Every piece of content Claude produces draws from its general training — which means it sounds like everyone and no one at the same time. If you've ever read Claude's output and thought "this is fine but it could belong to any brand," the problem isn't Claude. It's the absence of structured brand context.

A Brand Skill solves this. It's a modular instruction set — a folder containing a master file and supporting references — that Claude reads before writing anything for your brand. You upload it once via Customize > Skills in Claude.ai, and it auto-loads whenever your prompt matches the skill's description. No need to paste brand guidelines into every conversation. No need to remind Claude about your tone for the hundredth time.

The skill has five core files, a curated reference corpus of your best (and worst) work, and a QA rubric that catches voice drift before it reaches your audience. Building the first version takes 4–6 hours. Maintaining it takes 1–2 hours per quarter.

A Claude Pro, Team, or Enterprise plan with Skills enabled. You upload skills via Customize > Skills in Claude.ai — no additional settings required. On Team and Enterprise plans, an admin may need to enable Skills at the org level first. You do not need Code Execution turned on. Skills are instruction files, not executable code.

Style guides, tone docs, brand decks, founder manifestos, campaign briefs, customer support templates, social media guidelines — whatever exists. Most of it will be scattered across drives and docs. That's normal. The first phase is gathering it all.

Someone who can definitively say "that sounds like us" and "that doesn't." If you're building this for a client, get a named approver before you start. Nothing kills a brand skill faster than building it to one person's taste and having someone else reject it.

A Skill is a folder containing a SKILL.md file plus supporting documents. You ZIP the folder and upload it via Customize > Skills. Once uploaded, Claude reads the skill's description and auto-loads it when your prompt matches what the skill covers. You don't need to mention the skill by name every time — if you ask Claude to "write a product launch email" and your brand skill's description includes email campaigns, it loads automatically.

One important nuance about triggering: Claude only consults a skill when the task feels complex enough to warrant it. A short request like "write me a tagline" might get answered from general knowledge without the skill ever loading. Longer, more specific prompts are far more likely to trigger it. If you want to guarantee the skill fires, mention the brand by name or reference something specific from the skill, like a register level.

Pull together every piece of brand collateral you can find — not just the polished final versions, but the real working material. The style guide nobody updated after the rebrand. The support reply that earned a thank-you. The social caption that outperformed everything else that month.

Create a working folder with this structure:

For every asset you collect, write three notes: what to keep (specific elements that sound like the brand at its best), what to kill (elements that feel generic, forced, or dated), and what to learn (the rule Claude should extract). Then sort the keepers into four piles — brand identity, voice, visuals, and channel formats. These become your core skill files.

File: brand-foundation.md

This file tells Claude what kind of company it's writing for before it writes a single word. It needs to be short — under 800 words, readable in three minutes. Seven sections, nothing more.

Brand summary — one paragraph explaining what the company does, who it helps, and why anyone should care. Write it like you're explaining the company to a smart person at dinner, not filling in a template. Test: if you could swap in three competitor names and the paragraph still works, it's too generic.

Mission — one sentence. The thing the company is actually trying to do.

Audience — written as a character brief, not a media buy. Claude doesn't need "women 25–44 with disposable income." It needs emotional and situational context. Who are these people? What are they doing when they encounter the brand? What are they tired of?

Positioning — where the brand sits in the market in real terms. What's above, below, and next to it. What it deliberately isn't.

Personality traits — four or five traits, each with three calibration points. This is the single most important element in the entire skill. For each trait, write: what it means in practice, a line that nails it, a "too far" line that overshoots, and a "too flat" line that undershoots. Claude doesn't learn voice from adjectives. It learns from seeing the edges.

What the brand isn't — anti-patterns, banned-word territory, tonal mistakes that keep recurring. This section is often more useful than the personality traits. It catches the moment a brand drifts from playful into try-hard, from confident into cocky.

Pre-flight check — 4–5 gut-check questions Claude should ask itself before writing anything.

File: voice-and-tone.md

Voice is the brand on a normal day. Tone is the outfit it wears for the room. This file teaches Claude the difference.

Before-and-after pairs — for each personality trait from the foundation file, write 5–6 pairs. The "before" is generic or off-brand. The "after" is the brand at its best. Include a one-line note explaining why the "after" works. These pairs are how Claude calibrates. More contrast means less guessing.

Register spectrum — define a 1-to-5 scale that maps energy level to context. This is what most voice guides miss entirely.

Write 2–3 example lines at each register level — same message, different register. This gives your team a shared vocabulary: "write this at a 3" is faster and more precise than "make it sound more like us but not too much."

Banned words and phrases — a running list of specific words and clichés the brand doesn't use. Not "avoid jargon" — name the jargon.

Sentence-level rules — preferences about average sentence length, fragment usage, rhetorical questions, list style, and punctuation quirks. These micro-rules are where consistency actually lives.

File: visual-guidelines.md

Most visual guidelines fail because they're written for humans interpreting a mood board. Claude needs executable instructions.

Cover four areas: colours (hex codes with permitted usage ratios, not just "use orange for accents"), typography (font families, fallbacks, casing rules, character counts), layout rules (paragraph length in lines, spacing, CTA placement, white space), and image direction (lighting style, composition preferences, what subjects appear and what doesn't).

If your team generates HTML artifacts or landing pages through Claude, include CSS variables that encode the visual system. This makes the guidelines machine-readable, not just human-readable.

File: content-formats.md

A line that works in an Instagram caption can fall apart in a board update. For each content format you regularly produce, document: purpose, structure, which personality traits turn up or down, register level, one good example, and one anti-example.

Start with your three most-used formats. Add the rest as you go. If you end up with more than five formats, split into separate files to keep each one focused and maintainable.

This is the step most people skip, and it makes the biggest difference.

Select 5–8 pieces of published content that represent the brand at its best, spanning at least 2–3 different formats. Then select 3–5 pieces that represent what the brand used to sound like or what it sounds like on a bad day. Store them in a references/ subfolder with clear naming:

Claude doesn't learn voice only from rules. It learns from reading the work alongside the rules. The rules say what to aim for. The corpus shows what hitting the target looks like. Keep the total to 8–12 pieces — beyond that, you're adding noise.

Create a structured evaluation that Claude runs before delivering any output. This is not a vague checklist — it's pass/fail criteria that catch drift before your audience does.

The rubric should cover four areas: voice checks (scroll-stop opening lines, adjective density, sentence length variance, banned word flags), tone checks (channel-appropriate tone, warm-before-playful in problem scenarios), structure checks (format compliance, single CTA per section, paragraph length), and anti-pattern checks (could this belong to a competitor? does it sound like it was written to impress a CMO rather than reach a customer?).

Add explicit instructions telling Claude to review against the rubric before delivering any output, revise anything that fails, and note what changed.

The rest of the SOP — including the full SKILL.md assembly with frontmatter formatting, the complete folder structure ready for upload, the testing protocol with adversarial prompts, and the team onboarding and quarterly maintenance playbook — you can find here, a free gift from me :)

✍️ Consensus - Search engine that pulls evidence-based answers from scientific papers.

🌐Perplexity - AI answer engine with inline citations from real-time web search.

🖱️ Bardeen - Automate repetitive browser tasks without code.

⚙️ Relay.app - Workflow automation with built-in human-in-the-loop approval steps.

🖼️ Ideogram - AI image generator that nails legible text and typography in designs.

😍It was great, all I needed

🙂Good, not great

🙄It sucked