Writing Humanizer

Writing Humanizer

Concept of the skill

Writing humanization is the discipline of turning robotic or AI-patterned prose into clear, direct, reader-trustworthy text while preserving truth and intent — prose-quality repair, not authorship laundering. It works on a fixed set of primitives: AI tells (predictable filler, hedging, passive voice, hollow intensifiers), readability signals (sentence length, paragraph density, jargon load), rhythm signals (opening-word variety, sentence-length variance, paragraph cadence), audience register, and evidence specificity. The skill replaces "make it sound human" as a vague style request with a repeatable five-step edit pass — Tell Scan, Readability Check, Structural Rewrite, Rhythm Pass, Voice Calibration — that moves through those primitives in order: remove the tells, clarify who the actor is, improve readability, vary the rhythm, then calibrate tone to the surface (an API doc, an error message, a release note, a commit). Detector-style signals such as perplexity and burstiness are treated as rough diagnostic clues, never as proof of authorship and never as the target; the reader, not the detector, is the audience. The non-negotiable constraint is honesty: every edit must keep the text at least as clear and as accurate as it was, and the skill refuses to misrepresent authorship or promise detector-proof output.

Coverage

The full pipeline for transforming AI-generated or robotic text into clear, human-sounding prose:

• AI-tell detection and removal — Tier 1 zero-tolerance word list (delve, testament, crucial, vital, paramount, furthermore, seamless, robust, comprehensive, cutting-edge, foster, empower, leverage, harness, etc.); Tier 2 conditional list (utilize, facilitate, streamline, implement, optimize)
• Active-voice conversion — passive-to-active decision tree based on actor knownness; hedging-pattern removal table
• Readability scoring as diagnosis — Flesch-Kincaid Grade (target 8–10 for general docs), Gunning Fog Index (10–12 for technical docs), Flesch Reading Ease (60–70 for UI copy); the readability diagnostic tree for sentence length, word complexity, paragraph density, and nested clauses
• Sentence variety and rhythm — the 3-beat short-long-medium pattern; sentence-structure variety checklist (declarative, compound, conditional, question+answer, fragment); opening-word rotation rule
• Vocabulary diversity — repeat technical terms exactly, rotate generic verbs, avoid elegant variation; the abstract-vs-concrete table; the jargon decision tree
• Tone mapping — the formal-to-casual spectrum (1–5) and a context-tone table covering API documentation, error messages, UI tooltips, commit messages, PR descriptions, release notes, issue bodies, onboarding copy, empty states, and marketing copy
• AI-detector limits and prose fingerprints — how some detectors use probability signals such as perplexity and burstiness, why detector scores are not proof of authorship, and which clarity-preserving edits reduce robotic prose without gaming a review process
• Paragraph rhythm and structure — paragraph-length rules per context, the hook-body-landing pattern, the bullets-vs-prose decision tree
• Anti-patterns — over-qualification, repetitive transitions, the enumeration trap, hollow intensifiers
• The 5-step humanization workflow — Tell Scan, Readability Check, Structural Rewrite, Rhythm Pass, Voice Calibration

Philosophy of the skill

AI-generated text has consistent, detectable failure modes: excessive hedging, passive voice, hollow superlatives, monotonous sentence structure, and vocabulary that signals machine authorship. Left unchecked, these patterns erode trust with human readers, invite detector flags or editorial pushback, and produce text that is longer and less clear than necessary. Every agent in a system produces text that humans read, and every piece of that text reflects on the product.

The mental model: many detector-style systems look for probability patterns such as perplexity (word predictability) and burstiness (sentence-length variance), but those signals are review clues, not authorship proof. The humanizer's job is to shift the prose toward clarity, specificity, and natural rhythm without sacrificing honesty. Plain-language editing addressed the same underlying failure long before LLMs existed; this skill applies that discipline to AI-era prose fingerprints.

