Brand Voice

Brand Voice

Most "brand voice" is vibes — "friendly, modern, approachable". That's useless. This skill extracts a learned profile from 5+ real writing samples, then applies it to new content.

Instructions for Claude

This skill has three modes: extract (build a new voice profile), apply (use a saved profile), list (show saved profiles).

Detect mode from the user's input. If unclear, ask.

---

Mode: Extract

Step 1: Identify the Brand and Where to Save

Ask:

1. Brand name (used as the filename, e.g., "mintlooks" → mintlooks.md)
2. Save location: project (.claude/brand-voices/) or global (~/.claude/brand-voices/)?

Step 2: Collect Samples

Ask for AT LEAST 5 writing samples. More is better. Tell the user:

I need at least 5 samples to extract a voice. The more diverse the better:
- LinkedIn posts
- Instagram captions
- Newsletter excerpts
- Brand guidelines
- Customer-facing emails
- Existing video scripts

Paste them here, one per message or all in one — your call. Tell me when
you're done.

If the user provides fewer than 5, push back:

That's [N] samples — fewer than I'd want for a reliable profile. With this
few I'll guess more than I learn. Either:
1. Send more samples (recommended)
2. Continue with what we have, knowing the profile will be rougher

Step 3: Extract the Profile

Read all samples carefully. Extract:

Tone signature (3–5 words): Look at HOW the brand says things, not what they say. Examples:

• "Casual + technical" (Linear's docs)
• "Earnest + builder-first" (Vercel)
• "Sardonic + plain-spoken" (Honest Co's old captions)

Sentence structure patterns:

• Average sentence length (count from a few samples)
• Are paragraphs short or long?
• Does the brand use sentence fragments? Single-word sentences?
• Does it ask rhetorical questions? Use second-person ("you")?

Vocabulary signatures:

• Words/phrases the brand uses repeatedly (e.g., "ship", "build", "we")
• Words the brand seems to avoid (look for absent-but-expected — e.g., a tech brand never saying "synergy" or "leverage")
• Punctuation tics (em-dashes, semicolons, ellipses, line breaks as paragraphs)

Rhetorical moves:

• Does it open posts with stories? Stats? Questions? Bold claims?
• Does it use lists or prose?
• Does it use first-person ("I"), royal-we ("we"), or impersonal voice?
• Where does it land emotionally — humor, conviction, vulnerability, authority?

Do / Don't rules (3–5 each): Concrete behaviors, not vibes. "Do: open with a specific moment, not a general statement" beats "Do: be authentic".

Step 4: Write the Profile

Save to [location]/brand-voices/[brand].md:

---
brand: [Brand Name]
extracted: [date]
samples: [N]
---

# [Brand Name] — voice profile

## Tone signature
[3–5 word phrase]

## Sentence structure
- Avg sentence length: [N] words
- Paragraph style: [short/medium/long, line-break-heavy/prose]
- Fragments: [used/avoided] — [example if used]
- Questions: [rhetorical/genuine/avoided]
- POV: [I / we / you / impersonal]

## Vocabulary signatures
**Uses repeatedly:**
- [word/phrase]
- [word/phrase]

**Avoids:**
- [word/phrase] — [why we noticed]

**Punctuation tics:**
- [tic]

## Rhetorical moves
**Opens posts with**: [pattern]
**Builds arguments via**: [stories / lists / claims]
**Emotional register**: [where it lands]

## Do
- [concrete rule]
- [concrete rule]
- [concrete rule]

## Don't
- [concrete rule]
- [concrete rule]
- [concrete rule]

## Example reframings

The same idea, written 3 ways — only the third is on-brand.

**Idea**: [generic statement]

**Off-brand**: "[generic version]"
**Off-brand**: "[different generic version]"
**On-brand**: "[in the voice]"

## Source samples

[Quote a 1–2 line excerpt from each sample, with a tag like "LinkedIn post, Mar 2026"]

Step 5: Confirm and Save

Show the profile to the user. Ask:

Here's the profile I extracted. Anything off? I can refine it before saving.

Save to: [path]

On confirmation, write the file.

mkdir -p [location]/brand-voices
# Write the profile file

---

Mode: Apply

Step 1: Identify the Voice and the Content

Ask:

1. Which voice? (Or auto-detect from [user input] in [brand] voice.)
2. What content needs to be in that voice? (Paste or describe.)

Step 2: Load the Profile

# Try project-local first, then global
cat .claude/brand-voices/[brand].md 2>/dev/null || cat ~/.claude/brand-voices/[brand].md

If not found:

No voice profile found for "[brand]". Either:
1. Run /brand-voice extract to build one
2. List available voices: /brand-voice list

Step 3: Rewrite

Apply the profile to the content. Show:

• The original (for reference)
• The rewritten version
• 1–2 sentences explaining what you changed and why (which rules from the profile drove the changes)

Step 4: Iterate

Offer:

Want me to:
1. Adjust toward [more/less of some quality]
2. Try a different voice
3. Done

---

Mode: List

ls -la ~/.claude/brand-voices/ 2>/dev/null
ls -la .claude/brand-voices/ 2>/dev/null

Display the brands found, with a 1-line preview of the tone signature for each.

---

Hard Rules

1. Voice = patterns, not adjectives. "Friendly" is not a voice. "Opens with a specific time + place, uses 'we' not 'I', never uses exclamation marks" is.

1. Don't extract from <5 samples without warning the user. Small sample = unreliable profile.

1. Concrete rules only. Every Do/Don't must be observable in writing. "Be authentic" fails this test. "Don't use the word 'authentic'" passes.

1. The profile is a learning artifact, not a system. When applying it, treat the rules as strong defaults but break them if the content demands it. Note when you broke a rule and why.

1. Never invent samples. When showing "source samples", quote real lines from what the user provided.

Error Handling

Brand profile already exists:

A profile for "[brand]" exists at [path].

Options:
1. Overwrite — extract fresh from new samples
2. Append — add new samples to the existing profile
3. Cancel

User asks for a voice they describe but don't have samples for:

I can write in a *style* you describe (e.g., "punchy, contrarian, lots of
em-dashes"), but I won't save that as a "voice profile" — those need real
samples. Want me to:
1. Apply your description as a one-shot style (no save)
2. Help you collect samples and extract a real profile

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ImTaegan
• Source: ImTaegan/claude-ugc-skills
• License: MIT

Install and usage instructions live in the source repository linked above.