Easy~25 min setup

A Brand-Voice Review Before Anything Ships

Run every draft through a consistent voice-and-claims check before it publishes — flagging off-brand tone, unsupported claims, and missing disclaimers, by severity, with fixes.

Consistency isn't something you remember to do. It's a gate every draft passes through.

Voice drifts when each person edits by feel and busy weeks lower the bar. Store your voice and claim rules, and every draft gets the same review before it ships — off-brand tone, unsupported claims, missing disclaimers, all flagged by severity with the fix attached.

Voice doesn't break in one bad post. It erodes one rushed, slightly-off piece at a time, until the brand sounds like five different people.

A voice guide that lives in a doc nobody opens isn't a standard. A voice guide that gates every draft is.

The review catches what a checklist can; your judgment catches what it can't. The gate raises the floor — it doesn't replace the final human read on anything that goes out under your name.

At a glance
ComplexityEasy
Tools neededClaude Pro or above, Claude Desktop → Cowork mode, on macOS or Windows, Claude Projects
Time to build~25 min first time · ~2 min per draft after that
Best forMarketing lead or founder who owns the brand voice — Any business publishing content under one brand

What this solves

A voice guide nobody opens before publishing, so tone drifts and risky claims slip through one rushed draft at a time.

The problem

You have a brand voice guide. It’s a thoughtful document, and almost nobody opens it before they publish. So voice drifts — a social post that’s a little too hype, a blog intro with a cliché opener, a claim like “the best in the industry” that legal would wince at, a promo missing the disclaimer it needs. No single piece is a disaster, but across dozens of drafts from a few people in busy weeks, the brand starts sounding like several different companies, and an unsupported claim eventually creates a real problem. The guide exists; the gate doesn’t.

The fix is turning the guide into a review that actually runs on every draft, fast enough that nobody skips it.

Ingredients

  • Claude SubscriptionsCowork and Projects aren't available on the free plan
  • Platforms & ModesCowork runs in the desktop app only — not on web or mobile. This recipe reviews drafts against your standards and flags issues — file-grounded work that runs in Cowork
  • Claude ProjectsYour voice, your claim rules, and your required disclaimers are constant. A Project holds them so every draft gets the same review against the same standard
Built on these guides

How it works

1
A Claude Project

Every draft needs the same check, and your standards don’t change. A Project holds your voice and claim rules so the review is consistent across people and pieces — the guide becomes a gate instead of a document.

  1. Open Claude Desktop and click Cowork in the mode selector across the top (Chat · Cowork · Code).
  2. In the left panel, find Projects and click the + button.
  3. Choose Start from scratch. Name the project “Brand Review” and let Claude set up its folder.
  4. You’ll know it worked when the project appears in the left panel with its own folder and an instructions field.
~25 minutes the first time. ~2 minutes per draft after that
2
Set Up Your Workspace

Your project has a folder on your computer — that’s where the two working files live.

  1. Ask Claude, right in the project: “Create two files in this project’s folder: voice.md and claim-rules.md. Leave them empty — I’ll fill them in.” (Or create them yourself in any text editor and save them into the project folder.)
  2. Fill in both files using the descriptions below.
  3. Confirm Claude can see them: ask “List the files you can see in this project.” Both filenames should come back. If they don’t, see If It Doesn’t Work.

voice.md

Your voice as rules, not adjectives — including the patterns to forbid. “Be authentic” can’t be checked; “never open with a staged scene, never use these banned words, never stack three rhetorical questions” can. The companion brand-voice work goes deep on this.

Example: “Direct, concrete, no hype. Banned: ‘game-changing,’ ‘seamless,’ ‘in today’s fast-paced world,’ ‘delve.’ No fake-question openers. No three-part tidy summaries. Specific over sweeping.”

claim-rules.md

What you can and can’t say without backup, and the disclaimers certain content requires — so the review flags compliance issues, not just tone.

Example: “No superlatives we can’t prove (‘best,’ ‘leading’) without a citation. No specific result claims without a source. Anything mentioning ROI needs the standard disclaimer. No competitor names in a negative claim.”

3
Prompt Claude

Open your Project in Claude Cowork. Give Claude the specifics for this run, ask for the main output, then follow up for any additional pieces you want. The exact wording for each prompt — with what it’s asking for and why — is in What you actually type below.

