---
name: voice-master
description: >
  Apply the user's voice when writing or editing content. Activate when
  the user says "use the voice-master skill", "write this in my voice",
  "make this sound like me", or any equivalent instruction. Use for any
  content type: professional proposals, long-form writing, social posts,
  emails, or short-form copy. Do NOT activate for code, technical docs,
  or any output the user hasn't asked to be written in their voice.
---

## THIS IS A TWO-PART SKILL — BOTH PARTS ARE MANDATORY

1. **Voice matching** — use exemplars to match tone, vocabulary, rhythm
2. **Anti-AI-slop pass** — review output for AI tells and fix them

You MUST apply BOTH passes on every use. Never apply exemplar matching
without the slop pass. Never skip the slop pass because "the voice sounds
right." The slop pass catches what exemplars can't — structural patterns,
banned words, em dash abuse, uniform sentence length.

If you are editing existing text (not generating): run the full Step 6
audit. If generating: Steps 1-5 produce the draft, Step 6 audits it.
Both paths end with the audit. No exceptions.

---

## Companion Files

This is the **project-level** skill (workflow + AI-tell audit rules). The
**user-level** companion at `~/.claude/skills/voice-master/` has the
exemplars and voice-dimensions that this skill references. Always load
both locations when using this skill:

- **Exemplars:** `~/.claude/skills/voice-master/exemplars/` (index.md,
  social.md, professional.md, longform.md)
- **Voice dimensions:** `~/.claude/skills/voice-master/voice-dimensions.md`

## Overview

Generates content that sounds like the user — not an AI impression of the
user. The exemplars are the primary source of truth. The voice-dimensions
file is fallback guidance for edge cases the exemplars don't cover.
**When exemplars conflict with voice-dimensions, exemplars win.**

The final output must pass an AI-tell audit before delivery. No exceptions.

## Quick Mode

For short content, skip the exemplar research pipeline. Quick Mode runs
Steps 5–7 only (generate + audit + deliver). The voice comes from
internalized markers rather than fresh exemplar analysis.

**Auto-triggers** (self-detected unless overridden):
- Content request is <200 words of output
- Content type is email, reply, DM, or Slack message
- Caller explicitly says "quick" / "quick mode" / "just dash this off"

**Audit-only** (editing existing text, not generating):
- Skip Steps 1–5 entirely
- Run Step 6 (full enhanced audit) on the provided text
- Deliver the cleaned version

**Quick Mode generation (Step 5 lite):**
- Characteristic openers where natural ("Here's the thing," "Frankly,"
  "Not for nothing")
- Match register to audience (use Register Quick Reference below)
- Evidence-first, no windup, no padding
- Do NOT load exemplar files — use internalized voice markers only

**Override:** If the user says "full mode" or "use exemplars," run the
complete 7-step workflow regardless of content length.

## Workflow

### Step 1 — Read the exemplar index

Read `exemplars/index.md`. This lists all curated exemplars with tone,
formality (1–5), and domain metadata.

### Step 2 — Select 3–5 matching exemplars

Match on:
- **Medium** — proposal, long-form, social, email, short-form?
- **Tone** — direct, analytical, reflective, blunt, measured?
- **Formality** — 1 (inner circle) to 5 (formal/public)?
- **Domain** — technical, strategy, AI, security, general?

For consulting proposals and professional writing: formality 3–4, select
from `professional.md` and `longform.md`. For social/short-form: formality
1–2, select from `social.md`.

### Step 3 — Read the matched exemplar files

Read the full text of each matched exemplar. Use them as stylistic
reference during generation — sentence structure, vocabulary level,
characteristic phrases, reasoning patterns. **Do not copy content.**

### Step 4 — Read voice-dimensions (if needed)

Read `voice-dimensions.md` for edge cases not covered by the exemplars —
register scaling by audience, tone, sentence structure rules, vocabulary,
humor handling.

### Step 5 — Generate

Write the content. Apply what you learned from the exemplars:
- Evidence-first, not windup
- Mix short punchy statements with longer reasoning
- Characteristic openers where natural: "Here's the thing," "Frankly,"
  "Not for nothing"
- No padding, no transitions for their own sake
- Register scales with audience — collared shirt for formal audiences,
  casual for inner circle. Drop profanity entirely at formality 3+.

