Skip to content
Primores AI wiki

Building Brand Voice Skills for Claude: The Technical Foundation

Building Brand Voice Skills for Claude

TL;DR: Brand voice consistency in AI writing isn’t about better prompts — it’s about teaching Claude through examples. Skills work because LLMs perform in-context learning: they pattern-match from examples in the prompt rather than following abstract rules. Structure your skill with few concrete examples in SKILL.md and detailed references in separate files. Be specific — vague guidelines produce inconsistent output.

The Problem

Every time someone on your team opens Claude and re-explains your brand voice, they’re paying a tax:

  • Re-pasting the style guide
  • Re-typing “write this like us”
  • Getting output that sounds like a different brand

A well-built skill eliminates this. But most advice on building brand voice skills treats them as magic files. Understanding why they work — the underlying LLM mechanics — helps you build better ones.

Why Examples Beat Rules: In-Context Learning

LLMs don’t follow rules the way humans do. They perform in-context learning (ICL) — pattern-matching from examples provided in the prompt without any fine-tuning.

Research from Google DeepMind’s Many-Shot In-Context Learning paper (NeurIPS 2024) demonstrates:

  • More examples = better performance — scaling from few-shot to many-shot produces “significant performance gains across a wide variety of generative and discriminative tasks”
  • Examples overcome pretraining biases — the model’s default patterns can be overridden by sufficient examples
  • Format matters more than explanation — the structure of your examples constrains output more than verbose rules

The Prompt Engineering Guide summarizes the key finding: “the format you use plays a key role in performance, even if you just use random labels, this is much better than no labels at all.”

★ Insight ───────────────────────────────────── What this means for brand voice: When you show Claude “Here’s on-brand copy, here’s off-brand copy,” you’re providing training signal in the prompt itself. Each example narrows the distribution of acceptable outputs. One example = wide variance. Five examples = tighter band. The skill “compounds” not through magic but through statistical constraint. ─────────────────────────────────────────────────

The Instruction Hierarchy

Why do skills override conversational context? Research on The Instruction Hierarchy shows LLMs can be trained to prioritize different instruction sources:

  1. System/skill-level prompts — highest priority
  2. User prompts — medium priority
  3. In-context information — lowest priority

Skills work because they occupy the privileged position. But this hierarchy only functions when instructions are specific and unambiguous. Vague skill content competes poorly against specific user requests.

The Underspecification Problem

Research on What Prompts Don’t Say found a critical warning:

“Underspecified prompts are 2x more likely to regress over model or prompt changes, sometimes with accuracy drops exceeding 20%.”

Brand voice is inherently fuzzy. “Sound confident but not arrogant” means different things to different models and different contexts. This is why example-based skills outperform rule-based skills — examples are specifications, rules are abstractions.

Skill Architecture: SKILL.md + Reference Files

Per Anthropic’s official documentation, skills should follow this structure:

brand-voice/
├── SKILL.md           # Main instructions (lean, focused)
├── examples/
│   ├── on-brand.md    # Good examples by channel
│   └── off-brand.md   # Anti-patterns to avoid
├── references/
│   └── style-guide.md # Full brand guidelines
└── banned-words.txt   # Words that kill the voice

Critical insight from the docs: “skill.md should contain exactly four things: a brief description of what the skill does, the inputs it requires, the numbered process steps in execution order, and the output specification. Everything else — context, rules, examples, templates, domain knowledge — should live in reference files.”

Why? Skills load progressively:

  1. Description — always in context (for matching)
  2. SKILL.md body — loads when skill seems relevant
  3. Reference files — load only when Claude navigates to them

Large example collections in SKILL.md waste tokens. Put them in reference files.

A Proper Brand Voice SKILL.md

---
name: brand-voice
description: Applies [Company] brand voice to all written content. Use when writing marketing copy, social posts, email campaigns, or any customer-facing content. Ensures consistent tone across team members.
---


# Brand Voice Skill


## Who we write for
[Your ICP in 2-3 sentences — specific, not generic]


## Our voice in three words
1. [Word] — means [specific behavior]
2. [Word] — means [specific behavior]
3. [Word] — means [specific behavior]


## Process


1. **Read context**: Understand the content type and channel
2. **Load channel rules**: See `examples/[channel].md` for channel-specific patterns
3. **Check banned words**: Never use words in `banned-words.txt`
4. **Match examples**: Output should pattern-match the examples in `examples/on-brand.md`
5. **Verify**: Does this sound like examples, or like `examples/off-brand.md`?


