voice-extractor

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: voice-extractor
description: Extract and document someone's authentic writing voice from samples. Use when someone needs a "voice guide," wants to capture their writing DNA, or needs to train AI to write in their style. Also useful for ghostwriting, brand voice documentation, or onboarding writers.
---

# Voice Extractor

AI-generated content all sounds the same. The fix isn't better prompts — it's teaching the AI how you actually communicate.

This skill extracts your communication DNA from writing samples and produces a Voice Guide: documented, tested, and ready to use.

---

## Mode

Detect from context or ask: *"Quick voice snapshot, full Voice Guide, or full guide with examples?"*

| Mode | What you get | Best for |
|------|-------------|----------|
| `quick` | Top 5 voice characteristics + 3 do/don't rules | Fast style reference, single piece |
| `standard` | Full Voice Guide: tone, vocabulary, rhythm, structure | AI training, ghostwriting, brand documentation |
| `deep` | Full Voice Guide + 10 sample rewrites + writing rules checklist + AI training examples | Onboarding writers, building a brand voice system |

**Default: `standard`** — use `quick` if they just need a fast reference. Use `deep` if they're onboarding a ghostwriter or building a content team.

---

## Context Loading Gates

**Before extracting, collect:**

- [ ] **Writing samples** — minimum 3 samples OR 500 total words (see priority list below)
- [ ] **Purpose of voice guide** — AI training? Ghostwriter onboarding? Team alignment?
- [ ] **Confidence zones** — Any topics where they want to sound more/less authoritative?
- [ ] **Known anti-patterns** — Any words or phrases they already know they want to avoid?

**Sample priority (most → least authentic):**
1. Casual Slack or email (raw, unedited voice)
2. Podcast or call transcript
3. LinkedIn posts or articles
4. Website copy (often edited, less authentic)

**Minimum sample gate:** If samples total under 500 words, stop:
> "These samples are too short to extract reliable patterns. Please add 2-3 more — emails, Slack messages, or transcripts work best. The messier and more casual, the better."

Do not attempt full extraction from under 500 words. Offer quick mode instead.

---

## Phase 1: Sample Quality Assessment

Before extracting, reason through:

1. **Sample authenticity:** Are these samples from edited/polished contexts (website, press) or raw contexts (Slack, email)? More polish = less authentic voice.
2. **Sample variety:** Do the samples cover different contexts (professional, casual, educational)? Single-context samples produce single-dimension voice guides.
3. **Exclusion check:** Identify and flag patterns that are NOT the authentic voice:
   - Platform formatting tics (LinkedIn line breaks, Twitter brevity forcing)
   - Typos and autocorrect errors
   - Phrases borrowed from others (quotes, retweets)
   - Unusually formal writing (legal docs, press releases)
4. **Sample size adequacy:** Is there enough material for full mode, or should I use quick mode?

Output a sample assessment:
> "I have [X samples / Y words] to work with. Quality: [high/medium — why]. I'll use [full/quick] mode. Excluding: [any patterns and why]."

---

## Phase 2: Core Energy Extraction

Identify the fundamental communication mode:

**Role:**
- Teacher (breaks things down systematically)
- Challenger (pushes back on assumptions)
- Cheerleader (builds confidence and momentum)
- Straight-shooter (cuts through BS efficiently)

**Default energy:**
- Calm authority ("Here's what works.")
- High enthusiasm ("This is exciting — let me show you.")
- Understated confidence ("I've seen this a hundred times.")

**Recurring themes:** What topics appear unprompted across samples? These are the things they actually care about.

---

## Phase 3: Phrase Extraction (Systematic)

Scan all samples and extract:

**Transition phrases** (how they shift topics):
- Quote exact examples from samples
- Pattern: "Here's the thing...", "W