### Step 6 — AI-tell audit (mandatory, every time)

Before delivering, scan the output and eliminate any of the following:

**Banned words and phrases:**
- delve, leverage, utilize, ensure, robust, seamless, streamline
- clean (as filler/intensifier, e.g. "clean architecture" — OK for
  literal cleanliness), smoking gun, landscape, ecosystem (when not
  literal), holistic, synergy, empower, elevate, harness, foster
- pivotal, crucial, enhance, underscore, vibrant, testament, showcase,
  intricate, evolving, navigate, journey
- it's worth noting, it is important to note, it's important to
- in conclusion, in summary, to summarize
- cutting-edge, game-changing, transformative, revolutionary
- I'd be happy to, certainly, absolutely, of course
- this allows us to, this enables, this ensures

**Banned structural patterns:**
- Opening with "I" on the first sentence
- Three-part lists that follow the exact same grammatical structure
- Passive voice overuse ("it was determined," "it should be noted")
- Rhetorical questions that aren't genuinely rhetorical
- Em-dash overuse and wrong format. Two rules, both hard:
  1. **Format:** NEVER spaced. `word — word` is the #1 AI punctuation tell.
     Acceptable: `word—word` (bare unicode) or `word--word` (double hyphen).
     Preferred: restructure the sentence to use a period, comma, colon, or
     semicolon instead. If you must use an em dash, no spaces touching it.
  2. **Frequency:** Reach for a comma, period, colon, or semicolon first.
     Every time. Em-dashes are for emphasis or asides where nothing else fits.
     Max 1-2 per page, not per paragraph. Stacking them is an AI fingerprint.
  The anti-slop pass MUST catch and fix any spaced em dashes. This is a
  hard fail — if ` — ` appears anywhere in the output, the audit failed.
- Hedging openers ("It's worth considering that...")
- Sycophantic acknowledgments before answering
- "Importance" sentences — delete sentences stating impact, legacy,
  significance, or broader trends. Show why it matters, don't state
  that it matters.
- Contrast structures — delete "It's not X, it's Y" / "Not A. Not B.
  But C" / "Despite this..." patterns. AI cadence markers.
- Vague authority claims — delete "experts say," "industry reports,"
  "many believe," "studies show" without a named source. No source =
  no claim.
- General claims without evidence — replace with specifics or delete.
  No proof = doesn't belong.
- Sentence length bias — strong bias toward sentences under 16 words.
  Not a hard ban (user's voice includes longer reasoning chains) but
  flag and split where possible. If you can say it shorter, do.
- Universally applicable statements — delete sentences that could apply
  to 1000+ topics unchanged. Not specific to THIS subject = padding.

**Test:** Read each paragraph out loud. If it sounds like a polished AI
response, it needs a rewrite. If it sounds like a person thinking through
something and writing it down, it's right.

### Step 7 — Deliver

Output the final content directly. No preamble, no explanation of what
you did, no "here's the content written in your voice." Just the content.

## Register Quick Reference

| Audience | Formality | Profanity | Exemplar source |
|----------|-----------|-----------|-----------------|
| Inner circle / Genesis | 1–2 | OK | social.md |
| Professional peers | 2–3 | OK | professional.md |
| Formal / cold outreach | 3–4 | None | professional.md, longform.md |
| Public content | 4–5 | None | longform.md |

## Examples

### Single paragraph, professional proposal
**Input:** "Use the voice-master skill to write an intro paragraph for a
security section of a consulting proposal. Owner/founder audience."

**Action:** Index → select professional.md formality 3 exemplars + longform
formality 3 → generate at formality 3-4 → AI-tell audit → deliver.

### Social post
**Input:** "Write a LinkedIn post about this in my voice."

**Action:** Index → select social.md formality 2 exemplars → generate at
formality 4 (public content) → AI-tell audit → deliver.

### Full document section
**Input:** "Rewrite this section in my voice."

**Action:** Read the section, identify medium and tone needed → full
exemplar workflow → AI-tell audit paragraph by paragraph → deliver.

### Quick email reply
**Input:** "Write a reply declining the meeting. Keep it short."

**Action:** Quick Mode (auto-detected: email + short) → generate at
formality 3 with voice markers from memory → AI-tell audit → deliver.

### Audit existing text
**Input:** "Run the AI-tell audit on this paragraph."

**Action:** Audit-only path → Step 6 on provided text → deliver cleaned
version with changes noted.