## Output format
- Match the tone of examples exactly
- When uncertain, choose the more specific/concrete phrasing
- Flag any phrases that might need human review

The 5-Example Method

The most practical workflow for building the skill:

  1. Collect 5 pieces from the last 90 days:

    • 2 that felt completely on-brand
    • 1 that made you cringe
    • 2 that were just okay

  2. Run this prompt with your team:

    “Here are examples of our best and worst content. Ask us questions to define our brand voice. Challenge anything generic. Then turn it into a 1-page guide.”

  3. Extract patterns, not rules:

    • How do we open pieces? (example)
    • How do we handle humor? (example)
    • What do we never say? (specific words)

  4. Structure into reference files:

    • examples/on-brand.md — the 2 best pieces, annotated
    • examples/off-brand.md — the cringe piece, with notes on what went wrong
    • banned-words.txt — words that kill the voice instantly

Channel-Specific Sections

Your LinkedIn voice isn’t your email voice. Add channel files:

examples/
├── linkedin.md      # Hook styles, length, tone
├── email.md         # Subject lines, openings, CTAs
├── instagram.md     # Caption patterns, emoji rules
└── landing-page.md  # Hero copy, benefit framing

Each file should contain:

  • 2-3 real examples from that channel
  • Character limits and format constraints
  • What’s different about tone here vs. other channels

The Banned Word List

Every brand has words that kill the voice instantly:

banned-words.txt
# Add one word per line
# Include note on why


synergy          # corporate speak
game-changer     # overused, says nothing
leverage (verb)  # jargon
robust           # meaningless modifier
utilize          # just say "use"

The compounding effect: Each time Claude outputs something off-brand, add the offending word. The list grows, constraints tighten, output improves — without anyone writing new rules.

Maintenance Model

A skill with no owner gets stale. Assign one person to:

  • Weekly: Review content that went out, note any off-brand moments
  • Update: Add new banned words, better examples, tighter constraints
  • Test: Run the same prompt across team members, compare outputs

Time investment: 5-10 minutes per week. The skill compounds.

Testing Your Skill

Run the same three prompts across your team:

  1. A social caption for [specific product/announcement]
  2. An email subject line for [specific campaign]
  3. A campaign intro paragraph for [specific audience]

Compare outputs. If they sound like 3 different brands, the skill needs:

  • More examples (examples/on-brand.md)
  • Stronger anti-examples (examples/off-brand.md)
  • More specific constraints (banned-words.txt)

If they sound like one brand, the skill is working.

Common Mistakes

Mistake 1: Rules instead of examples

Wrong: “Write in a confident but approachable tone” Right: [Paste actual on-brand paragraph] — “Match this tone”

Mistake 2: Everything in SKILL.md

Wrong: 2000-word SKILL.md with all examples embedded Right: Lean SKILL.md with references to example files

Mistake 3: Generic descriptions

Wrong: “Helps with brand content” Right: “Applies [Company] brand voice when writing LinkedIn posts, email campaigns, or landing page copy. Use when any team member creates customer-facing content.”

Mistake 4: Assuming automatic loading

Wrong assumption: Claude reads the skill before every response Reality: Claude sees skill descriptions, loads full content only when relevant. Make your description specific enough to trigger correctly.

Key Takeaways

  • Skills work through in-context learning — pattern-matching, not rule-following
  • Examples beat rules — format and structure constrain more than explanations
  • Be specific — underspecified prompts regress 2x more often
  • Use reference files — keep SKILL.md lean, put examples elsewhere
  • Maintain continuously — assign ownership, compound the banned word list
  • Skills occupy privileged position in instruction hierarchy, but only when specific
  • tools/claude-skills — Complete technical guide to building Skills
  • glossary/prompt-engineering — Techniques for effective prompting
  • marketing/ai-marketing-case-studies — Real-world AI marketing examples
  • marketing/ai-tells-in-sales-copy — The editing/audit-side discipline. Brand-voice Skills define what to sound like (LLM-instruction side); ai-tells defines what not to sound like in client-facing copy (review-pass side). The two are complementary disciplines — instructions plus audits.
  • marketing/ai-human-voice-prompting — The techniques-and-empirics layer underneath this page. This page covers the SKILL.md architecture and 5-Example Method; ai-human-voice-prompting covers the six broader techniques (voice-profile document, few-shot, banned-words, lived-experience anchors, voice-from-audio, audience-mode review) plus the 2026 detection landscape and platform-specific tactics for LinkedIn, X, TikTok, and cold email.

Sources