This skill is not about making text casual. A technical API doc can be fully humanized while staying formal. It is also not about adding contractions alone — structural uniformity is the stronger fingerprint. And it is not a substitute for copywriting; persuasion and conversion are separate concerns owned by a copywriting skill.

Clear, direct, and human. AI tells are bugs in the prose.

When to Use

Use this skill whenever you are tasked with:

1. Writing new documentation, reports, or communicative user text.
2. Editing existing copy or documentation.
3. Reviewing PR descriptions, issue bodies, or release notes.
4. "Humanizing" or adjusting the tone of any drafted text.
5. Writing error messages, UI copy, tooltips, or onboarding instructions.
6. Rewriting text flagged as AI-generated or robotic.
7. Writing commit messages, changelogs, or release notes.

1. Remove AI Tells and Cliches

Eradicate filler words and overused phrases that immediately signal AI authorship.

Tier 1 — Always remove (zero tolerance):

• delve, dive into, navigate (unless literally navigating a UI)
• testament, crucial, vital, paramount
• furthermore, moreover, consequently, in conclusion
• seamless, robust, comprehensive, cutting-edge
• foster, empower, elevate, leverage, harness
• it's important to note that, it's worth mentioning
• in today's world, in the ever-evolving landscape

Tier 2 — Remove unless technically precise:

• utilize → use; facilitate → help / enable; streamline → say what improves
• implement is OK in code context; optimize is OK when measurable

Before: Let's delve into the comprehensive architecture, which serves as a testament to our robust backend. After: Here is the backend architecture.

2. Enforce Active Voice and Direct Phrasing

Convert passive constructions into active ones. Remove hedging and unnecessary introductory clauses.

Passive-to-Active Decision Tree

Is the actor known?
├── Yes → Name them: "The server rejects invalid tokens"
│         (not "Invalid tokens are rejected")
└── No → Is the reader the actor?
    ├── Yes → Use imperative: "Click the button to submit"
    │         (not "The button should be clicked")
    └── No → Passive is acceptable: "The migration was applied at 03:00 UTC"

Hedging Language to Remove

Before: In order to ensure that the synchronization process completes successfully, it is important to check your internet connection. After: Check your internet connection before syncing.

3. Readability Scoring

Use readability formulas to diagnose problems, not as optimization targets. Over-optimizing for a score produces choppy, unnatural prose.

When to Apply Each Formula

Readability Diagnostic Tree

Text feels hard to read?
├── Long sentences (avg > 20 words)?
│   └── Break sentences. Target 12–18 word average.
├── Multi-syllable words (> 3 syllables frequently)?
│   └── Replace with plain equivalents where meaning is preserved.
├── Dense paragraphs (> 4 sentences)?
│   └── Split. Use bullets for lists of 3+ items.
└── Nested clauses (2+ commas per sentence)?
    └── Unpack into separate sentences or use a colon + list.

Word-Level Simplification

4. Sentence Variety and Rhythm

Monotonous sentence structure is the #1 tell of AI-generated text. Humans naturally vary sentence length, structure, and opening words.

The 3-Beat Rhythm Pattern

Alternate sentence lengths in a roughly short–long–medium pattern. This creates a natural cadence that mirrors how people actually write.

Monotonous (AI pattern):

The dashboard shows your profit. The dashboard shows your costs. The dashboard shows your revenue. The dashboard shows your orders.

Natural rhythm:

The dashboard shows your profit at a glance. Below the headline number, you will find a full cost breakdown — production cost, ad spend, transaction fees, and refunds, each with its own trend line. Revenue and order counts sit in the sidebar.

Sentence Structure Variety Checklist

Opening Word Rotation

Never start 3+ consecutive sentences with the same word. Especially avoid The (the most common AI opener), This, It, There. Rotate We / You too.

5. Vocabulary Diversity

Synonym Rotation Rules

1. Repeat technical terms exactly — your domain term (an acronym like MRR or CAC, or any other established industry abbreviation) is always written the same way; never call the same concept production costs in one place and fulfillment expenses in another.
2. Rotate generic verbs — shows, displays, lists, includes can rotate for non-technical usage.
3. Avoid elegant variation — do not call the same thing three different names. Clarity beats variety.

