A useful agent should not only know the task. It should learn how you work: your tone, your preferred words, your level of directness, your rhythm, your judgement, and the phrases that make something sound like you.
Why voice matters.
Most people start by asking an agent to write an email, a note, or a briefing. The first draft may be accurate, but it often sounds generic. It may be too polished, too corporate, too soft, too eager, or too unlike the person who is meant to send it.
Your voice is not decoration. It carries your intent. It tells people how serious you are, how warm you are being, how much uncertainty exists, and what kind of relationship you want to maintain.
This is not impersonation.
The aim is not to let the agent pretend to be you without review. The aim is to help it prepare better drafts for you to check.
Use safe examples. Do not feed it private emails, client messages, HR material, confidential strategy, or other people's personal data just to make the style better. Start with writing you are comfortable using for training.
Build a voice card.
A voice card is a small, readable set of rules. It says how you usually write, what you avoid, how direct you want to be, and when the agent must ask before changing tone.
Part
What it captures
Example
Voice principles
The broad feel of your writing
Plain, warm, direct, practical
Preferred phrases
Words that sound natural to you
"Let's make this useful", "the safe first step"
Banned phrases
Words that make the draft sound wrong
Corporate filler, empty hype, fake enthusiasm
Structure rules
How you like ideas ordered
Start with the point, then explain the risk, then give the next action
Approval rules
When the draft is not allowed to leave
Anything with commitment, criticism, money, people impact, or legal risk needs review
Teach your house language.
Everyone has language that means something specific. You may talk about a "safe folder", a "stop-line", a "public-first start", a "Chief of Staff twin", "good enough", "judgement", or "what good looks like".
Write those meanings down. A glossary keeps the agent from flattening your language into generic AI words.
Teach your decision style.
Your agent also needs to learn how you make decisions. Not so it can decide for you, but so it can help you gather the right evidence, use the right tools, slow down at the right moments, and challenge you when something material is missing.
Write down the processes you use, the tools you check, the people you ask, the evidence you trust, the evidence you treat carefully, and the stop-lines that mean the agent must pause. A good agent should know when to prepare options, when to ask questions, when to raise a structured challenge, and when to leave the decision with you.
Where to put the customisation.
The safest place depends on the tool. For local project work, use visible project instructions such as AGENTS.md or CLAUDE.md. For a Chief of Staff twin, use visible Library files such as tone-of-voice.md, house-language.md, and decision-style.md. For skills, keep the skill text readable. For custom instructions or memory, only store guidance you are comfortable keeping beyond the current task.
Do not hide the voice guide. If you cannot inspect it, you cannot improve it.
Copy these into your agent.
Use these prompts to build the voice gradually. The agent should show you the guide before it stores or writes anything.
Keep correcting it.
The first voice guide will be rough. That is fine. Every time the agent sounds too generic, too polished, too formal, too casual, or too unlike you, correct the guide.
The goal is not perfect imitation. The goal is fewer bad drafts, clearer meaning, and a safer route from your intent to usable words.