AI Field Manual
FM-MK-03 · Guide

Marketing  ·  Fundamentals

Check copy against your brand voice guide

Five writers, three agencies and a product team that writes its own release notes. Each of them has read the voice guide; none of them sounds quite like the others. Editing everything yourself doesn't scale, and "make it sound more on-brand" isn't feedback anyone can use. This guide checks copy against the guide rule by rule, with a reason for every suggested change.

Status
Published
Reading time
About 11 minutes
You'll need
Your brand voice guide, the draft copy, and an approved AI assistant
Last reviewed
November 2026
Before you start

What AI is for here, and what it isn't

Voice guides describe a way of writing in principles ("plain, warm, confident") and examples. Applying them is a judgment call made thousands of times, one sentence at a time. That's why voice drifts as more people write: each writer's judgment is slightly different, and nobody has time to check every line.

AI is good at the checking. It can hold the whole guide in mind, compare each sentence against it, and point to the specific rule a sentence breaks. That turns "this doesn't sound like us" into feedback a writer can act on: "uses jargon ('leverage'), guide section 2: say 'use'."

What it tends to do badly is restraint. Asked to improve copy, it rewrites everything into its own smooth, slightly generic style, which is a voice of its own and not yours. The method here keeps it to marking departures and suggesting minimal edits. Deciding when the guide should be broken stays with you.

Step 1

Make the guide checkable

Most voice guides are written to inspire, not to check against. Before you use one with AI, make sure it has:

  • Numbered principles, each with a one-line description and a do/don't pair. "Confident: say what the product does, without hedging. Do: 'Reports update every hour.' Don't: 'Reports should hopefully update regularly.'"
  • A word list: words you use, words you avoid, and preferred terms ("sign in", not "log in").
  • Mechanics: spelling variant, capitalization of product names, numbers, punctuation choices.
  • Tone by situation: how the voice flexes for an apology, a legal notice, a celebration.

If your guide lacks some of these, the companion template has the structure. Fill it in once; every check after that gets better.

Step 2

Run the check

Give the AI the context as well as the copy. The same sentence can be right on a product page and wrong in an apology, so the channel, the audience and the situation change what counts as a departure. Paste the guide with its section numbers, so every row can cite one.

The voice check prompt
Check the draft below against the voice guide. Your job is to find
departures, not to rewrite.

VOICE GUIDE: [paste, with numbered sections]
CONTEXT: [channel, audience, situation, e.g., "product page, new
  prospects" or "apology email after an outage"]

FOR EACH DEPARTURE, a table row with:
- the original text (exact)
- the guide section it departs from
- what the problem is, in one line
- a suggested edit, changing as few words as possible
- confidence: CLEAR (the guide is explicit) or JUDGMENT (it depends
  on interpretation)

THEN, separately:
- CLAIMS TO CHECK: any objective claim (numbers, "fastest", "best",
  "guaranteed", comparisons) that needs evidence or legal review.
  These aren't voice issues; list them anyway.

RULES: Do not rewrite sentences that follow the guide. Do not change
facts, product names or legal wording. If the draft is consistent with
the guide, say so.

DRAFT:
[paste]

The "change as few words as possible" instruction keeps the writer's voice in the copy. Their sentence with one word changed is better feedback than your sentence.

Step 3

Review the suggestions

Go through the table. Accept the clear departures, which are usually word-list and mechanics issues, quickly. Spend your time on the "judgment" rows:

  • Is the guide right here? A playful voice that works on a product page may be wrong in an outage notice. The situation-tone part of the guide should cover this; if it doesn't, the guide needs a new line.
  • Is the suggestion better or just different? If the edit doesn't fix a real departure, reject it.
  • Is the writer doing something deliberately? A short, blunt sentence that breaks the "warm" principle may be exactly right for the moment.

Send the claims list to whoever checks substantiation or legal wording. In marketing, objective claims need evidence before they're published; a voice check is a good place to catch them, but not the place to clear them. Clearing a claim means looking at the data behind it, not just confirming that some data exists:

  • The evidence: who was measured, how many, how, and when. Self-reported estimates from a small beta group are weaker evidence than measured results from typical customers.
  • The typical result, not only the best one. If the claim quotes a maximum, what did most people get?
  • The reader's takeaway: what a reasonable customer will believe they'll get. Qualifiers like "up to" don't reliably fix a misleading impression. In an FTC-commissioned study, many consumers expected to achieve the maximum result an "up to" claim states.

Approve wording that matches the takeaway to the evidence.

Step 4

Give feedback the writer can use

Send the writer the table, trimmed to the edits you accepted, with the guide section for each. Writers learn the guide from specific, repeated feedback far faster than from rereading it.

Keep a note of rows you rejected repeatedly. Those point to places where the guide is unclear or out of date, and fixing the guide is better than overriding the check every time.

Step 5

Worked example: the agency landing page

Fictional example. An agency delivers landing page copy for a new scheduling product. The brand's guide has four principles: plain, warm, confident and brief. It uses US spelling and "sign up" rather than "register".

The check returns eight rows. Six are clear and quick:

Original Section Suggested edit
"Leverage intelligent scheduling" 1 · Plain (avoid jargon) "Use smart scheduling"
"Register today" Word list "Sign up today"
"It should help you save time" 3 · Confident (no hedging) "It saves you time"

…and three more like them. Two are judgment calls. The headline, "Your calendar, finally under control", breaks "warm" (it implies the reader's calendar is out of control). The marketer keeps it: it tested well, and a little edge in a headline is fine. The other, a long paragraph flagged under "brief", gets cut by half.

The claims list flags one line: "Teams save 5 hours a week." It goes to the legal reviewer with the data behind it: a survey of 40 beta users who estimated their own time saved. Looking at the responses, the median estimate was 1.5 hours a week; 5 hours was the highest figure, reported by three users.

"Up to 5 hours" would be literally true, but a reader would reasonably expect something close to 5 hours, which most beta users didn't report. The approved line describes the typical result and the kind of evidence: "Beta users estimated saving a median of 1.5 hours a week", with a note on the page: "Self-reported by 40 beta users, [month and year]." The team also starts measuring time saved in the product, so a stronger claim can be supported later. The agency gets the table and the reasons the same afternoon.

Step 6

Record the review

Voice check record
Piece: [title], [channel]     Writer: [name]     Reviewer: [name]
Guide version: [v]     Context: [audience, situation]

ACCEPTED EDITS
Original | Section | Edit | Reason

JUDGMENT CALLS
[Original] — [section] — kept / changed — because [ ]

CLAIMS SENT FOR REVIEW
[claim] — evidence [ ] — reviewer [ ] — outcome [ ]

GUIDE NOTES
[section that needs clarifying, and why]

What stays with you: when breaking the guide is the right call.

Esc

Saved in this browser

Reading list

Contents

Reference

Keyboard shortcuts

/ or Ctrl K
Search the manual
↑ ↓
Move through search results and index entries
Enter
Open the selected result
← →
Switch discipline tabs; previous / next entry in a workflow card
Esc
Close any dialog
?
Show this list