humanize

name	humanize
description	Convert AI-written text to more human-like writing through subtle edits. Use when text reads "too AI", when the user mentions "humanize", "sounds robotic", "AI-written", "make it natural", or when editing for a more conversational voice.

Humanize Text

Make AI-generated content read like it was written by a human through targeted, subtle edits.

Why? AI-written text has telltale patterns—formulaic transitions, passive voice, overly balanced sentences—that make it feel mechanical. This skill fixes those patterns without rewriting the whole document.

Model Coverage: Tested on Opus 4.5, Sonnet 4, Haiku. See TESTING.md for details.

Quick Start

1. wc -w FILE (baseline) → 2. Scan for patterns → 3. Make 10-20 targeted edits → 4. Verify ±20 words & 0 em-dashes → 5. Check burstiness

Supporting Files: REFERENCE.md (pattern tables) | TESTING.md (evaluation scenarios)

Content Integrity Rules (CRITICAL)

[!CRITICAL] The Cardinal Rule: Transform style, never fabricate content.

Humanization edits HOW something is expressed, not WHAT is expressed. Every technique must work only with material already present in the source text.

Allowed (Style Transformation)

• Restructure sentences (change order, split, combine)
• Replace words with synonyms that preserve meaning
• Change punctuation and sentence boundaries
• Add/remove contractions (domain-appropriate)
• Convert passive to active voice
• Vary sentence lengths by restructuring existing content
• Add hedging to soften existing claims ("X is true" → "X appears to be true")
• Convert existing lists to prose or vice versa

Forbidden (Content Fabrication)

• Invent personal anecdotes, opinions, or experiences not in source
• Add fake citations, names, dates, or statistics
• Create metaphors that introduce claims not in the original
• Insert "I've seen...", "In my experience..." unless source has them
• Make up specific details to replace vague ones
• Add editorial commentary ("surprisingly", "disappointingly") unless source expresses that sentiment

The Source Material Test

Before any edit, ask: "Is this information already in the source text?"

• If YES → transform freely
• If NO → do not add it

How It Works

Step 1: Read and Scan for AI Patterns

Read the target file and identify common AI-writing tells. Priority patterns to scan first:

1. "By [gerund]" — "By implementing...", "By training..."
2. "That [noun]" connectors — "That shift...", "That vulnerability..." (linking sentences)
3. Indirect speech — "The field is shifting...", "Research suggests...", "A study identifies..."
4. Em-dashes (—) — Humans rarely use them; AI overuses them
5. "This [verb] that" — "This suggests that...", "This demonstrates that..."
6. Subordinate smoothness — "while maintaining X", "thereby reducing Y" (too-smooth connectors)
7. Framework intro pattern — "The [X] framework mitigates/addresses..." (robotic combo)
8. High-risk phrases — "framework provides" (7x), "maintaining high" (6.4x), "eliminating the need" (5.4x)
9. "For X, Y does Z" — "For real-time applications, X optimizes..." (formal opener)
10. Colon definition splits — "X decouples A from B: it maintains..." (explanatory colons)
11. Too-simple declaratives — Short, direct sentences can also trigger "robotic formality"

See REFERENCE.md for the complete pattern detection table.

Step 2: Apply Targeted Edits

Make 10-20 edits across the document. Do NOT rewrite entire sections.

Step 2b: Word Count Verification (MANDATORY)

[!CRITICAL] THE ±20 WORD RULE: Final word count must be within ±20 words of the original. Always measure against the ORIGINAL document, not previous iterations.

AI models inherently summarize. You must fight this bias by restructuring, not condensing.

Workflow:

1. Measure: wc -w PATH before editing
2. Edit: Apply targeted changes
3. Verify: wc -w PATH after editing
4. If > 20 word change: STOP. Revert and restructure instead of cutting/padding

How to preserve word count:

• Restructure sentences: "By analyzing X" → "When analyzing X" (same length)
• Expand expressions: "X happens" → "X happens because Y, which means Z"
• Never pad with filler ("meaningfully", "smartly", "actively")

Step 2c: Em-Dash Count Verification (MANDATORY)

[!CRITICAL] NO NEW EM-DASHES: Em-dash count must NOT increase. Target: 0.

