voice-injection-rewriter

What this skill does

# Voice Injection Rewriter

> Transform AI-generated text into authentic, voice-consistent content — not by faking humanity, but by applying real voice patterns from a specific person or brand.

## When to Use This Skill

Use this skill when you need to:

- **Rewrite AI drafts** to match a specific person's voice
- **Strip AI fingerprints** from generated content before publishing
- **Enforce voice consistency** across AI-assisted content production
- **Bridge the gap** between AI efficiency and authentic brand expression
- **Post-process any AI output** (blog posts, emails, social, landing pages)

This skill is NOT a generic "humanizer." It requires voice input — either from `brand-voice-learner` output, ClawFu brand memory, or voice samples you provide.

**Why this matters:** Generic humanizers add fake imperfections (random typos, forced contractions) to trick detectors. That's an arms race you lose. This skill applies YOUR voice patterns — which produces naturally human text because it IS the voice of a real human.

---

## Methodology Foundation

**Core Principle:** AI text sounds artificial not because it lacks typos, but because it lacks a specific person's vocabulary, rhythm, opinions, and structural habits. The fix is voice injection, not cosmetic imperfection.

**Sources:**
- NN/g Voice and Tone research
- Brand voice analysis methodology (ClawFu `brand-voice-learner`)
- AI detection pattern research (GPTZero, Originality.ai signal analysis)
- WhatsApp IA NDD Camp community insights on AI content workflows

**The 201 insight:** Most people try to make AI text "sound human" (generic). The actual skill is making AI text "sound like ME" (specific). The first is commodity. The second is craft.

---

## What Claude Does vs What You Decide

| Claude Does | You Decide |
|-------------|------------|
| Applies voice patterns to rewrite text | Which voice profile to use |
| Strips AI detection signals | How far to deviate from the original |
| Identifies voice mismatches | Final approval of tone and accuracy |
| Suggests voice-consistent alternatives | Where factual precision overrides voice |
| Runs enforcement checklists | Publication context and audience |

---

## Instructions

### Prerequisites

Before using this skill, you need ONE of these:

| Voice Source | How to Get It | Quality |
|-------------|---------------|---------|
| **ClawFu brand memory** | `get_brand [name]` — pre-stored voice profile | Best |
| **brand-voice-learner output** | Run the analysis skill on 10+ writing samples | Very good |
| **3-5 writing samples** | Paste examples of the target voice | Good |
| **Voice description** | "Analytical, dry, French, no buzzwords, short sentences" | Minimum viable |

---

### Phase 1: Voice Profile Load

Start by establishing the voice you're targeting.

**If using ClawFu brand memory:**
```
Load brand voice for [name]. Apply it to rewrite the following text.
```

**If providing samples:**
```
Here are 3 examples of my writing voice:
[Sample 1]
[Sample 2]
[Sample 3]

Extract the voice patterns, then rewrite this AI-generated text to match:
[AI text]
```

**If describing voice:**
```
My voice profile:
- Register: [analytical/casual/authoritative/conversational]
- Sentence length: [short punchy / mixed / long flowing]
- Contractions: [always / sometimes / never]
- Opinions: [strong and stated / subtle / neutral]
- Vocabulary: [technical / accessible / mixed]
- Signature moves: [specific habits — e.g., "Ce n'est pas X. C'est Y."]
- Forbidden words: [list any words you never use]

Rewrite this to match:
[AI text]
```

---

### Phase 2: AI Fingerprint Detection

Before rewriting, identify what makes the text sound AI-generated.

