---
title: "Make AI Write in Your Voice: PM Setup"
description: "How to make AI write in your voice at work: real writing samples instead of adjectives, hard rules it enforces, banned phrases, and a portable voice file."
canonical_url: "https://builderscamp.com/guides/tools/make-ai-write-in-your-voice"
date_published: "2026-09-27"
date_modified: "2026-09-27"
author: "Andre Albuquerque"
publisher: "Builders Camp"
guide_class: "tools"
---

# How to make AI write in your voice: samples, hard rules and banned phrases

**TL;DR:** To make AI write in your voice, give it three to five real samples of your writing instead of adjectives, a short list of hard rules it must follow, and a banned-phrase list with a replacement for each. Keep that voice file separate from the quarterly context about your job, so it stays portable, and test it by checking whether a brand-new chat drafts in your voice without being told.

AI writes in a generic voice because generic is the average of everything it has read. Anthropic's [prompting best practices](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices) recommend including 3 to 5 examples for best results, and for voice that advice is the whole method: show the model how you write instead of telling it.

The second half of the method is older than AI. George Orwell's essay Politics and the English Language, published in Horizon in April 1946 according to [the Orwell Foundation's text](https://www.orwellfoundation.com/the-orwell-foundation/orwell/essays-and-other-works/politics-and-the-english-language/), ends with six rules for plain writing, and the first is a banned-phrase rule. Samples show the model your voice; bans and rules keep it from drifting back to the average.

## Why do adjectives fail as style instructions?

"Write in a concise, direct, friendly tone" is the most common voice instruction and the least useful. Every one of those words describes a range, and the model picks the middle of it. Your version of direct might mean leading with the decision and never apologising for bad news. Someone else's means short sentences with exclamation marks.

Samples carry what adjectives cannot: how long your sentences are, how you open a message, whether you use bullets, how you deliver a slip, what you never bother to say. Anthropic's guidance makes a related point about format: "The formatting style used in your prompt may influence Claude's response style." A prompt full of your own writing pulls the output toward your writing.

## Which writing samples should you use?

Pick three to five pieces of the kind of writing you most want help with, written under normal conditions and good enough that you would send them again. For most PMs that means work writing, not blog posts:

| Sample | What it teaches the model |
|---|---|
| A weekly update to your stakeholders | Your default structure, length and level of detail |
| An update that delivered bad news | How you handle a slip, a risk or a changed plan |
| A decision memo or recommendation | How you argue, and where you put the ask |
| A short reply to a senior person | Your register when the stakes are higher |

Vary them on purpose. Five good-news updates teach the model one mood. Mix good and bad news, short and long, peer and exec, so it learns what stays constant in your writing and what changes with the situation. Then ask the model to describe the patterns it sees in the samples, and edit that description; you will usually find it names habits you did not know you had.

## What hard rules should the model enforce?

Samples teach tendencies. Rules teach the things you never compromise on. Keep the list short, because a long list of rules gets applied unevenly, and write each one as something checkable:

1. Every recommendation states the decision in the first sentence.
2. Every number carries its source and the date it was pulled.
3. Status updates never exceed what fits on one phone screen.

Rules like these do two jobs. They shape the draft, and they give the model something to check its own draft against. Ask it to confirm each rule before it hands you the text, so rule breaks get caught before you start reading.

## How do banned phrases help, and how do you write them?

A banned-phrase list removes the tics that make a draft sound machine-written or corporate: "I hope this finds you well," "exciting news," "circle back," "moving forward," "just wanted to flag." George Orwell's rule from [Politics and the English Language](https://www.orwellfoundation.com/the-orwell-foundation/orwell/essays-and-other-works/politics-and-the-english-language/) still applies: "Never use a metaphor, simile or other figure of speech which you are used to seeing in print." Stock phrases are how a model fills space when it has nothing specific to say, and banning them forces specifics.

Pair each ban with what to do instead. Anthropic's guidance on steering output is explicit: "Tell Claude what to do instead of what not to do." A ban on "just wanted to flag" works better as "open with the risk itself." A ban on "exciting news" works better as "open with what changed and who it affects." The replacement is the part the model actually learns.

