Claude Preferences: How to Set Instructions for Claude
Where Claude's preferences live now, what to write in them, three examples, and how they fit with Projects and skills.

Claude preferences are standing instructions you write once and Claude applies to every conversation. The field used to ask "What personal preferences should Claude consider in responses?"; in 2026 it's called Instructions for Claude, and you reach it from your initials > Settings.
If you went looking for the old box and couldn't find it, that's why. Below: where it is now, what to write, three examples, and how it fits with Projects and skills.
Where are preferences in Claude settings?
Click your initials in the lower-left corner, choose Settings, and find the field labeled Instructions for Claude (Anthropic, updated September 16, 2026). Type, save, and every new conversation starts with it.
A few naming changes explain most of the "I don't have that" searches:
- Old label: "What personal preferences should Claude consider in responses?" Older guides call it profile preferences and send you to Settings > Profile. It's the same feature; Anthropic's current guide goes straight from Settings to Instructions for Claude.
- Cowork users: if you set "Global instructions" in Claude Cowork, they're now part of Instructions for Claude. Anthropic suggests checking the merged text is still right.
- Capabilities and Memory: memory settings now sit under Settings > Your Claude (their own Settings > Memory page in the older layout); they're separate from your instructions. More on that in the Claude memory guide.
What should you put in Claude preferences?
Write the briefing you'd give a sharp new colleague on day one, in four parts:
- Who you are. Role, field, level. "Senior backend engineer, mostly Go and Postgres" sets how technical Claude should be.
- How you want answers. Specific enough to change behavior: "start with the answer, explain only if I ask" beats "be concise".
- What to stop doing. Usually the highest-impact part. "No preamble." "Don't hedge every sentence." "When editing my writing, fix problems but keep my voice."
- Standing context. Tools and terms you reuse: "I deploy on Vercel." "'The app' means our iOS client."
Leave out anything that changes month to month. Your current project, this week's priorities and the deal you're working go stale fast, and Claude applies stale instructions with full confidence. That context belongs in a Project or in a layer built to stay current.
A template you can adapt
I'm a [role] working in [field/stack]. I'm [beginner/intermediate/expert] in [area].
Answers:
- Start with the answer, then explain if needed
- Be direct: tell me what's weak and why
- No preamble, no summary at the end
Don't:
- Rewrite my voice when editing my writing
- Hedge every statement
- Ask for confirmation on obvious next steps
Context I reuse: [tools, recurring terms, typical scenarios]
What are good Claude personal preferences examples?
Good examples are short, specific and mostly about what to change. Three you can copy and edit:
Developer
Full-stack developer, TypeScript and Python, deploy on Vercel and Fly.io.
Show code first, explanation after, and only explain non-obvious parts.
Use pnpm, never npm. Prefer small diffs over rewrites.
Don't add comments that restate the code. Flag security issues even if I didn't ask.
Marketer
Head of content at a B2B SaaS company; audience is technical founders.
Write in plain US English, short paragraphs, no buzzwords.
When I share a draft, list the three biggest problems before suggesting edits.
Never invent statistics. If a claim needs a source, say so.
Founder
Solo founder, pre-seed, building a scheduling tool for clinics.
I make decisions fast: give me a recommendation first, then the trade-offs in bullets.
Push back when my plan has an obvious flaw.
Assume I know startup basics; skip definitions.
Keep it to the instructions that cover most of your conversations. If you can't read the whole field in about a minute, it's too long: it loads into every chat, and the rule you care about gets buried.
Preferences vs Project instructions vs skills
Claude has three personalization layers, and they load together rather than overriding each other:
| Layer | Scope | Use it for |
|---|---|---|
| Instructions for Claude | All your conversations | Who you are, how you work, standing rules |
| Project instructions | Chats inside one Project | Context for one client, product or body of work |
| Skills | When invoked or relevant | Tone, format and repeatable behaviors |
Projects are available on every plan, and free accounts can create up to five (Anthropic). Each Project's instructions and files sit on top of your Instructions for Claude, so you don't repeat the general part per Project.
What happened to Claude styles?
Styles are moving to skills. According to Anthropic's migration note, custom styles become skills automatically but start disabled, so you turn on the ones you want under Customize > Skills and call them with a command such as /concise-pirate-style. The default Concise, Explanatory and Formal styles go away, the Learning style survives as an installable skill, and the Styles menu disappears once migration finishes. If you'd rather skip skills, describe the tone you want in Instructions for Claude instead.
Common mistakes to avoid
- Too vague. "Be helpful and concise" changes nothing. Every line should be specific enough that you'd notice if Claude ignored it.
- Project facts in the global field. "This is a React app using Supabase" will show up when you ask Claude to plan a trip. It goes in Project instructions.
- Set and forget. Review the field every month or two, and add a line whenever you catch yourself making the same correction twice.
- Expecting it in Claude Code. The Claude app's instructions don't configure Claude Code, which reads
~/.claude/CLAUDE.mdfor personal preferences; see what CLAUDE.md is.
Hand-written instructions also have a ceiling. In an August 2026 Unabyss benchmark, a hand-written CLAUDE.md reduced follow-up questions but barely moved cost, time or output quality, while structured context from the user's real work cut cost by 36% and time by 38% (What Structured AI Context Is Worth). Instructions tell Claude how you like to work; they can't tell it what you're working on.
Where do Claude preferences stop?
They stop at the edge of Claude. Open ChatGPT or Cursor and none of your instructions are there; you'd write equivalent text in each tool's settings and keep every copy in sync by hand.
Unabyss keeps your personal context in one place, built from Gmail, Notion, Slack and the other apps where it already lives, and serves it to Claude, ChatGPT and Cursor over MCP. To make Claude check that context before it guesses, paste one of the ready-made blocks from the Claude and ChatGPT global instructions guide into your Instructions for Claude field. For the bigger picture of what persists between chats, see how to give Claude persistent context.
→ Set your context once and use it in every AI tool with Unabyss