The CLAUDE.md File That Locks Your Brand Voice (Full Anatomy + Copyable Template)

Claude Code for Founders · Episode 2: The File. 42 seconds. The actual CLAUDE.md that holds the brand voice. Watch on YouTube.

The drift problem. Your brand voice degrades one quarter at a time.

Here is how brand voice actually dies. In Q1, you brief your contractor: "We sound direct, no-nonsense, confident." They write from that. In Q2, a different person picks it up and works from memory: "direct, professional, polished." In Q3, the agency rounds it off to "professional, premium, refined." Nobody made a wrong decision. Each step was a reasonable paraphrase of the last one. But four quarters later, the brand sounds like a different company.

This is drift, and it is structural. Every time the brand voice gets re-explained from memory, it shifts slightly toward the mean. The mean is corporate. So the voice flattens, post by post, until what ships is indistinguishable from every other brand using the same AI tools with the same defaults.

We wrote up the broader version of this in Brand as Code, Episode 1 of this series. The short version: stateless AI plus no brand spec equals predictable slop. This post is about the fix, in detail. The file.

What a CLAUDE.md brand-voice file actually is.

A CLAUDE.md is a markdown file that lives at the root of a project. Claude Code reads it automatically before doing anything in that directory. It is the single most leveraged file in a content operation, because it is the one document every tool downstream agrees on.

For brand voice, the file has four sections. That is the whole spec. Not thirty pages. Four sections.

1 · Voice attributes

The three to five adjectives that describe how the brand actually writes. The discipline here is specificity. "Premium" and "innovative" are not voice attributes, they are aspirations that every brand claims. "Direct, confident, no-nonsense" is a voice. It tells the model to lead with claims, cut hedges, and skip the throat-clearing. If your attributes could describe your competitor, they are too vague.

2 · Tone register

The default posture the writing takes. Executive, not academic. First-person, not third. Active voice, not passive. Concrete, not abstract. These are binary choices that should be the same in every output. The tone section is where you settle the arguments once so they never get re-litigated per piece of content.

3 · Forbidden words

The list of phrases that get rejected at the gate. This is the most powerful section because it is the most enforceable. The model knows these are off-limits, so it routes around them. Synergy. Robust. Transformative. Cutting-edge. Seamless. Each brand keeps its own list, but every list does the same job: it blocks the words that signal AI-generated filler. The forbidden list is your firewall against reverting to the mean.

4 · Required structures

The shapes the writing must take. Concrete nouns over abstract claims. Falsifiable statements over hedges. Lead with the claim, never with a windup. One thesis per piece. Where the forbidden section blocks the bad, the required section enforces the good. Together they bracket the output into the brand's actual voice.

Why the file beats the brand guide.

A traditional brand guide is a document humans are supposed to read and then remember. That fails for three reasons. People do not re-read it. Contractors never receive it. And the AI tools that now produce most of the content never see it at all. The guide is a reference that depends on human memory, and human memory drifts.

A CLAUDE.md file is different in kind. It is not a reference, it is an input. The AI reads it programmatically, every session, before producing anything. There is no memory step to fail. The spec is loaded into context before the prompt fires, which means the brand voice is enforced at the moment of production rather than checked after the fact. That is the entire difference between a guide and a file: one hopes to be remembered, the other is read by construction.

How every tool reads the same file.

The leverage compounds because the file is not read by one tool. It is read by all of them. Claude Code reads it when drafting copy. The voice linter reads it to fail a commit that ships a forbidden word. The video pipeline reads the brand tokens it references. The voiceover generator reads the tone. Every surface that produces brand output pulls from the same spec, which is why the output stays consistent across channels. The LinkedIn post, the Reel caption, and the email subject all sound like the same brand because they are all reading the same file underneath.

This is the part that does not fit in a 42-second video, so here it is in full: consistency across channels is not a discipline you maintain, it is a property that falls out of having one source of truth. You do not keep the channels aligned by being careful. They stay aligned because they all read the same file.

A copyable CLAUDE.md starter template.

Here is a minimal brand-voice block you can drop into a CLAUDE.md at the root of your repo and edit. Replace the values with your brand. The shape is what matters.

# Brand Voice

voice:
  attributes: [direct, confident, no-nonsense]
  tone: executive, not academic

forbidden:
  - synergy
  - robust
  - transformative
  - cutting-edge
  - seamless

required:
  - concrete nouns over abstract claims
  - falsifiable statements over hedges
  - lead with the claim, never a windup
  - one thesis per piece

structure:
  - no hedge openers ("in today's landscape")
  - no stacked adjectives
  - no hollow conclusions

That is roughly twenty lines and it will already change every draft your AI produces. The full version we run per brand adds channel-specific notes, example rewrites, and links to the brand color tokens the video pipeline uses, but the block above is the load-bearing core. Start there.

How to start.

1. Run the free voice scan. The scanner on the homepage reads your existing site and outputs your voice profile in about ninety seconds. That profile is the first draft of your file.
2. Drop the template above into a CLAUDE.md at the root of wherever you run Claude Code. Edit the four sections to match your brand.
3. Audit your live content against it. Point whystrohm-audit at your site and see where the drift already is.
4. Re-run weekly. The file is versioned in git, so it evolves as the brand learns about itself. The drift stops because the spec stopped moving.

This is Episode 2 of Claude Code for Founders. Episode 1, Brand as Code, covers why the file matters at all. Episode 3 covers how Claude Code remembers you between conversations, so the file is not the only thing carrying context.