Verify: grep -o '—' PATH | wc -l (before and after)

If em-dashes increased, replace with: periods, commas, colons, or parentheses.

Step 2d: Anti-Detection (Burstiness & Perplexity)

AI detectors measure statistical uniformity. Disrupt rhythm and predictability using only existing content.

1. Burstiness (Sentence Length Variation):

[!CRITICAL] AI maintains uniform 12-18 word sentences. Human writing has HIGH variance. This is the #1 detection signal.

Target per 10 sentences: 2-3 very short (2-6 words), 2-3 very long (25-40 words), 4-5 medium. See REFERENCE.md.

Techniques:

• Split: "The model processes data and outputs results" → "The model processes data. Then it outputs results."
• Combine: "X works. Y helps." → "X works, and when combined with Y, it improves significantly."

2. Vocabulary Entropy:

Replace 3-5 "AI-typical" words per paragraph with rarer synonyms that preserve meaning exactly. See REFERENCE.md.

[!CAUTION] Synonym must have EXACT same meaning. If unsure, keep the original.

3. Visual Structure: Vary paragraph shapes (dense → bullets, short paragraphs → merged).

Step 2e: Lexical Diversity

AI text has measurably lower vocabulary diversity. Fix by varying word choice using only synonyms that preserve meaning.

1. Connector Audit: Each connector should appear max 2 times per 1000 words. If more, replace 50% or restructure. See REFERENCE.md.

2. Verb Repetition: If any verb appears 3+ times in 500 words, vary it. See REFERENCE.md.

3. Noun Phrase Variation: After first reference, vary: "The transformer" → "this approach" → "it"

[!CAUTION] Never change meaning. Only vary when semantically equivalent.

Step 2f: Punctuation Diversity

Humans use more varied punctuation than AI. Increase variety by restructuring. See REFERENCE.md for targets.

Techniques:

• Questions: "The implications are significant" → "What are the implications? They're significant."
• Semicolons: "X is fast. Y is slow." → "X is fast; Y is slow."
• Parentheses: "The approach, which is unconventional, works." → "The approach (unconventional as it is) works."

[!CAUTION] Questions must not imply answers not in the source.

Condensing vs. Restructuring

[!CRITICAL] Most common failure mode. Condensing removes words; restructuring rearranges them.

Condensing (❌)	Restructuring (✅)
"Long sentence" → "Short sentence"	"Long sentence" → "Reworded long sentence"
Removes words	Changes arrangement
Net content loss	Same content, different pattern

When tempted to condense: Expand expressions, add supporting detail, or break into multiple sentences.

Multi-Pass for Long Documents (2000+ words):

1. Scan high-frequency patterns
2. Fix sentence rhythm
3. Verify no new patterns created
4. Word count check (MANDATORY)

Transition Replacements: See REFERENCE.md.

Key Rules:

• Never use em-dashes. Replace with periods, commas, colons, or parentheses
• Questions only for topic transitions, not rhetorical pauses
• Keep formal register in academic writing (no contractions)
• Remove filler: "It is worth noting that" → just state the thing

Step 3: Add Human Personality

[!IMPORTANT] Removing AI patterns is not enough. Detectors also flag text that lacks "rhetorical flourishes" and feels "impersonal." You must ADD human touches using only existing content.

Inject Personality (without fabricating):

• Mild surprise: "Interestingly," or "Curiously," before a finding (if the finding IS interesting)
• Conversational asides: "—and this matters—" or "(worth noting)"
• Direct address: "Here's the thing:" or "Notice that..."
• Occasional informality: "pretty effective" instead of "effective", "a lot" instead of "significantly"
• Opinion hedging: "seems to", "appears to" (humans hedge more than AI)

Disrupt S-V-O Order:

• Invert occasionally: "Effective, this approach was not." → only when natural
• Lead with result: "A 10% gain—that's what the model achieved."
• Fronted adverbials: "In practice, the system fails." instead of "The system fails in practice."

