# How do I teach an AI agent my writing style?

Give it real samples of your writing, a short note on your habits, and a way to see what you change before you send. The free route is a style note in custom instructions plus samples in a Claude or ChatGPT project, or a rules file such as CLAUDE.md for a coding agent; a tool such as Ownhand builds a Hand from 5 to 10 samples and learns from the exact text you send.

By Jordan Gibbs, founder of Ownhand at Thalient Labs. Updated 2 October 2026.

Every major assistant has a free place to describe how you write. That gets you part of the way. The hard part is that style lives in small habits you don't notice: how long your messages run, whether you start with a greeting, how often you hedge. A model copies those from your own sentences far better than from a description of them.

In 2025, researchers tested this. Wang and colleagues asked models to imitate more than 400 real authors, using a few samples of each. The models did fairly well on structured writing like news and email, and worse on casual posts. A detection tool still classed most imitations as machine-written. The share it judged human stayed under 55 percent and was often under 20 percent. Samples of your writing beat a description of it. What you change before you send is the other half.

## What your style is made of

Before you teach it, know what you are teaching. These are the habits Ownhand measures from your samples, and they are a good checklist for any method:

- Length: how many words your messages usually run, and how long your sentences are.
- Contractions: "don't" or "do not".
- Openers and closers: a greeting or none, a sign-off or none.
- Capitals and punctuation: lowercase starts in chat, exclamation marks, emoji.
- Hedging: "I think" and "maybe", or flat statements.
- Warmth: whether you thank people and say how you feel.
- Word length: short plain words, or longer formal ones.

The same draft lands differently for different writers. Here one AI draft was rewritten by two demo writers built from different samples:

These examples are real output from the live API. Each AI draft was rewritten by a demo Hand: a fictional writer (an engineer, a founder, an ops lead, a designer or a consultant) built from a handful of samples.

**AI draft:** "I apologize for the delay in getting back to you. I've been dealing with a few urgent matters, but I'll have the slides over to you by end of day tomorrow."

**Demo Hand, engineer:** "sorry for the delay. slides by eod tomorrow"

**AI draft:** "I apologize for the delay in getting back to you. I've been dealing with a few urgent matters, but I'll have the slides over to you by end of day tomorrow."

**Demo Hand, ops:** "Slides by end of day tomorrow. Sorry for the delay."

Same facts, same apology. The engineer writes in lowercase with shorthand. The ops lead leads with the deliverable and puts the apology last. A note saying "short and casual" fits both, so it can't tell them apart.

## Step 1: pick your samples

1. Use only things you wrote and sent yourself. Skip anything an AI drafted or someone else edited.
2. Cover the places you write in. A few chat messages, a few emails, a pull request description or a doc if you write those.
3. Keep the short ones. A two-line message shows how short you go. That's often the habit models miss.
4. Pick 5 to 10. Past that, choose for variety over volume.
5. Keep the typos and lowercase. Paste them as they were sent. Cleaning them up teaches the cleaned-up version.

## Step 2: write a short style note

Samples show the habits; a note tells the model which ones to keep. Use numbers and real words in place of adjectives:

```
My writing: chat messages are 1 to 2 sentences, no greeting, often lowercase.
Emails are under 120 words, start with the point, sign off with my first name.
I use contractions. I never use exclamation marks or emoji at work.
Words I never use: delighted, kindly, per my last email.
When you draft for me, match the samples I gave you over this note.
```

## Option A: Claude

1. Open Settings and add your note to the instructions for Claude. Anthropic says they apply to all your conversations.
2. For writing work, create a project. Put the note in its instructions and add your samples to it. Free accounts can have up to 5 projects.
3. Memory is on by default for Free, Pro and Max on the web, desktop and mobile. Check what Claude saved, and edit it, under Settings, then Memory.
4. In Claude Code, put the note in `~/.claude/CLAUDE.md` for every project, or in `CLAUDE.md` at a repo's root for that repo.

## Option B: ChatGPT

1. Open Settings, then Personalization. Turn on customization and paste your note into Custom Instructions. OpenAI allows 1,500 characters on Free and Go, and 5,000 on Plus, Pro and business plans.
2. Make a project for a set of samples. Open its menu and choose Project settings. Add instructions there. They apply only inside that project and override your global ones. Add your samples as files.
3. Under Settings, Personalization and Memory, check what ChatGPT has saved about you. Depending on your account, it can also reference past chats.

## Option C: coding agents

- Cursor: put the note in User Rules (Customize, then Rules) for every project. User Rules apply to Agent chat. They don't apply to inline edits. For a team, add a project rule in `.cursor/rules` as an `.mdc` file, or use `AGENTS.md`.
- GitHub Copilot: repository instructions go in `.github/copilot-instructions.md`. Personal instructions only reach Copilot Chat on GitHub.com, JetBrains IDEs, and the Copilot CLI. GitHub's support table doesn't list pull request summaries, so test before you count on either file for pull request text.
- Any agent that reads `AGENTS.md` or `CLAUDE.md` can take the same note. Keep it short, since Claude Code loads its memory files into every session.

