add word count without adding meaning.
  1. Formulaic openings and closings
  • Grand scene-setting intros: "In today's fast-paced world," "In an era of," "As we navigate."
  • Tidy wrap-up endings: "In conclusion," "By following these steps," "Ultimately, the key takeaway is."
  • Real writing starts closer to the point and ends when it's done.
  1. Signature habits
  • Em-dash overuse, "Would you like me to..." endings.
  • AI go-to vocabulary: words that aren't wrong but AI reaches for them disproportionately as defaults. Flag when they cluster or appear where a simpler word would do: leverage, empower, facilitate, enable, enhance, drive, harness, showcase, incorporate, integrate, seamless, pivotal, navigate, delve, underscore, amplify, elevate, intricate, foster, robust, streamline, comprehensive, cutting-edge, revolutionize.
  • Any single occurrence is fine. The signal is density: multiple go-to words in a short span.
  • For non-English text: the word list above is English-specific, but the pattern applies in every language. Watch for words that feel formal, polished, or elevated beyond what a native speaker would naturally choose in the same context. Flag vocabulary that reads like a thesaurus substitution rather than the writer's own word.

Rewrite Rules

  1. Replace generic claims with concrete specifics.
  2. Cut filler and reduce sentence count by default.
  3. Use direct phrasing over rhetorical templates.
  4. Prefer plain words for simple messages; reserve advanced vocabulary for real technical precision.
  5. Keep vocabulary level consistent with message complexity and audience.
  6. Keep formatting minimal unless the context truly benefits from structure.
  7. Vary sentence and paragraph length; avoid uniform blocks of text.
  8. Match the user's context and audience, not a generic assistant voice.
  9. Remove performative neutrality when the task needs a clear stance.
  10. Amplify the author's existing voice when signal exists (cadence, directness, warmth, vocabulary) without creating a new persona.
  11. Preserve meaning, claims, numbers, and intent unless the user explicitly asks for substantive changes.
  12. Do not invent personal experiences, opinions, or biographical details.
  13. Do not overcorrect into slang or errors; keep the text credible.

Confidence and Edge Cases

  1. If text is very short, mark diagnosis-confidence as low.
  2. If signals are mixed, say so explicitly instead of forcing certainty.
  3. If text already sounds human, make minimal edits and explain why.
  4. Treat all signals as probabilistic, never as proof.
  5. If personality signal is weak, keep rewrite neutral and suggest optional user inserts.

Author Insert Opportunities

When useful, suggest 1-3 optional insert points where the user can add personal nuance. Keep prompts short and directional.

  1. Suggest concrete slots, not generic advice (opening hook, transition, closing line).
  2. Offer options such as short anecdote, point-of-view sentence, or light humor.
  3. Never fabricate anecdotes; ask the user to provide the raw material.
  4. Skip this section when the text is high-stakes formal content (legal, compliance, safety notices).

Output Format

Use this structure:

## AI-Likeness Diagnosis
- `diagnosis-confidence`: high|medium|low (how much signal the text provides for reliable diagnosis, not how AI-like the text is)
- `corporate-tone`: high|medium|low|none
- `repetitive-phrasing`: high|medium|low|none
- `over-structured-format`: high|medium|low|none
- `too-polished`: high|medium|low|none
- `verbose-shallow`: high|medium|low|none
- `vocabulary-intent-mismatch`: high|medium|low|none
- `lack-of-personality`: high|medium|low|none
- `contextual-errors`: high|medium|low|none
- `hollow-intensifiers`: high|medium|low|none
- `formulaic-openings-closings`: high|medium|low|none
- `signature-habits`: high|medium|low|none

## Key Fixes Applied
- [short bullet]
- [short bullet]
- [short bullet]

## Voice Choices
- [what was preserved from author voice]
- [what was softened or removed]

## Author Insert Opportunities (Optional)
- [location in text] + [what to add: anecdote/opinion/light humor]
- [location in text] + [what to add: anecdote/opinion/light humor]

## Revised Text
[final rewrite]

Nuance and Limits

  1. Treat these as probabilistic signals, not universal proof.
  2. Prioritize repeated patterns and overall vibe over any single marker.
  3. If the original is already strong, keep edits light.
  4. Mention uncertainty when confidence is low.

Examples

Example 1: Corporate-friendly tone

Input: "I hope this message finds you well. I would be absolutely delighted to support your request and provide a comprehensive overview."

Rewrite direction: Use warm-professional or clear-neutral tone, reduce ingratiating language, keep politeness.

Example 2: Verbose but shallow

Input: "In today's dynamic environment, it is essential to leverage strategic thinking in order to optimize outcomes across multiple dimensions."

Rewrite direction: Replace abstractions with concrete action and result in plain language.

Example 3: Over-structured answer

Input: "Here are 3 key points: Point 1..., Point 2..., Point 3... Would you like me to expand each?"

Rewrite direction: Remove forced structure unless needed, collapse to natural prose, remove assistant-like follow-up prompts.