Break Impersonal Tone:

• Replace "The field is shifting" → "Researchers are shifting the field" (add human actors)
• Replace "Research suggests" → "Three recent papers suggest" (specificity)
• Replace "The implication is clear" → "What does this mean? It means..." (question form)

Vary Grammar (Break "Correct but Unvaried"):

• Avoid repeating sentence structures: if three sentences use "X [verb]s Y", restructure one
• Break parallel semicolon lists: "A does X; B does Y; C does Z" → "A does X. Meanwhile, B does Y. And C? It does Z."
• Use sentence fragments occasionally: "The result? Better accuracy."
• Try rhetorical inversion: "Effective, this was not." (sparingly)
• Interrupt with asides: "The model—surprisingly—failed at basic counting."
• Break colon splits: "X decouples A from B: it maintains..." → "X separates A and B. This lets it..."

Fix "Too Simple = Robotic":

• Very short declaratives trigger detection too: "The focus is on X." feels robotic
• Add texture: "The focus? X." or "What's the focus here? X."
• Combine with adjacent sentence to add flow
• Or add mild opinion: "The focus, rightly, is on X."

Fix Indirect Speech (Still Heavily Flagged):

• "A study identifies..." → "Smith et al. found...", "Recent work shows...", or just state the finding
• "A protocol called X becomes necessary" → "You need X" or "X becomes essential"
• "Research suggests..." → Name the researchers or say "Three papers this week show..."
• Add human actors: "The field is shifting" → "Researchers are rethinking..."

Humanize Headings (if editing full documents):

• Overly clean headings trigger detection ("Multimodal Grounding and Internal Mechanics")
• Add slight informality: "How Models Actually See" instead of "Visual Processing Mechanisms"
• Use questions: "Why Do Models Fail at Counting?" instead of "Enumeration Failures"
• Keep some formal, vary others—consistency in heading style is itself a tell

Step 4: Vary Your Edits

[!CAUTION] Don't create new patterns. If you replace every "However" with "But", that's just a different pattern. Mix it up:

• Some "However" → "But"
• Some "However" → start sentence differently
• Some "However" → merge with previous sentence using ", but"
• Some "However" → leave as-is

Step 5: Final Verification Checklist

[!IMPORTANT] Complete ALL checks before submitting. For detailed validation scenarios, see TESTING.md.

Content Integrity (DO FIRST):

• No anecdotes/experiences fabricated
• No citations/statistics invented
• All synonyms preserve exact meaning

Quantitative (MANDATORY):

• Word count within ±20 words of original
• Em-dash count ≤ original (target: 0)
• Connectors ≤ 2 per 1000 words each
• Sentence length varies (<6 and >25 word sentences present)

Style:

• No 2+ consecutive paragraphs start same way
• Technical terms and citations preserved
• Contractions match domain register

Quick Validation:

wc -w FILE                    # Word count
grep -o '—' FILE | wc -l      # Em-dashes (target: 0)

Final Test: Does the edited version claim anything the original didn't? If yes, revert.

Examples

Example 1: Formulaic Opening

• Before: "A systematic evaluation of 53 large language models has revealed that longer reasoning chains do not reliably produce better answers."
• After: "A systematic evaluation of 53 large language models revealed something counterintuitive: longer reasoning chains don't reliably produce better answers."

Example 2: "This suggests" Pattern

• Before: "This method proves particularly effective in mathematical reasoning, suggesting that the dichotomy between imitation and exploration is artificial."
• After: "Works especially well for mathematical reasoning, which suggests the imitation vs. exploration dichotomy might be artificial."

