The Client's Brand Voice in One File: Stop Rewriting Every Prompt

How to keep each client's brand voice in one file that every AI prompt reads: what goes in it, how to fill it, prompt patterns, an example and review rules.

Written by the Miriad team with AI assistance. Figures link to their sources; the example is illustrative, not a real case. Updated 2026-10-21.

TL;DR: Keep each client's brand voice in one dated file: tone, words they use, words they never use, approved claims, three approved samples and a log of past corrections. Every prompt for that client attaches the file instead of retyping instructions. The output stays closer to the client's voice, a new team member has something to read on day one, and corrections stop repeating.

An account manager opens a new chat to draft a client's newsletter and types the same paragraph she typed last week: "Friendly but professional, no exclamation marks, never say 'cheap', the founder is called Kati, mention the free parking." Last week she forgot the parking. The week before, a colleague wrote "affordable", which the client has now crossed out three times. The instructions exist, but they live in people's heads and old chats. This article belongs to our guide to marketing agency AI and staying different when everyone uses it.

What is a client voice file, and why keep it in one file?

A client voice file is a single document that holds everything a writer, human or AI, needs to sound like that client: tone, vocabulary, banned words, approved claims, samples and past corrections. One file means one place to update, one attachment for every prompt, and no version drift between team members or tools.

Without it, voice instructions are scattered across the brand guidelines PDF, the kickoff call, an email from the client's marketing manager and the memory of whoever wrote last month's posts. Each prompt picks up a different subset, and the client notices. With it, the account manager attaches the same file to every prompt, and the client's corrections go into the file instead of into one person's notes.

The file also makes AI output easier to review. When a draft deviates, the reviewer can point to a line in the file rather than argue about taste.

What goes into a client's brand voice file?

A voice file needs seven sections: who the client is in two lines, the audience, tone rules with examples, words to use, words never to use, approved claims with sources, and three approved samples. Add a feedback log at the end. Keep it under two pages, so it fits in any prompt and gets read.

A template you can copy:

(Client name) — voice file          version: (date) · owner: (account manager)

WHO THEY ARE        2 lines
AUDIENCE            who reads, what they already know, what they worry about
TONE RULES          each with a "like this / not like this" pair
WORDS TO USE        the client's own terms, product names spelled exactly
WORDS NEVER TO USE  and what to say instead
APPROVED CLAIMS     each with source and approval date
SAMPLES             three short pieces the client approved, with dates
FEEDBACK LOG        date · what the client changed · the rule it implies

The "words never to use" section does more work than any tone description. "Friendly but professional" means something different to every writer; "never 'cheap', say 'fairly priced'" means the same thing to everyone. The approved claims section stops AI from inventing new ones: if a claim is not in the list, it does not go into the draft without asking.

If the client has brand guidelines, link them at the top, but do not copy the whole document. The voice file is the working extract.

How do you fill the file from what the client already has?

Fill it from evidence, not from a questionnaire alone. Collect the client's brand guidelines, ten pieces of content they approved, their customers' reviews, and the transcript of the kickoff call. Ask AI to draft the file from those, then go through it with the client in a short call and date the approved version.

Steps for client onboarding:

  1. Collect: brand guidelines, ten approved posts or pages, twenty customer reviews, the kickoff call transcript, any past corrections by email.
  2. Extract: ask an AI tool to list recurring words, sentence length, words never used, and claims made, quoting the source for each.
  3. Draft: fill the template from that list. Mark anything you inferred rather than found.
  4. Approve: walk the client through the file in fifteen minutes. Their reactions to the "words never to use" list are the most useful part.
  5. Date and store: save it in the client's folder with the version date.

Keep it current with the feedback log. Every time the client changes a draft, write down what changed and the rule behind it. Once a quarter, move repeated rules from the log into the main sections and date the new version.

How do prompts use the voice file?

Every prompt for a client attaches the voice file and refers to it by section, instead of restating the voice in the prompt. The prompt then only describes the task. Ask the model to flag anything in its draft that conflicts with the file, so the reviewer sees the risky lines first.