## How the options compare

| Method | Holds samples | Fits the occasion | Learns from what you send | Cost |
| --- | --- | --- | --- | --- |
| Style note in custom instructions or a rules file | No, a description only | Only if you write it in | No | Free |
| Claude or ChatGPT project with samples | Yes | One project per kind of writing | No | Free or your plan |
| Claude or ChatGPT memory | No | No | Only what you say in the chat | Free or your plan |
| Ownhand | Yes, 5 to 10 to start | 12 occasions, from `chat_dm` to `pr_description` | Yes, from the exact text you send | $1 free, then about half a cent a rewrite |

Start with the free options. They're enough if you write in one or two places, use one assistant, and edit every draft yourself anyway. A project with good samples and a tight note gets most people most of the way.

## Step 3: let it see what you send

This is the step the built-in options leave out. You fix a draft in your email client or Slack, and the assistant never sees the fix. Next week it makes the same mistake. Memory catches what you tell it in the chat. It doesn't catch what you change after.

Close the loop by hand: paste your sent version back into the chat, say what you changed and why, and ask the model to update your note. It works if you keep it up.

Ownhand makes it part of the flow. After you send a draft, your agent reports the exact text you sent. A reason you give in your own words, like "too formal", becomes a rule right away. A change you keep making becomes a rule once it shows up twice. A rule can apply everywhere or to one occasion. You can pin or reject any rule and roll your Hand back to an earlier version. Details are in [how Ownhand learns from what you send](https://ownhand.dev/answers/how-ownhand-learns-from-edits).

## Teaching Ownhand

1. Connect it to your agent (the command is under Install below, and other clients are on the Connect page).
2. Say "Set up my Ownhand Hand." Your agent asks for 5 to 10 things you wrote and how you'd describe your writing. Then it builds the Hand.
3. Ask for something as you would anyway, like "reply to Dana in my voice". Your agent drafts it and sends it to the write tool with the right occasion. You see the result before anything goes out.
4. Send it, edited or not. Your agent reports what you sent, and the Hand learns.

Your samples, the style card built from them, and the rules learned from your feedback are stored with your account. Threads your agent passes for a reply are used for that one request and never saved. For the full picture, see [whether Ownhand is safe to use](https://ownhand.dev/answers/is-ownhand-safe).

## Follow-up questions

### How many writing samples does an AI need to learn my style?

Five to ten short pieces you sent is a good start. They just need to cover the places you write. Ownhand asks for 5 to 10, and builds its style card once it has at least three texts you wrote. More samples help less than samples from the right places.

### Can I just describe my style in a sentence?

You can, and it helps a little. A line like "casual and short" leaves the model to guess what casual and short mean for you. Your own sentences show it your length, your openers, your punctuation. A description rarely does.

### Does ChatGPT or Claude memory learn my writing style on its own?

Partly. Both can remember the preferences you state, and ChatGPT can draw on past chats. Neither sees the version you edited and sent from your email or Slack. The changes you make after the chat are lost.

### Should I use AI-written drafts as samples?

No. Use only things you wrote and sent yourself. A sample an AI drafted teaches the model its own habits back. That's the opposite of what you want.

### Is my writing used to train models if I use Ownhand?

Ownhand doesn't train its own models on your data, and one person's writing never shapes another person's Hand. Feedback and style card updates go to a learning model. Its provider may use them to improve its models, so read the privacy page before you add samples.

## Sources

Checked on 2 October 2026. If something here is out of date, email [hi@thalientlabs.ai](mailto:hi@thalientlabs.ai) and we will fix it.

- [Claude Help Center: Understanding Claude's personalization features](https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-features): account-wide instructions in Settings, project instructions, free accounts limited to 5 projects
- [Claude Help Center: Chat search and memory](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context): Settings > Memory, on by default for Free, Pro and Max
- [Claude Code docs: How Claude remembers your project](https://code.claude.com/docs/en/memory): ~/.claude/CLAUDE.md for all projects, ./CLAUDE.md per project
- [OpenAI Help Center: ChatGPT custom instructions](https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions): Settings > Personalization; 1,500 characters on Free and Go, 5,000 on paid plans
- [OpenAI Help Center: Projects in ChatGPT](https://help.openai.com/en/articles/10169521-projects-in-chatgpt): project instructions override your global custom instructions
- [OpenAI Help Center: Memory in ChatGPT](https://help.openai.com/en/articles/8590148-memory-faq): saved memories and reference chat history
- [Cursor docs: Rules](https://cursor.com/docs/context/rules): project rules in .cursor/rules as .mdc files, User Rules, AGENTS.md
- [GitHub Docs: Support for different types of custom instructions](https://docs.github.com/en/copilot/reference/custom-instructions-support): which Copilot features read repository, path-specific and personal instructions
- [Wang et al., Catch Me If You Can? Not Yet (EMNLP Findings, 2025)](https://arxiv.org/abs/2509.14543): LLMs imitating everyday authors from a few samples, 400+ authors, 40,000+ generations per model
- [Ownhand source code](https://github.com/jordan-gibbs/ownhand): what a Hand measures, how rules are learned, the occasions

## Install

In Claude Code it is one line. You sign in through your browser, so there is no API key to copy.

```
claude mcp add --transport http --scope user ownhand https://ownhand.dev/mcp
```

For Claude, ChatGPT, Cursor, VS Code, Codex, Gemini CLI, Windsurf, Zed, Goose and Cline, see [Connect](https://ownhand.dev/connect). New accounts get $1 of free credit. After that you pay as you go, and each rewrite costs about half a cent.

More answers: [all questions](https://ownhand.dev/answers). Full reference: [Docs](https://ownhand.dev/docs).

---

Page: https://ownhand.dev/answers/teach-ai-my-writing-style. Index for agents: https://ownhand.dev/llms.txt