humanizer（人性化）

Humanizer

Rewrite AI-generated text so it reads like a skilled human wrote it. Not paraphrasing — structural transformation that changes how the text *feels*.

When to Use

• Text sounds robotic, generic, or overly polished
• Content flags AI detection tools
• Blog posts, marketing copy, or emails feel "off"
• Need to match natural human writing patterns
• After AI-generated first drafts before publication

When NOT to Use

• Code cleanup → use slop-remover agent instead
• Grammar/style review without rewriting → use copy-editor skill
• Writing in a specific brand voice → use brand-voice + style-writer
• Creating SEO content from scratch → use seo-content (has built-in humanization)
• After style-writer without providing the Style DNA report → the humanizer will treat intentional stylistic devices (anaphora, fragment pairs, zero-hedging, zero-em-dash styles) as AI tells and replace them with its own AI tells. If you must humanize voice-matched content, provide the Style DNA report (style-forensics output) as style context (see Phase 0)

Methodology

Four phases: Style Context → Diagnose → Transform → Verify.

Phase 0: Style Context Detection

Before diagnosing, check if the input has style constraints that override default rules.

Detection order:

1. User explicitly provides a Style DNA report (*-style-dna.md from style-forensics), brand-voice guide, or style constraints
2. User mentions the text was written with style-writer or matches a specific author
3. A *-style-dna.md file exists alongside the source content

If style context is found, read it and extract these overrides:

Zero counts are constraints. If the Style DNA reports 0 em dashes, 0 parenthetical asides, or 0 hedging instances, these are hard constraints — do not inject what the author never uses.

Store as STYLE_OVERRIDES for Phase 1-3. When no style context exists, all defaults apply and the skill works exactly as before.

If style context is detected but not provided: Ask with AskUserQuestion: "This text appears to be voice-matched. Do you have a Style DNA report or brand-voice file I should respect? Without it, I may rewrite intentional stylistic devices."

Phase 1: Diagnose

Read the input text and score it against five detection signals.

Signal scan:

Prohibited words scan:

• Check against references/prohibited-words.md
• Count banned word clusters per paragraph
• Flag banned phrases

Structural pattern scan:

• Intro-body-conclusion essay template?
• Mechanical "on one hand / on the other" balance?
• Formulaic "Challenges and Future Prospects" section?
• Everything grouped in threes?
• Negation-assertion pattern? Distinguish two forms:

- Mechanical (rewrite): "This isn't just X. This is Y." / "It's not about X. It's about Y." — full-sentence negation followed by full-sentence assertion with parallel structure - Rhetorical fragment (preserve): "Not a question. A system." / "Not theory. Practice." — short fragment pairs used for emphasis. These are a deliberate device, not an AI tell - If STYLE_OVERRIDES lists fragment pairs as a signature move → always preserve

• Bolded Inline Header: colon pattern?

Style-aware check (when STYLE_OVERRIDES exist):

Before flagging a pattern, check whether the style context explains it:

• Anaphoric repetition (e.g., "Context means X" ×5) → check Style DNA Signature Devices for anaphora/repetition patterns. If documented, mark as intentional device, not a detection signal
• Consistent fragment pairs → check Style DNA for fragment stacking patterns
• "Not X. Y." contrast patterns → check Style DNA Signature Devices for this exact pattern
• Uniform sentence structure in a passage → check Style DNA sentence length distribution for documented rhythm patterns

Rule: A pattern explained by style context is not an AI tell. Only flag patterns that exist *despite* (or *absent*) style context.

Report format:

## Diagnosis

**AI Confidence:** [High/Medium/Low] — [1-sentence justification]

**Signals flagged:**
- Perplexity: [OK/FLAG] — [detail]
- Burstiness: [OK/FLAG] — [detail]
- Entropy: [OK/FLAG] — [detail]
- Stylometry: [OK/FLAG] — [detail]
- Coherence: [OK/FLAG] — [detail]

**Banned words found:** [count] — [list top offenders]
**Banned phrases found:** [count] — [list]
**Structural patterns:** [list patterns detected]

Early exit: If AI Confidence is Low (0-1 signals flagged, no banned word clusters), stop and tell the user the text already reads human. Don't transform text that doesn't need it.

Phase 2: Transform

Apply transformations from references/humanization-playbook.md in this order:

Step 0: Lock factual anchors

Before any transformation, identify and protect content that must survive unchanged:

• Specific numbers, dollar amounts, percentages, dates
• Proper nouns: people, companies, products, places
• URLs, file paths, code blocks, technical terms
• Direct quotes and attributed statements
• Statistics and data points ("45,000 rows", "$14,000/month", "130,000 times")

These are untouchable. External humanizer tools destroy meaning by replacing "20 million views" with "20 black views" or "$14,000" with garbled text. Our advantage is understanding what words mean — never trade accuracy for style.