#### AI Detection Signals (What to Flag)

**Structural signals:**
- Uniform sentence length (AI defaults to 15-20 words per sentence)
- Predictable paragraph structure (topic sentence → 3 supporting points → transition)
- Symmetrical lists (every bullet same length, same grammatical structure)
- Perfect parallelism in headers

**Vocabulary signals:**
- Corporate buzzwords nobody actually says: "delve," "landscape," "leverage," "robust," "streamline," "facilitate," "comprehensive"
- Filler openings: "In today's rapidly evolving...", "It's important to note that...", "In order to..."
- Over-hedging: "It's worth noting," "One might argue," "It could be said"
- Transition addiction: "Moreover," "Furthermore," "Additionally," "However" (every paragraph)

**Tonal signals:**
- No opinions (AI stays neutral by default)
- No contractions (AI under-uses them)
- Equal weight to everything (AI doesn't emphasize or deprioritize)
- Sanitized language (no edge, no personality, no risk)

**Flag template:**
```
AI FINGERPRINT SCAN:
□ Uniform sentence length? [yes/no]
□ Buzzword count: [N] instances
□ Transition density: [N] per paragraph
□ Contractions present? [yes/no]
□ Opinions/personality present? [yes/no]
□ Any sentence you'd never actually say? [list]
```

---

### Phase 3: Voice Injection Rewrite

Apply the voice profile to the flagged text. This is the core operation.

#### The 5-Pass Rewrite

**Pass 1 — Kill AI vocabulary**

Replace every AI-default word with the voice-appropriate equivalent:

| AI Default | Generic Human | Voice-Specific (example) |
|------------|--------------|-------------------------|
| "Delve into" | "dig into" | Depends on voice — could be "creuser," "regarder de plus près," "explorer" |
| "Leverage" | "use" | Voice may prefer "exploiter," "utiliser," or "s'appuyer sur" |
| "Landscape" | "space" | Voice may prefer "marché," "écosystème," or just name the thing |
| "Robust" | "strong" | Voice may prefer "solide," "fiable," or a domain-specific term |
| "Moreover" | "Plus" | Voice may prefer "Et," "D'ailleurs," or nothing at all |
| "Ensure" | "make sure" | Voice may prefer "vérifier," "s'assurer," or just drop it |
| "Comprehensive" | "full" | Voice may prefer "complet," "exhaustif," or cut the word entirely |
| "It is important to note" | (cut) | Replace with direct statement of the thing |

**Pass 2 — Break rhythm uniformity**

- Find the longest streak of same-length sentences. Break it.
- Add 2-3 short punches per section (under 8 words). "That's the point." "Not even close."
- Allow 1-2 longer flowing sentences if the voice uses them
- Fragments are OK if the voice uses fragments
- Dashes—for emphasis—if the voice does that
- Parentheticals (if the voice thinks while writing)

**Pass 3 — Inject voice-specific patterns**

Pull from the voice profile:
- Signature phrases or rhetorical moves
- Preferred sentence openers
- Opinion-stating patterns ("Ce n'est pas X. C'est Y." / "Here's the thing" / "Look,")
- Domain-specific vocabulary the voice always uses
- Structural habits (bullets vs. flowing prose, numbered lists vs. paragraphs)

**Pass 4 — Contraction and register pass**

- Apply contraction level matching the voice (some voices never contract, some always do)
- Check register: is the formality level consistent with the voice?
- Verify pronoun usage matches (some voices are "nous," some are "je," some are "on")

**Pass 5 — Read-aloud test**

Read the full text aloud. For each sentence ask:
- Would [person] actually say this?
- Does it flow at their pace?
- Is there any word they'd never use?

If any sentence fails, rewrite that sentence from scratch in the voice.

---

### Phase 4: Output Format

#### Standard Output

```markdown
## Original (AI-generated)
> [The input text]

## Rewritten (Voice: [Name/Brand])
[The rewritten text]

## Changes Made
- **Vocabulary**: [3-5 key word swaps]
- **Rhythm**: [What changed structurally]
- **Voice markers**: [What was injected from the voice profile]
- **Removed**: [What AI signals were stripped]

## Voice Consistency Score
[High/Medium/Low] — [one-line explanation]
```

#### Quick Mode (For batch processing)

If rewriting multiple pieces, skip the detailed output:

```markdown
[Rewritten text only]

---
*Voice: [Name] | Swaps: [N] | Score: [H/M/L]*
```

--