How to make AI write like you: a voice skill built from your own edits

Ask an assistant for an email in your voice and you get a competent note that could have come from anyone. It opens with “I hope this finds you well”, says “rest assured” somewhere in the middle and closes on a summary nobody asked for. With nothing specific to go on, the model reaches for the phrases that are safe for everybody.

The usual fix makes it worse. People describe their voice in adjectives, and “direct but warm” is how nearly everyone would describe themselves. An adjective gives the assistant a mood, and it gives you nothing to check the draft against.

So stop describing your voice and show it, then let your own edits do the rest. The method takes five things you wrote, two short prompts and a few rounds of crossing out. It ends with a voice skill the assistant picks up whenever you ask it to draft in your name.

Use the assistant your employer provides, signed in with your work account. Pick five things you wrote yourself and are free to share, and choose them for range rather than polish. Two emails, a note to your team, a reply to a supplier and a short post say more about you than five versions of the same status update. Paste them into a new chat, then type this.

Each sentence does one job. “Do not imitate these” stops the assistant copying the surface of your samples, their topics and their stock lines, into everything it writes next. “Rules a stranger could follow” makes the description concrete: how you open, how long your sentences run, whether you ask before you tell, how you sign off. A stranger cannot follow “warm”, but they can follow “signs off with Thanks and a first name”. “Then draft a short note” tests the rules while your samples are still on the screen.

Read the rules before you read the draft. A rule that would fit a colleague as well as you is too vague, so ask for it to be made specific or cut it.

The first draft will be closer, and still wrong in places. The usual tell is an opener you have never used in your life. Cross out every phrase that is not you, however small, then paste your edited version back and type the second prompt.

Repeat this over a few drafts and the never list becomes the longest part of the skill. That is expected, and it is the useful part. A voice skill is mostly what you cross out.

Then keep both lists together: in the same chat, ask for the rules and the never list to be saved as a voice skill. Where that lives in Claude and ChatGPT is set out further down this page.

Examples carry what adjectives cannot. Anthropic's prompting guidance names three to five relevant, varied examples as one of the most reliable ways to steer format, tone and structure [1], and its guidance on writing skills says examples convey style and detail better than descriptions do [2]. Five samples of different kinds sit at the top of that range.

Asking for rules instead of an imitation makes the result checkable. A rule is a statement you can read, disagree with and correct, and the editing step depends on exactly that.

Naming phrases works because the ones an assistant falls back on are a short, repeated set. Each one you name becomes a rule it can check before it writes, which no adjective ever gave it. Where the reason for a cut is not obvious, add it in a few words: Anthropic's guidance is that Claude does better with the reason behind a constraint than with a bare ban [1]. “Opens with the point, never with a pleasantry” teaches more than a list of banned greetings.

The loop keeps the skill honest. OpenAI's advice for any skill is that when it later gets something wrong, you ask for that correction to be added to the skill [3]. Every draft you fix makes the next one closer.

The two prompts above cover everyday writing. Two variations come up often enough to keep beside them.

The first keeps two registers apart. If the two sets of rules barely overlap, save them as two skills: Anthropic advises keeping each skill to one workflow [4], and OpenAI one job per skill [5].

The second is the habit that keeps the skill current. Claude updates an existing skill's files when you ask it to edit them [6], and in ChatGPT the same request adds the correction to the skill [3]. Check the change in a fresh chat rather than the one you edited it in.

Both products save this kind of standing instruction as a skill: a folder whose one required file, SKILL.md, carries a name, a description and the instructions [4, 5]. The assistant reads only each skill's name and description until a request matches, then loads the rest [5, 7], so a long never list stays out of your other chats. These are the paths as checked on 11 October 2026.

The Claude column is from Anthropic's help centre and documentation [4, 6, 7, 8, 9], and the ChatGPT column from OpenAI's Help Center, Academy and documentation [3, 5, 10, 11, 12]. Anthropic's help centre lists skills on every Claude plan, Free included [8]; OpenAI documents skills for eligible plans [10].

A skill is the right home because it loads only when you are drafting. Instructions for Claude, by contrast, apply to all your conversations [13], so a never list kept there rides along on every spreadsheet question as well.

Where an administrator has switched off user-created skills, as Claude Team and Enterprise owners can [14], the project route still works for one person. Create a project for drafting, paste both lists into its instructions and do your writing there. You lose the automatic loading and keep everything else.

If your organisation is still choosing between the two assistants, our comparison of ChatGPT Enterprise and Claude Enterprise sets out what each vendor publishes on plans and admin controls.

A voice skill is a set of instructions the assistant reads, and that sets its limits. Plan for these before you rely on it.

A voice skill is a good first skill because the person who owns it is the only judge it needs. The company version is harder: many people writing skills, some of them shared, and someone deciding which switches are on, who may publish to the organisation's library and who owns each shared skill. Rolling out AI work tools across a company, with the training, controls and measurement that make them pay, is core work for us. On a rollout we would set those switches and owners up as controls, under our AI governance implementation work. At a large international marketing agency we contribute engineering to an internal Claude Code skill platform that now carries around 600 production skills, written across the business by the people who do the work.

We take small fixed-scope statements of work as well as larger builds, and a fixed-scope rollout can start within a day of the scope being signed. If that would help, book a call from the end of this page or email enquiries@1aym.com.