Should I use Ownhand or a style prompt in custom instructions?

Start with a style prompt in the custom instructions your tool already has, since Claude, ChatGPT, Cursor and GitHub Copilot all offer one for free and it is enough for a few fixed habits; move to Ownhand when you want drafts that sound like your real messages, change for each occasion and thread, and improve from the text you actually send.

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

A style prompt is a note you write once about how you want text to sound. You save it where the assistant reads it before every chat. Every major assistant has a place for one, and it costs nothing. Ownhand is a paid tool that rewrites drafts from samples of your own writing. Most people should try the free note first. This page shows where to put it in each tool, how to write one that works, and where it runs out.

Where the free options live

Claude

• Instructions for Claude. Click your initials in the lower left, open Settings, and fill in Instructions for Claude. It applies to every chat, on every plan.
• Projects. Open Projects, create one, and use Set project instructions. You can also add files to the project's knowledge, so a project called "Writing as me" can hold real samples. Free accounts get up to five projects.
• Memory. In Settings, then Memory, turn on Generate memory from chats. Claude saves topics as you chat, keeps a separate memory per project, and lets you edit or delete what it saved. It is on Free, Pro and Max.
• Claude Code. Put your style note in ~/.claude/CLAUDE.md for all your projects, or in a project's own CLAUDE.md. Claude Code loads both at the start of every session.

ChatGPT

• Custom instructions. Open Settings, then Personalization, make sure Enable customization is on, and fill in Custom instructions. Free and Go hold 1,500 characters; Plus, Pro, Business, Enterprise and Edu hold 5,000.
• Projects. In a project, open the menu, then Project settings, and add instructions. They apply only in that project and override your global custom instructions, so you can keep one project per kind of writing.
• Memory. Settings, Personalization, Memory. ChatGPT decides what to keep from your chats and shows a summary you can correct.

Cursor and GitHub Copilot

• Cursor. User rules in Cursor Settings apply across all your projects. Project rules live in .cursor/rules and go into version control. Cursor also reads AGENTS.md. The old .cursorrules file is legacy, so move it into a project rule.
• GitHub Copilot. Repository instructions go in .github/copilot-instructions.md, with path-specific files under .github/instructions and support for AGENTS.md. Personal instructions are set in Copilot Chat on GitHub.com and take priority over repository and organization ones.

Write a style prompt that works

Most style prompts fail because they describe a voice in adjectives. "Casual, direct, friendly" fits a million people, so the model writes the average of them.

Research points the same way. One 2025 study found that few-shot prompting, where the model sees examples, gave up to 23.5 times higher style-matching accuracy than zero-shot. An EMNLP 2025 paper says personal style is subtle and hard to specify through prompts.

1. Paste two or three short things you sent, as they went out, typos and all. Label each one with where it went.
2. Give numbers in place of adjectives: how long your Slack messages run, whether you use greetings, whether you sign emails.
3. List the openers and words you never use. The list of AI writing tells is a good start.
4. Add one line per setting you write in often, such as "PR descriptions: what changed, why, how to test".
5. Test it on three real messages and fix the note where the drafts miss.

Writing as me:
- Slack: one or two lines, lowercase is fine, no greeting, no sign-off.
- Email to people I know: first name, the point in line one, sign "J".
- Never write "I hope this finds you well" or "just circling back".
Samples:
[Slack] can you look at the deploy logs before 3? something's off with retries
[Email] Hi Sam, the contract is signed. Invoice goes out Friday. J

When a style prompt is enough

If you mostly want the model to stop sounding stiff, a note like this gets you a long way. It suits people who write most messages themselves and use AI for a first pass, people who write in one or two settings, and anyone who doesn't want another tool or another bill. It costs you nothing to try, and you can edit it whenever a draft misses.

Where it falls short

A note is fixed. It says the same thing to a Slack DM, a cold email and a PR description, unless you keep a separate note or project for each. It doesn't see the thread you're replying to. So it can't tell that this channel is all lowercase, or that your manager writes in two lines. It never learns from what you sent. When you rewrite the draft in Gmail before sending, that edit stays in Gmail.