A prompt pattern for your prompt library:

Attached: (client) voice file, version (date).
Task: write (deliverable) about (topic) for (channel), max (length).
Use only claims from "Approved claims". If you need a claim that is
not there, write [CLAIM NEEDED] instead of inventing one.
After the draft, list every line that might conflict with
"Tone rules" or "Words never to use", and why.

The same pattern works for every deliverable, so the prompt library stays short: one prompt per deliverable type, and one voice file per client. In most AI tools you can keep the file inside a separate project per client, so it is attached automatically. Tool choice by workflow step is covered in best AI tools for a marketing agency.

Example: one voice file per client for twelve local businesses

A worked example shows the file at agency scale. Below, a small content team writes blog posts for twelve local-business clients, each with its own voice file and its own review step. The figures are illustrative; the routine is the part to copy, especially the rule that one article never serves two clients.

Example (illustrative, not a real case): A content team of three writes four blog posts a month for each of twelve local-business clients: a dentist, two restaurants, a plumber, a bookshop and seven others, 48 posts a month in total. Each client has a voice file of about a page and a half. Every draft starts from a prompt that attaches the file. Before publishing, an editor fact-checks the body, the title, the meta description and the image alt text against the client's facts file. The same article is never adapted for two client sites, even when two clients would like the same topic; each gets its own angle from its own voice file and facts. After three months, the feedback logs hold between four and nineteen entries per client, and the editor moves the repeated ones into the main sections.

Which rules make the review step non-negotiable?

The review step is required by more than quality. Search engine policies treat mass-produced pages without added value as spam, the EU AI Act ties some disclosure exemptions to human editorial responsibility, and copyright protects what people add to AI output. A voice file makes drafts better; only a named reviewer makes them publishable.

Google's guidance says that generating many pages without adding value may violate the scaled content abuse policy, and that titles, meta descriptions, structured data and alt text should be fact-checked before publishing (Google, 2026). Its spam policies list content from white-label services and freelancers as third-party content under the site reputation policy (Google, 2026). If you sell white-label content that other agencies publish on their clients' sites, that applies to your output too.

Under Article 50(4) of the EU AI Act, applicable from 2 August 2026, AI-generated text published to inform the public on matters of public interest must be disclosed, unless a person holds editorial responsibility after human review (AI Act Art. 50, 2024). And the US Copyright Office's 2025 report says prompts alone do not make output protectable, while human selection, arrangement and modification can be (US Copyright Office, 2025). For the client, the edits a reviewer makes are part of what they buy. The same voice file approach works for your own agency's positioning; slogans for a marketing agency shows where a slogan fits in it.

If you want a guided route to a written brand system, voice rules included, our Brand Identity Prompt Pack ends in a one-page guideline plus a short brand brief for future AI chats. Get notified when it launches. For the wider picture, return to the marketing agency AI guide.

FAQ

How long should a client's brand voice file be?

Short enough that people read it and every prompt can include it, usually one to two pages. The examples and the "words never to use" list matter more than long descriptions of tone. If the file grows past two pages, move background material into a separate document and link it, keeping the voice file as the working extract.

What is the difference between brand guidelines and a voice file?

Brand guidelines describe the whole brand: logo, colours, typography and usually some general tone advice. A voice file is the working extract for writers and AI: concrete words to use and avoid, approved claims, samples and a log of the client's corrections. It links to the guidelines but is written to be attached to every prompt.

Who should own each client's voice file?

The account manager for that client owns it: they update it after feedback and date each version. The client approves the main sections. Writers and editors can propose changes through the feedback log, but only the owner edits the main file, so there is always one current version that every prompt reads.

Can AI build the voice file for us?

AI can draft it from the client's approved content, reviews and call transcripts, quoting a source for each rule. It cannot decide what the client wants to sound like. Treat the AI draft as a starting point, mark anything inferred, and go through it with the client before using it in prompts.

Sources

Related articles

Products from Miriad

BrandBrand Identity Prompt Pack

Build a brand system with Claude or ChatGPT, one decision at a time

29 USD

Get notified

See all prompt packs