Concrete vs. Abstract

Jargon Decision Tree

Is the term domain-specific (industry vocabulary your audience knows)?
├── Yes → Keep it. Add a tooltip or parenthetical on first use.
│         Example: "production cost (what your fulfillment partner charges per order)"
└── No → Is there a simpler word with identical meaning?
    ├── Yes → Use the simpler word.
    └── No → Keep the term but define it inline.

6. Tone Mapping Framework

Tone is not one-size-fits-all. Match the tone to the context and audience.

The Formal-to-Casual Spectrum

Legal/Compliance ←──── Technical Docs ←──── Product Copy ←──── Internal Comms ←──── Chat
      1                     2                    3                   4                5

Context-Specific Tone Rules

7. AI-Detector Limits and Prose Fingerprints

Treat AI-writing detector scores as probabilistic review signals, not proof. OpenAI retired its AI Text Classifier because of low accuracy and warned that classifiers should not be primary decision tools. Turnitin's own guidance suppresses low positive scores below its 20% threshold to reduce false positives. Stanford HAI reported that widely used detectors misclassified more than half of non-native-English TOEFL essays as AI-generated in one study.

That does not mean detector-like signals are useless. Low-specificity prose, uniform sentence length, hollow intensifiers, and predictable transitions are still real writing problems. Fix them because they hurt readers, not because a score changed.

Clarity-Preserving Fingerprint Fixes

What NOT to Do

• Do not insert typos or grammatical errors.
• Do not use obscure vocabulary just to raise perplexity.
• Do not sacrifice clarity for detection avoidance — clarity always wins.
• Do not apply these techniques to code comments or strict API docs (precision over personality there).
• Do not promise that text will pass a detector, and do not claim human authorship when AI was materially involved.

Source Notes

• OpenAI's AI Text Classifier notice says the tool is unavailable because of low accuracy and should not be used as a primary decision-making tool.
• Turnitin's AI Writing Report guidance suppresses scores above 0% and below 20% to reduce false-positive risk.
• Stanford HAI's detector-bias report documents elevated false positives for non-native-English writing.
• Digital.gov plain-language guidance supports active voice, audience-first wording, and clear sentence structure as readability practices independent of AI detection.

8. Paragraph Rhythm and Structure

Paragraph Length Rules

The Hook-Body-Landing Pattern

Strong paragraphs follow this structure:

1. Hook sentence — short, direct, states the point (5–10 words)
2. Body sentences — develop the point with evidence or detail (1–3 sentences)
3. Landing sentence — wraps up or transitions (optional, keep short)

Weak (no hook): When you consider the various factors that contribute to the overall profitability of an order, including the cost of goods sold... (reader lost by word 10)

Strong (hook-body-landing): Order profit depends on three cost layers. Production cost covers what your fulfillment partner charges. Transaction fees come from your payment processor. Add refunds, and you get real margin.

When to Use Bullets vs. Prose

Is it a list of 3+ items?
├── Yes → Are items parallel in structure?
│   ├── Yes → Bulleted list
│   └── No → Fix parallelism first, then bullet
└── No → Prose (2 items can stay inline with "and")

9. Anti-Patterns Reference

Over-Qualification

Adding unnecessary caveats that weaken every statement.

Repetitive Transitions

AI text chains paragraphs with the same connectors. Vary or delete.

The Enumeration Trap

AI loves "There are three key aspects:" followed by "First, ... Second, ... Third, ..." Humans just list them: Order profitability depends on production cost, transaction fees, and refund rates.

Hollow Intensifiers

Remove words that add emphasis but no meaning: very, really, extremely, highly, absolutely. If important needs strengthening, pick a stronger word (critical, required). If fast needs a number, give the number.