4
Review What Comes Back

Check three things before you trust the review:

  1. Severity is sensible. A review that flags everything as a blocker is as useless as one that flags nothing. Confirm the blocker/should-fix/nice-to-have split matches real risk — an unsupported claim is a blocker; a slightly long sentence isn’t.
  2. Claims and disclaimers were actually checked. The voice part is easy; the valuable part is catching “we’re the leading provider” with no proof, or a promo missing its disclaimer. Make sure the review covered claims, not just tone.
  3. Fixes are specific, not scolding. “This is too salesy” doesn’t help. “Replace ‘revolutionary platform’ with what it actually does for the user” does. If the feedback is vague, ask for the before/after rewrite.

Before it publishes: the gate raises the floor; it doesn’t replace your read. You’re still responsible for anything that goes out under your name, and claim or legal questions should go to whoever owns them. The review catches the patterns; you catch the things judgment sees.

What you actually type

Name your files explicitly the first few runs, and ask Claude to show its work on anything that matters.

Prompt AGenerateCopy this
Ask for the review
Review this draft against `voice.md` and `claim-rules.md` [paste]. Flag every issue by severity — blocker, should-fix, nice-to-have. For each: quote the problem, say which rule it breaks, and give a specific before/after fix. Call out any unsupported claim or missing disclaimer separately.
Prompt BFollow upCopy this
Follow up
Apply all the blocker and should-fix changes and give me the clean version, leaving the nice-to-haves for me to judge.
What you get back

A review of your draft with every issue sorted into blocker, should-fix, or nice-to-have — each one quoting the problem text, naming the rule it breaks, and giving a specific before/after fix. Unsupported claims and missing disclaimers are called out separately from tone issues. The follow-up returns a clean version with the blockers and should-fixes applied, leaving nice-to-haves for your judgment. If everything comes back a blocker, or the claims check is missing entirely, it missed — see If It Doesn't Work.

What this does not do
  • It isn't legal review. It catches a claim your rules forbid; whether a disclaimer actually satisfies a regulator is a question for whoever owns compliance.
  • It only enforces rules you've written down. 'Be authentic' gates nothing — when a pattern slips through, the fix is adding it to `voice.md`, not blaming the review.
  • It doesn't replace the final human read. The gate raises the floor; you're still the last set of eyes on anything that ships under your name.

If it doesn’t work

  • No Cowork tab in Claude Desktop — update the app to the latest version and confirm you’re on a paid plan; Cowork isn’t on the free tier. On Windows, Cowork also needs the Virtual Machine Platform feature enabled — if the tab still won’t appear, that’s the fix.
  • Claude can’t see voice.md or claim-rules.md — the files aren’t in the project’s folder, or they’re in a different folder than the one the project owns. Open the project, check which folder it points to, and move the files there. Then re-run “list the files you can see.”
  • Everything comes back a blocker — your files don’t tell Claude what severity means, so it escalates to be safe. Add the line to claim-rules.md: unsupported claims and missing disclaimers are blockers; tone and banned-word issues are should-fix; style preferences are nice-to-have. Severity is a rule like any other — write it down.
  • The review covers tone but skips claims — the prompt named one file and not the other, or claim-rules.md is too thin to check against. Name both files in every review, and if claims still get missed, run them as their own pass: “Now check only the claims and disclaimers against claim-rules.md.”

Extra credit

Small additions that pay back the next time you run it.

  • Run it as a plugin — the brand-voice approach can be packaged so a /slop-check-style command lints any draft on demand, and an enforcement hook checks content automatically. See the Plugins guide.
  • Shared standard — keep voice.md in a shared Drive folder so every writer is reviewed against the same rules. See the Connectors guide.
  • Sharpen the rules — every time something off-brand slips through, add the pattern to the file. The gate gets stricter as you learn what you keep missing.

“Brand voice isn't lost in one bad post. It erodes one rushed draft at a time.”

What this teaches you about Claude Cowork

The recipe is one application. The principles apply to everything you’d hand to Claude.

A gate beats a guide. A voice document only works if it’s applied, and applied consistently. Turning it into a review that runs on every draft — fast enough that nobody skips it — is what actually keeps voice consistent, instead of hoping people remember the doc.

Rules a machine can check are rules that hold. “Be authentic” can’t gate anything; “never use these words, never open this way” can. Writing voice as forbidden patterns is what makes the review reliable, draft after draft.

Severity makes feedback usable. A flat list of complaints gets ignored. Sorting issues into blockers, should-fixes, and nice-to-haves tells a busy writer exactly what must change before publishing and what’s optional — which is what gets the review acted on, not argued with.

Who this is for

Marketing lead or founder who owns the brand voice in Any business publishing content under one brand (5–100 employees).

The pain: Voice drift and unsupported claims slipping through in busy weeks

The outcome: Every draft passes the same voice-and-claims gate before it ships

Published June 16, 2026 · 0 views

Run a Brand-Voice Review Before You Publish | Eikon Digital