Example 3: Passive + Formal

• Before: "The deployment of Large Reasoning Models has been hampered by their tendency to apply uniform computational resources."
• After: "Large Reasoning Models have a problem: they apply the same computational effort whether you ask them to add two numbers or prove a theorem."

Example 4: Conclusion Softening

• Before: "This week's research reflects a shift from unbounded reasoning capability toward calibrated cognitive efficiency."
• After: "The week's theme: unbounded reasoning isn't always better."

Example 5: "By [gerund]" Pattern

• Before: "By employing a margin policy gradient loss and rejection sampling, CompassJudger-2 attempts to create a generalist judge that rivals larger models."
• After: "CompassJudger-2 uses margin policy gradient loss and rejection sampling to create a generalist judge rivaling larger models."

Example 6: Framework Redundancy

• Before: "The RefCritic framework employs a long-chain-of-thought critic module trained via reinforcement learning."
• After: "RefCritic employs a long-chain-of-thought critic module trained via RL."

Example 7: Result Phrasing

• Before: "This approach achieves a 23.2% improvement in success rates on novel software environments compared to static baselines."
• After: "The result: 23.2% better success rates on novel software environments."

Example 8: Adding Questions

• Before: "However, applying these techniques to open-ended domains has remained elusive due to the lack of verifiable signals."
• After: "However, applying these to open-ended domains has remained elusive. Why? No verifiable signals to anchor the training."

Example 9: Preserving Word Count While Removing "By [gerund]"

• Before (32 words): "By analyzing synchronous discourse in human-AI triads, researchers found that the educational value of these agents lies not in their ability to generate content, but in their capacity to alter the structure of reasoning."
• After (32 words): "When analyzing synchronous discourse in human-AI triads, researchers found that the educational value of these agents lies not in their ability to generate content, but in their capacity to alter the structure of reasoning."
• Note: Pattern change achieved by substituting "By" → "When" without restructuring, padding, or cutting. Same word count, improved tone.

Quality Guidelines

• Preserve meaning: Edits change tone, not content
• Stay subtle: 10-20 targeted edits, not a full rewrite
• Maintain expertise: Knowledgeable but not robotic
• Don't over-correct: The problem is overuse and uniformity, not formality itself
• First-reference rule: Keep context on first mention; only shorten after established

Domain-Specific Calibration: See REFERENCE.md.

Academic/Research Warnings:

• Never pad with hollow adverbs ("meaningfully", "smartly")
• Keep technical terminology, section structure, and citations intact
• Vary your pattern replacements (don't swap all "By [gerund]" with "When [verb]")

Troubleshooting

Problem	Solution
Word count changed >20 words	STOP. Revert. Restructure instead of cutting/padding.
Em-dash count increased	STOP. Replace new em-dashes with periods, commas, colons, or parentheses.
Still detected as AI (98%+)	Increase burstiness aggressively; add more punctuation variety; vary vocabulary more.
Fabricated content	Revert. Review Content Integrity Rules. Only transform what exists.
Text too casual	Scale back conversational asides; keep original phrasing.
New repetitive pattern	Vary replacements; use different fixes for same issue.
Compounding issues	Always measure against ORIGINAL document, not previous iteration.

Edge Cases

Edge Case 1: Short Document (<200 words)

• Apply only 3-5 edits maximum
• Focus on the most egregious patterns first
• May not hit all burstiness targets; that's okay for short content

Edge Case 2: Technical Jargon-Heavy Text

• Do NOT replace domain-specific terms with synonyms
• Focus on structure (transitions, sentence flow) rather than vocabulary
• Example: "The LLM utilizes attention mechanisms" → keep "attention mechanisms" but change "utilizes" to "uses"

Edge Case 3: Already Human-Like Text

• If detector scores <70% AI, minimal changes needed
• Focus only on obvious patterns (em-dashes, "By [gerund]")
• Risk: over-editing good text makes it worse

Cross-Article Consistency: When editing multiple articles, vary replacements across articles; don't use "The key insight:" in every one.