## Why keep your voice separate from your job context?

Your voice survives a job change. Your current product area, your metrics, your stakeholders and this quarter's priorities do not. If both live in one file, you will rewrite the whole thing every quarter and eventually stop maintaining it, and the voice degrades with the stale context around it.

Use a simple test for every line: would it still be true if you moved to a different company tomorrow? If yes, it belongs in the voice file. If no, it belongs in a separate context file you refresh each quarter. That split is also what lets one voice file work across every assistant and agent you use, which is the foundation of a [personal AI operating system](https://builderscamp.com/guides/tools/build-a-personal-ai-operating-system).

## How do you test whether the model has your voice?

Open a brand-new chat, load the voice file the way you normally would (custom instructions, a project, an agent's instruction file), and give it a real task without mentioning style: "Draft this week's update on the onboarding redesign from these notes." Read the result as if a colleague sent it.

If you would send it with only factual edits, the setup works. If you keep making the same kind of change, that change is a missing rule or a missing banned phrase: add it and test again. Two or three rounds usually settle it. For the updates themselves, [AI for stakeholder updates](https://builderscamp.com/guides/tools/ai-for-stakeholder-updates) covers structure, and the glossary entry on [few-shot prompting](https://builderscamp.com/guides/glossary/few-shot-prompting) explains why examples steer a model so strongly.

## What is the honest limit of an AI voice setup?

A voice file makes drafts sound like you. It does not make them say what you would say. The model can match your sentence rhythm while getting the decision, the emphasis or the political read of a situation wrong, and a draft in your voice is more dangerous than a generic one, because it is easier to send without reading. Treat every draft as a first draft, and keep the judgement calls for yourself.

## Where does Builders Camp cover this?

Building your AI Operating System is a 2 week Builders Camp bootcamp with 3 live sessions, directed by Inês Lourenço, and part of both the AI Agentic Builders Expert Track and the Product Leadership Track. Its published topics include context layer design, reusable PM skills, agent orchestration, and automation and quality control. See the [Building your AI Operating System bootcamp](https://builderscamp.com/bootcamps/building-your-ai-operating-system) for dates and format.

## Frequently asked questions

### Why does describing my style with adjectives not work?

Because adjectives like 'concise', 'friendly' or 'direct' mean different things to different people, and the model fills them with its own average. A few real samples show sentence length, openings, how you handle bad news and what you leave out, none of which an adjective carries.

### How many writing samples should I give the model?

Anthropic's prompting guidance recommends three to five examples for best results, and that range works well for voice too. Pick samples of the same kind of writing you want the model to produce, and make them varied: one good-news update, one bad-news update, one decision memo.

### Should I use my best writing or my typical writing as samples?

Your best typical writing: pieces you would be happy to send again, written under normal time pressure. Heavily polished one-off pieces teach the model a voice you cannot sustain, and the gap shows the first time you edit its draft.

### Where should the voice file live?

In a plain text or markdown file you own, outside any single tool. Paste it into custom instructions, attach it to a project, or point an agent at it. If you change assistant or employer, the file moves with you.

### What is the difference between a voice file and a system prompt?

A system prompt is where instructions get loaded; a voice file is one of the things you load there. Keeping the voice in its own file means you can reuse it across assistants, projects and agents without rewriting it inside each one.

### How do I know the setup is working?

Open a brand-new chat, give it a real task without mentioning style, and read the draft. If you would send it with only factual edits, the voice file is working. If you keep rewriting the same kind of sentence, turn that rewrite into a new rule or a new banned phrase.

## Sources

- [Anthropic: Prompting best practices (Claude Platform Docs)](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
- [George Orwell: Politics and the English Language (The Orwell Foundation)](https://www.orwellfoundation.com/the-orwell-foundation/orwell/essays-and-other-works/politics-and-the-english-language/)

## How this guide was made

Researched from Builders Camp's bootcamp, track and masterclass material and the sources listed on this page, drafted with AI, and fact-checked against every source cited.