Humanizer

Humanizer · Make Text Sound Like a Human Wrote It

Operating principle

LLMs regress to the statistical mean. Humans are weird, specific, and inconsistent.

The fundamental AI tell: text that emerges from nowhere, addressed to no one, with no stake in its claims. Human writing reveals a mind behind it. If the reader can't picture a specific person writing this, it's not done.

A ruthless editor who despises AI slop. Don't just remove bad patterns. Replace them with something that has a pulse.

When to invoke

• After section-drafter produces workspace/{task_id}/draft.md (Stage: section-drafted)
• Before fact-check-and-citation (so fact-checker validates human-form claims)
• Auto-triggered by repair-orchestrator when AI-Slop score >50 OR ≥5 patterns detected
• User can manually invoke /humanize on existing prose

Three modes

Five voices × five purposes (25 combinations)

Voice + purpose combine layered. Voice handles personality; purpose handles register.

The 43 patterns

Full descriptions in references/style/ai-tells-43.md. Summary table:

Content patterns (P1-P8)

• P1 Significance Inflation ("represents a pivotal moment")
• P2 Notability Name-Dropping (listing pubs vs reporting what they said)
• P3 Superficial -ing Phrases ("ensuring", "fostering")
• P4 Promotional Language ("nestled", "vibrant", "boasts")
• P5 Vague Attributions ("experts argue", "studies show")
• P6 Formulaic Challenges Sections ("Despite X, faces Y")
• P7 AI Vocabulary cluster ("delve", "additionally", "crucial", "landscape", "tapestry")
• P8 Copula Avoidance ("serves as" instead of "is")

Language & style (P9-P18)

• P9 Negative Parallelisms ("not just X, it's Y")
• P10 Rule of Three (forced triadic lists)
• P11 Synonym Cycling (entity referred to 3 ways in a paragraph)
• P12 False Ranges ("from X to Y" non-spectrum)
• P13 Em Dash Ban — zero tolerance (U+2014)
• P14 Boldface/Formatting Overuse
• P15 Structured List Syndrome (every paragraph is a bullet list)
• P16 Title Case in Headings
• P17 Curly Quotes (ChatGPT fingerprint) + Oxford comma overuse
• P18 Formal Register Overuse (bureaucratic when conversational fits)

Communication (P19-P21)

• P19 Chatbot Artifacts ("I hope this helps", "Certainly!")
• P20 Knowledge-Cutoff Disclaimers ("As of my last training")
• P21 Sycophantic Tone ("Great question!")

Filler & Hedging (P22-P30)

• P22 Filler Phrases
• P23 Excessive Hedging (stacked: "could potentially possibly")
• P24 Generic Positive Conclusions ("future looks bright")
• P25 Hallucination Markers (overly specific fabricated dates)
• P26 Perfect/Error Alternation (mixed human+AI editing)
• P27 Question-Format Section Titles ("What makes X unique?")
• P28 Markdown Bleeding (**bold** in non-md contexts)
• P29 Comprehensive Overview Opening ("In this article, we will explore")
• P30 Uniform Sentence Length (15-25 words throughout)

Emerging 2026 (P31-P43)

• P31 Elegant Variation / Noun-Phrase Cycling
• P32 Collaborative Communication Leaking ("we will explore")
• P33 Placeholder Text ([Your Name], 2025-XX-XX)
• P34 Chatbot Reference Markup Leaking (citeturn0search0)
• P35 UTM Source Parameters from AI Tools (?utm_source=chatgpt.com)
• P36 Sudden Style/Register Shift (mixed authorship)
• P37 Overattribution / Source-Listing as Content
• P38 Paragraph-Reshuffling Immunity (swap-and-still-works)
• P39 "Whether" Summary Sentence Endings
• P40 Symbolic Gloss / Meaning-Telling
• P41 Infomercial Engagement Hooks ("The catch?", "The brutal truth?")
• P42 Erratic Inline Bolding
• P43 Treadmill Effect (low information density / restatement loops)

The Burstiness Principle

AI detectors measure "burstiness" — sentence length variance. Human writing has HIGH burstiness; AI has LOW.

Target patterns:

• Mix short (3-8 words), medium (12-20 words), long (25-40 words) in every paragraph
• Never have 3+ consecutive sentences of similar length
• Let a sentence run long when the thought needs room to breathe, winding through qualifications before landing
• A fragment, used sparingly, can land a point.

⛔ Do NOT raise burstiness by mass-inserting terse fragment-closers. Burstiness is a coefficient of variation (σ/μ) with no floor on short sentences, so stacking 2-5 word fragments ("Pure leakage." "Stage first." "Two in three, gone.") mechanically lifts the score AND drops ai_slop_score under the gate — but a human/LLM reviewer reads a run of terse closers as the single most visible AI tell (it flagged exactly this on the 2026-06-29 batch). Cap: at most ~1 fragment per 150-200 words, and NEVER as a repeated paragraph-ending device. Raise burstiness mainly by letting some sentences run LONG, not by piling up fragments. If you can only clear <20 by stacking fragments, the draft needs real rewriting, not staccato.

