How to Build a Brand Voice Guide That Works
Stop describing good. Start forbidding bad.
You've coached Claude, corrected it, fed it your writing — and it still comes back grammatical, on-topic, and unmistakably AI. "Make it sound human" is a target a model can't aim at, so it defaults to the average of every piece of marketing copy ever written. Slop is that average.
The fix is to convert taste into rules a search box could check. Slop has fingerprints — the staged scene-setting opening, the rhetorical-question chain, the tidy three-part recap, the word delve. Models are far better at never do this exact thing than at be more authentic.
Copy your tics, get a caricature. Copy your judgment, get your voice.
Generated content is held to a stricter bar than your own keyboard. A pattern doesn't earn a pass just because you write that way sometimes — and stripping slop must never strip the connection to a real reader. Clean but cold is its own failure.
Four steps to a working voice file
A Claude Project + two files + a prompt. No connectors, no plugins required to start.
Pick the tool
Create a Project in Claude Cowork. Name it "Brand Voice" or "Content." The Project is where your rules live so they apply on every generation without re-briefing.
Set up the workspace
One folder, two files: voice-constraints.md (the rules) and good-examples.md (2–3 pieces of your own writing you'd publish). Connect the folder to the Project.
Prompt Claude
Ask for content and name the constraint file explicitly the first few times. Treat every banned word and named failure pattern as a fail — not a preference.
Review what comes back
The file does most of the filtering. You make the final call: banned-word sweep, opening read-aloud, and a check that the rules didn't strip the connection.
What voice-constraints.md actually contains
Three sections. Hard rules, not a personality profile.
Banned words & phrases
The words that announce a machine wrote it. Starter list ships with the template. Add the ones specific to your industry that make you wince.
Named failure patterns
Structural slop moves, each with an example: the stacked-detail opening, the fake contrast pair, the rhetorical-question chain, the tidy recap.
What you actually sound like
Short. Framed as judgment, not quirks. Not "I use dashes" — but "I'm direct, I correct myself when I overstate, I use real numbers, I don't hype."
Stuck writing the file from scratch?
Don't start from blank. Copy the bundled template at voice-constraint-guide-TEMPLATE.md, then attach 2–3 pieces of your own writing and ask Claude to describe the judgment in how you argue while flagging any banned-word patterns your own writing leans on. Add those to the ban list — do not write exceptions that let them through.
The point of reading your writing is to capture how you think, not to whitelist your habits. Fifteen minutes instead of a blank-page standoff.
Words and phrases that announce a machine
A starter set — add your own industry offenders. Do not use unless directly quoting a source.
- In today's fast-paced world
- It's important to note
- At its core
- X isn't just about Y, it's about Z
- That being said
- To summarize
- Key takeaways
What you actually type in the Project
Use these verbatim. The pattern: name the constraint file, treat it as hard rules, and ask for a self-grade on anything important.
Generate content
"Write a 60-second video script promoting [topic]. Apply voice-constraints.md as hard rules — every banned word and failure pattern is a fail, not a preference. Match the judgment in good-examples.md, not the surface style."
Fix the file, not the draft
"This opening is a staged scene — that's the failure pattern in section 2 of my constraints. Rewrite it, and if the constraint file doesn't already forbid this clearly enough, tell me how to tighten the rule."
Self-grade before delivery
"Before you give me the final version, run it against voice-constraints.md line by line and flag anything that violates a rule. Show me the check, then the clean draft."
Three things to check yourself
The constraint file does most of the filtering. You make the final call on the parts a search box can't see.
Run the banned-word sweep
Search the draft for your top offenders. Every catch makes the file sharper — either tighten the rule or add the word to the list.
Read the opening out loud
Slop hides in the opening: staged scenes, "every business owner knows" claims, atmosphere before any point. If the first two sentences could open anyone's article, they fail.
Confirm the rules kept the emotion
Constraint files can be applied too hard — leaving copy that's clean but cold. The goal is a real person with your judgment, not a sanitized press release.
The rule before anything publishes: you read every piece that goes out under your name. The constraint file gets you from slop to a strong draft in one pass — it doesn't replace the last read. That final human pass is where you confirm the connection is real and the voice is yours.
Compounding moves once the file works
Small additions that pay back the next time you ship anything.
Reuse everywhere
One file, every surface — scripts, emails, social, landing pages.
Shared Drive folder
Put the file in Google Drive so everyone's drafts come out consistent.
Version the file
v1, v2, v3. See how your definition of slop sharpened over months.
Run it as a plugin
The brand-voice-noai plugin packages a skill, a /slop-check linter, an enforcement hook, and a reviewer agent.
Want the full pattern catalog?
This page is the procedure. The companion piece names every pattern your constraint file needs to forbid — the eight structural fingerprints, the journey model the writing has to move through, the review gate before anything ships.
"When output is consistently wrong, the durable fix is almost never a cleverer prompt. It's better stored context."
Three principles that compound
The recipe is one application. The principles apply to every constraint you'd put on a model.
Context beats prompting
Coaching evaporates when a conversation ends. A file stored in a Project is context Claude references on every generation. When output is consistently wrong, the durable fix is stored context — not a cleverer prompt.
Forbidding beats describing
"Be more human" is a target a model can't hit. "Never write delve, never open with a staged scene, never end with a tidy recap" are pass/fail tests it can apply every time. Convert taste into rules a search box could check.
Imitate judgment, not tics
Feed samples and you get a caricature — dashes, short lines, quirks dialed up. The move that works is the opposite: tell Claude to copy how you think — your directness, your skepticism, your willingness to correct yourself — and explicitly forbid it from copying your surface habits.
