Claude Custom Instructions & Project Instructions: A Setup Guide (2026)

Custom instructions vs. project instructions: the difference

They stack, and they're for different things.

• Custom instructions are profile-level. They apply to every conversation you have with Claude, on every topic. This is the home for universal rules: who you are, your tone preferences, how you want answers formatted. Profile-level custom instructions work on every plan, including free.
• Project instructions are scoped to a single Project. They apply only inside that workspace and stack on top of your profile. This is the home for what's specific to one client, one job, or one kind of work. Projects require a paid plan.

The rule of thumb: put what's universal in custom instructions ("be concise," "always verify before recommending"), and put what's specific in project instructions ("this project is for Acme's marketing site; here's the brand voice"). Don't repeat your universal rules in every project.

A good signal that you should create a Project at all: you expect more than ten conversations in that context over time. Below that, custom instructions alone are enough.

How to write instructions Claude actually follows

The most common failure isn't that Claude can't follow instructions — it's that the instructions are vague, bloated, or buried. Four rules fix most of it:

1. Be behavioral, not adjectival. "Be concise" is weak. "When I ask for a plan, give numbered steps with time estimates" is something Claude can act on. Write like you're onboarding a new hire.
2. Keep profile instructions short. They're sent with every message, so a bloated set eats the space Claude needs to actually work. Keep profile-level instructions tight — a few hundred words, not two thousand — and push detail into projects.
3. Lead with the most important rules. Models weight the top of the instructions more heavily. Put your hard constraints first, not at the bottom of a wall of text.
4. Say what not to do. Explicit boundaries ("never invent statistics," "don't pad with caveats") prevent the most annoying behaviors more reliably than positive instructions alone.

If those four rules sound familiar to developers, it's because they're the exact discipline behind a working CLAUDE.md file — the Claude Code equivalent of custom instructions. Same skill, different surface.

A simple template

A custom-instruction set that works tends to cover four things:

• Identity — who you are and what you do ("I'm a solo financial-services consultant; my clients are emerging fund managers").
• Behavior — how you want Claude to operate ("research first, then advise; web-search before giving any pricing or product claim").
• Format — how you want answers ("default to short prose; use a numbered list only when I ask for steps").
• Boundaries — the hard nos ("never fabricate sources; flag anything you're unsure about").

Keep it specific to you. Generic instructions produce generic results.

Custom instructions, CLAUDE.md, and skills — how they relate

There's one idea under three names, depending on where you're working:

• In the Claude app, standing context lives in custom instructions and project instructions.
• In Claude Code or Cowork, it lives in a CLAUDE.md file — the same always-on briefing, in your repo or workspace.
• For repeatable tasks (not just context), you add skills — on-demand know-how Claude loads when a matching job appears.

Get the standing context right first; it's the foundation everything else builds on.

The fastest way to get a set that works

Writing good instructions from a blank page is harder than it looks — most people either write too little (and stay generic) or too much (and get ignored).

The free CoworKit Builder does it for you: describe your role and how you work in plain English, and it generates a structured context file and scores it, so you can see what's strong and what's vague before you paste it into Claude. The output works equally well as custom instructions in the Claude app or as a CLAUDE.md in Claude Code.

And if you want Claude to do recurring work — not just understand you — CoworKit's Starter Kits bundle a tuned context file with ready-to-run skills for specific jobs like running a job search or a consulting practice.

FAQ

What's the difference between Claude custom instructions and project instructions?

Custom instructions are profile-level and apply to every conversation; project instructions are scoped to a single Project and stack on top. Put universal rules in custom instructions and project-specific context in project instructions.

Do custom instructions work on the free Claude plan?

Yes. Profile-level custom instructions are available on every plan, including free. Projects (and project instructions) require a paid plan.

Why does Claude ignore my custom instructions?

Usually because they're vague, too long, or buried. Write behavioral rules instead of adjectives, keep the profile set short, lead with your most important constraints, and state what not to do. The same fixes apply when Claude ignores a CLAUDE.md.

How long should custom instructions be?

Profile-level instructions should stay short — roughly a few hundred words — because they're sent with every message. Push longer, situation-specific detail into project instructions instead.

Is a CLAUDE.md the same as custom instructions?

They're the same idea on different surfaces. Custom instructions are how you give the Claude app standing context; a CLAUDE.md does the same job in Claude Code and Cowork. The writing discipline is identical.

Can a tool write my custom instructions for me?

Yes. The free CoworKit Builder generates and scores a structured context file from a plain-English description of your role, so you start from a strong, specific draft instead of a blank page.