Step 1: Kill prohibited words

• Replace every banned word/phrase using substitution tables
• If 3+ banned words cluster in a paragraph, rewrite the whole paragraph
• Don't just swap synonyms — rephrase the sentence

Step 2: Break the structure

• If essay template detected: start with the most interesting point, cut the summary conclusion
• If everything is in threes: use the actual count (2, 4, 5, whatever fits)
• If mechanical negation-assertion found: delete the negation, keep only the positive claim. Preserve rhetorical fragment pairs ("Not X. Y.") — these are intentional devices, not AI tells
• If sections are suspiciously uniform length: merge short ones, split long ones
• Replace generic headers ("Overview", "Key Features") with specific ones

Step 3: Inject burstiness

• Identify runs of 3+ sentences with similar length
• Break them: shorten one to under 6 words, extend another past 25
• Add a single-sentence paragraph where emphasis helps
• Allow one fragment per 500 words

Step 4: Add voice (respect STYLE_OVERRIDES — skip any injection the style prohibits)

• Insert at least one opinion or stance per 500 words
• Add at least one caveat or limitation per 1000 words ("This won't work if...") — skip if STYLE_OVERRIDES indicate zero-hedging voice (author presents claims as settled fact)
• Replace smooth transitions with opinion bridges
• Add parenthetical asides where natural (like this) — skip or limit if STYLE_OVERRIDES specify low aside frequency (e.g., "max 2 per 1200 words")
• Use rhetorical questions sparingly (one per 1000 words max)

Step 5: Make it specific

• Replace "many companies" with named examples
• Replace "significantly improved" with actual numbers
• Replace "experts say" with named sources or remove the claim
• Convert abstract benefits to concrete outcomes

Step 6: Conversational pass (respect STYLE_OVERRIDES punctuation rules)

• Add contractions (it's, don't, won't, can't, we're)
• Use "you" and "your" for direct address
• Start some sentences with "And" or "But"
• End with prepositions when natural
• Add em dashes for rhythm and texture — skip if STYLE_OVERRIDES specify zero em dashes (some voices rely entirely on commas and periods)
• Write how you'd explain it over coffee — then tighten

Step 7: Final perplexity boost

• Read each paragraph aloud mentally
• Where every word feels predictable, swap one for an unexpected-but-accurate choice
• Prefer the specific over the generic, the concrete over the abstract
• Don't overdo it — one surprise per paragraph is enough

Phase 3: Verify

Run the detection self-test after transformation.

Verification checklist:

[ ] All factual anchors preserved (numbers, names, stats, quotes unchanged)
[ ] No 3+ banned words in any single paragraph
[ ] Sentence lengths vary (some under 5 words, some over 25)
[ ] At least one personal opinion or experience per 500 words
[ ] At least one specific number or named example per 500 words
[ ] At least one limitation or caveat per 1000 words (SKIP if STYLE_OVERRIDES = zero-hedging)
[ ] Not everything grouped in threes
[ ] No "In today's..." openings or "In conclusion..." closings
[ ] No mechanical negation-assertion patterns — rhetorical fragment pairs are OK
[ ] Each paragraph has different rhythm than neighbors
[ ] No section suspiciously same length as another
[ ] No engagement bait ("Let that sink in", "Read that again")
[ ] No sycophantic phrases ("I hope this helps")
[ ] Punctuation matches STYLE_OVERRIDES constraints (em dashes, asides, etc.)
[ ] Intentional rhetorical devices from style context preserved (anaphora, fragments)
[ ] Would I say this out loud to a smart friend?

If any check fails, return to Phase 2 and fix the specific issue.

Output Format

Present the humanized text with a brief change summary.

## Humanized Version

[Full rewritten text]

---

## Changes Applied

**Banned words replaced:** [count]
- [word] → [replacement] (×[count])

**Structural changes:**
- [Change 1]
- [Change 2]

**Voice injections:**
- [What was added and where]

**Verification:** All checks passed / [list remaining concerns]

Calibration by Content Type

Different content types need different humanization intensity.

Adjust transformation intensity based on this table. Documentation gets light touch; blog posts get full treatment.

Principles

1. Transform, don't paraphrase. Synonym swapping without structural change is shallow and detectable.
2. Imperfect on purpose. Natural writing has rough edges. Don't over-polish.
3. Specificity is the best humanizer. Concrete details are hard to fake and easy to trust.
4. Voice > vocabulary. Swapping words helps; adding genuine perspective transforms.
5. Respect the content. Don't inject humor into serious topics or casualness into formal contexts.
6. The coffee test. If you wouldn't say it to a smart friend over coffee, rewrite it.

Integration

Complementary skills in the writing pipeline:

[AI draft] → humanizer → copy-editor → brand-voice

• Before humanizer: Any AI-generated first draft
• After humanizer: copy-editor for grammar/style polish, brand-voice for voice consistency
• Instead of humanizer: seo-content (has built-in humanization for new content)
• For code: slop-remover agent (different domain, same concept)

Pipeline with voice-matched content:

[source samples] → style-forensics → Style DNA report
[AI draft] → style-writer (with Style DNA) → humanizer (with same Style DNA) → copy-editor

Never run humanizer after style-writer without providing the same Style DNA report as style context. The humanizer will destroy the voice work otherwise.