10. The Humanization Workflow

When tasked with humanizing text, follow this 5-step process:

1. Tell Scan. Read the text once. Highlight every word from the Tier 1 removal list (section 1). Mark passive constructions and hedging phrases.
2. Readability Check. Estimate average sentence length. If over 20 words, flag for splitting. Count paragraphs over 4 sentences — flag for breaking.
3. Structural Rewrite. Break long sentences. Convert inline semicolon lists to bullets. Apply the hook-body-landing pattern to weak paragraphs. Fix passive voice using the decision tree (section 2).
4. Rhythm Pass. Read the rewrite aloud (mentally). Check for:
   • 3+ consecutive sentences starting with the same word → rotate openers
   • All sentences the same length → vary (short-long-medium)
   • All sentences the same structure → mix declarative, conditional, fragment
5. Voice Calibration. Match tone to context using the tone-mapping table (section 6). For technical docs, keep precision over personality.

Output Expectations

When asked to revise text or provide new text following these guidelines, only output the revised text. Do not provide metacommentary like "Here is the humanized version:" unless explicitly asked to explain your changes.

Verification

Before finalizing any humanized text, confirm:

• Zero Tier 1 AI tells remaining (section 1 word list)
• No passive voice where active is possible (section 2 decision tree)
• Average sentence length between 12–18 words
• No paragraph exceeds 4 sentences (docs) or 2 sentences (UI copy)
• No 3+ consecutive sentences start with the same word
• Sentence lengths vary (not all within 2–3 words of each other)
• Technical jargon is explained on first use or tooltipped
• Tone matches the context (section 6 tone table)
• Concrete details used instead of abstract claims (section 5)
• No metacommentary in output ("Here is the humanized version:")
• No claim that the output is detector-proof or human-authored when AI was materially involved

Do NOT Use When

Skill Graph context

Classification

• Subject: design
• Public: true
• Domain: design/content
• Scope: Writing and editing human-readable prose — docs, PRs, issues, release notes, errors, UI copy, commits, tooltips, support replies — especially when text sounds robotic, padded, monotonous, or overly formal: AI-tell removal, active voice, hedging reduction, readability diagnosis, sentence rhythm, vocabulary variety, tone mapping, paragraph rhythm, the bullets-vs-prose choice, and the 5-step humanization workflow. Portable across any written communication; principle-grounded, not repo-bound. Excludes documentation routing/type selection, code-identifier naming, and in-product UI-text pattern catalogs (microcopy).

When to use

• this PR description sounds AI-generated — strip the tells and rewrite it concisely
• rewrite this onboarding paragraph in the active voice with shorter average sentence length
• audit this release-notes draft for Tier 1 AI tells (delve, leverage, comprehensive, testament)
• this paragraph starts every sentence with The dashboard — rotate the openers and vary the rhythm
• humanize this error-message body so it stops sounding like a corporate FAQ
• drop the hollow intensifiers and over-qualification from this tooltip
• the docs team flagged this prose as robotic — apply the 5-step humanization workflow
• Triggers: humanize this text, sounds AI-generated, strip the AI tells, make this read like a human wrote it, this prose is robotic

Not for

• decide kebab-case vs camelCase for this new database column
• draft the marketing headline for the pricing page with strong persuasion
• restructure this doc into a tutorial format with progressive disclosure
• rewrite this UI button label so it names the actual action instead of saying Submit
• rename this React component across all call-sites in the repo
• audit this WCAG 2.2 contrast violation on the dashboard
• Owned by microcopy

Related skills

• Verify with: linguistics
• Related: linguistics, microcopy, semantics

Concept

• Mental model: |
• Purpose: |
• Boundary: |
• Analogy: Writing humanization is like audio mastering: the recording may already contain the right notes, but mastering removes hiss, balances loudness, and restores dynamics so a listener can trust what they hear.
• Common misconception: |

Keywords

• AI-tell detection, AI-tell removal, prose humanization, passive-to-active voice, hedging-pattern removal, readability scoring diagnosis, sentence-rhythm pattern, 3-beat sentence variety, hook-body-landing paragraph, tone mapping framework