The Perplexity Principle

AI text has LOW perplexity (predictable word choices). Human text has HIGHER (more surprising).

Increase perplexity naturally:

• Choose the second or third word that comes to mind, not the first
• Use domain-specific jargon or slang appropriate to the audience
• Make unexpected analogies from personal experience
• Occasional informal transitions ("Anyway,", "So here's the thing:", "Look,", "Thing is,")

Soul Injection (11 techniques)

Full descriptions in references/style/soul-injection-11.md. Brief:

1. Have actual opinions — "This API is frustrating" > "The API has limitations"
2. Honest uncertainty — "I'm not sure this is right, but..."
3. Specific sensory detail — "2am with cold coffee and a stack trace" > "complex"
4. Shared experience callback — "You know that feeling when..."
5. Allow tangents — brief digressions signal thinking, not algorithm
6. Dramatic paragraph variation — 4 sentences, then 1 line. Like this.
7. Imperfect opening — start mid-thought: "So I was looking at the logs..."
8. Break parallel structure — three items same grammar, fourth different
9. Use callbacks — reference earlier in piece
10. Self-correct — "well, technically..." signals real-time thinking
11. End without wrapping up — not every piece needs a neat conclusion

AI-Slop scoring formula

Copy code
score = 4 × patterns_hit + 25 × (1 - burstiness_normalized) + 15 × vocabulary_blacklist_ratio
clamped to 0-100

Score interpretation:

Implementation: scripts/validate/ai_slop_score.py.

Workflow (in /article pipeline)

Phase 1: Detect

bashCopy code
python -m scripts.validate.ai_slop_score \
    --input workspace/{task_id}/draft.md \
    --json > workspace/{task_id}/ai-slop-pre.json

Read result. If score < 20: skip humanizer (already clean).

Phase 2: Spawn humanizer agent

Invoke agents/humanizer.md (Edit-only tool whitelist — physically cannot Write). Pass:

• --mode rewrite
• --voice {brand-guideline.voice} (from project)
• --purpose {brand-guideline.purpose}
• --score (compute final score)
• --iterate 3 (convergence)

Humanizer agent reads draft.md + this catalog + soul-injection-11.md, then Edits the draft.md in place.

Phase 3: Re-detect

bashCopy code
python -m scripts.validate.ai_slop_score \
    --input workspace/{task_id}/draft.md \
    --json > workspace/{task_id}/ai-slop-post.json

Compare pre vs post. If post-score still >50, escalate to repair-orchestrator.

Phase 4: Frontmatter handoff

Update workspace/{task_id}/state.json:

• stage: humanizer-complete
• ai_slop_pre: 67
• ai_slop_post: 18
• iterations: 2

Next stage: fact-check-and-citation.

Final quality checklist (humanizer agent enforces)

1. Read aloud test — does it sound like a person talking?
2. Check the opening — boring overview sentence? Hook instead
3. Check the ending — generic positive wrap-up? Cut or replace with specific
4. Count "delves" — kill any AI blacklist words
5. Zero em dashes — search for U+2014, replace with commas/colons/hyphens
6. Sentence length audit — 3+ similar in a row → vary
7. "Who wrote this?" test — can reader picture a specific person?

What this skill does NOT do

• ❌ Modify quoted material, Citation Capsules, References section
• ❌ Make claims more specific by inventing facts (that's fact-checker's red line)
• ❌ Change headings (linker rule applies — headings sacred)
• ❌ Add new sections (only existing prose)
• ❌ Skip the post-rewrite re-score check

Composition

Copy code
section-drafter
  ↓ draft.md (Stage: section-drafted)
ai_slop_score (detect mode)
  ↓ ai-slop-pre.json
humanizer agent (rewrite mode)
  ↓ draft.md edited in place (Stage: humanizer-complete)
ai_slop_score (re-detect)
  ↓ ai-slop-post.json
  ↓
  → if post-score <50: fact-check-and-citation
  → if post-score ≥50: repair-orchestrator level 2
  → if post-score ≥70: repair-orchestrator level 3 (section regen)

See also

• references/style/ai-tells-43.md — full pattern definitions with triggers + fixes + examples
• references/style/soul-injection-11.md — 11 humanization techniques
• references/style/banned-words.md — AI vocabulary blacklist
• references/style/voices/{voice}.md — per-voice detailed rules
• references/style/purposes/{purpose}.md — per-purpose modifiers
• agents/humanizer.md — the Edit-only subagent
• scripts/validate/ai_slop_score.py — formula implementation