3 levels control how Claude behaves in Cowork. Most people only use one.
Ever notice Cowork is great one day and completely off the next? It’s not random. Claude now carries some memory over from your chats, but memory is a loose summary. It won’t reliably hold your rules, your project details or the exact way you want a task done.
That’s what these three levels are for. Most people set up one, if that. Set up all three and Claude stops guessing.
Good To Know
Cowork is on paid plans (Pro, Max, Ultra, Team and Enterprise). Anthropic is also merging Cowork into regular Claude chat, so if you don’t see a Cowork option, describe the task in any chat. All three levels still apply.
01
Rules for everything
Global instructions are the baseline Claude follows in every task. You set them once. Tell Claude who you are and how you always want things done.
In the new layout, paste them into Instructions under Settings, then Your Claude. In the older layout, go to Settings, then Cowork, and click Edit next to Global instructions, paste and save.
Only put things here that are true for all your work. Anything project-specific goes in Level 2.
02
Your project folder
This is where most people drop off, and it matters most. When you work from a folder on desktop, Claude can read the files in it and use the folder’s instructions as project context. Different folders get different context, so your client work and your internal reports behave differently.
Think of these files as onboarding docs for a new hire who rereads them before every task. Make a folder for your project and add these three.
Folder Setup
Put the three files at the top, then add /templates/ for examples you like, /references/ for brand guides and briefs, /drafts/ for work in progress and /outputs/ for finished work. Working on the web? Cowork projects hold their own files, instructions and memory, so add the same three files there.
If your global instructions say "casual" and this folder needs "formal," don’t make Claude guess. The working-rules file above says the folder wins for this project.
03
The work, done the same way every time
Skills are saved workflows. Claude reads each skill’s description and pulls in the right one when your request matches, or you can ask for it by name. No re-explaining.
The easiest way to make one is to describe it in a new chat. Claude asks about your process, builds the skill file, and you save it. It then shows up under Customize (being renamed Plugins), then Skills, where you can switch it on or off. Skills need code execution turned on.
Build small skills by job instead of one giant "writing" skill. A client-email skill, a team-update skill and a blog-draft skill each get their own rules, and Claude loads only the one that fits.
With all three, Cowork stops feeling like a chatbot and starts feeling like someone who works with you.
01
Always ask for a change log
It’s already in the working-rules file. You can review every file Claude touched instead of digging through folders.
02
Put real examples in /templates/
Claude matches examples better than it follows abstract rules. Drop in a report you love and tell Claude to use it as the model.
03
Update the files as you go
Every time Claude gets something wrong, add a rule. Every time your role changes, update about-me.md. The people who get the most out of Cowork treat these as living documents.