Memory does not close that gap. It saves what the model decides to keep from your chats, and it never sees the version that went out. The same draft also lands differently for different writers, and that's hard to put in a note. Two real rewrites of one stiff draft:

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.

Both are short and polite. A note saying "casual and brief" could give you either one, or neither. The difference lives in habits like lowercase, "eod", and "sorry" where the other writer says "Apologies". Those come from samples.

What Ownhand does differently

Your agent writes the draft as usual, then sends it to Ownhand with the occasion. Ownhand rewrites it from your Hand, checks it, and sends it back for you to read. Nothing goes out without your OK. After you send it, the agent reports the exact text you sent. Your Hand updates from the difference. The details are in how Ownhand learns from what you send.

Use both

They work together. Keep your custom instructions for how the assistant talks to you, and let Ownhand handle text that goes out under your name. Paste the agent blurb from the docs into your custom instructions, CLAUDE.md or AGENTS.md, so the agent knows when to call it. For setup in a specific app, see how to make Claude write like you or how to make ChatGPT write in your voice.

Follow-up questions

Can I paste my own writing samples into custom instructions?

Yes, within the limits. ChatGPT custom instructions hold 1,500 characters on Free and Go plans, and 5,000 on paid plans. That fits two or three short messages. A Claude or ChatGPT project has more room, since you can add files as reference material.

Does memory in ChatGPT or Claude learn my writing style?

It can pick up preferences you state, like wanting short replies. Both apps decide on their own what to remember from your chats. Neither one sees the message you finally sent in Slack or Gmail, so your edits there never reach it.

Can I build a custom GPT that writes like me?

You can't do this on a personal account any more. OpenAI's help center says Free, Go, Plus and Pro accounts can no longer create or publish new GPTs. OpenAI plans to retire custom GPTs in favor of plugins. Business, Enterprise and Edu workspaces can still create them where settings allow.

Do I have to drop my custom instructions to use Ownhand?

No. Keep them for how you want the assistant to talk to you. Ownhand only handles text you send as yourself. Add the agent blurb from the docs to your instructions so the agent knows when to call it.

Which costs less?

Custom instructions are free on every plan that has them. Ownhand gives new accounts free credit, then charges per rewrite and per feedback, with no subscription. If you send a handful of messages a week, the free option may be all you need.

Sources

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

• Claude Help Center: Claude's personalization features: Instructions for Claude, project instructions, skills
• Claude Help Center: creating and managing projects: project instructions and knowledge; free accounts get five projects
• Claude Help Center: chat search and memory: Settings > Memory, per-project memory, Free, Pro and Max
• Claude Code docs: how Claude remembers your project: CLAUDE.md and AGENTS.md, loaded at the start of every session
• OpenAI Help Center: ChatGPT custom instructions: Settings > Personalization; 1,500 or 5,000 characters by plan
• OpenAI Help Center: memory in ChatGPT: what memory draws on and how to review it
• OpenAI Help Center: projects in ChatGPT: project instructions override global custom instructions
• OpenAI Help Center: creating and editing GPTs: personal accounts cannot create new GPTs; planned retirement
• Cursor: rules: project rules in .cursor/rules, user rules, .cursorrules is legacy
• Cursor docs: rules: rule types and when each applies
• GitHub Docs: adding repository custom instructions for Copilot: .github/copilot-instructions.md, path-specific files, AGENTS.md
• GitHub Docs: about customizing Copilot responses: personal, repository and organization instructions, and their order
• How well do LLMs imitate human writing style? (arXiv, 2025): few-shot prompting gave up to 23.5 times higher style-matching accuracy than zero-shot
• Wang et al., LLMs still struggle to imitate the implicit writing styles of everyday authors (Findings of EMNLP 2025): personal style is hard to specify through prompts
• Ownhand source code: occasions, the reply fingerprint, the checker and the feedback loop