Claude Cowork Instructions vs. CLAUDE.md: What Goes Where
Claude Cowork instructions vs. CLAUDE.md comes down to what loads where. Anthropic's Cowork help pages describe global, folder and project instructions, and the pages we checked never mention CLAUDE.md. Claude Code does load CLAUDE.md at the start of every session. So keep a few must-hold rules in global instructions and the detailed rules in CLAUDE.md.
Below you'll find what the official documentation says (checked on 9 October 2026), what happened in our own tests and a table of what goes where.
Does Claude Cowork read CLAUDE.md? What the official docs say
Start with the Cowork side. Anthropic's help article "Get started with Claude Cowork" describes two kinds of instructions:
- Global instructions are standing instructions that apply to every Cowork session. You set them in Settings > Cowork by clicking "Edit" next to Global instructions. The same article adds a note: in the new Claude experience, global instructions are part of Instructions for Claude in Settings > General, and they apply to every conversation, not only to Cowork.
- Folder instructions add project-specific context to Cowork when you select a local folder on desktop. The article also says Claude can update them on its own during a session. Anthropic's Cowork product guide says they live in the Customize panel of the desktop app.
Menu and setting names were checked on 9 October 2026. The Claude desktop app changes them from time to time, so if you can't find a setting, search for it by name.
Cowork projects add a third layer. According to the Cowork documentation, a project holds local folders, standing instructions applied to every session in that project, useful links and its own memory. Memory is a fourth layer: in sessions in the cloud, Claude starts from what it already remembers from your chats.
Now the Claude Code side. Claude Code's memory documentation is clear: CLAUDE.md files are loaded at the start of every session. Your personal file, ~/.claude/CLAUDE.md, applies to all your projects. A CLAUDE.md in the project folder applies to that project. Files in the folders above it load too, and files in subfolders load once Claude works there.
The two worlds meet in one place. Cowork's overview says it uses the same agentic architecture that powers Claude Code. And Claude Code's memory page has one note about Cowork: in Cowork sessions on your desktop, imports in your personal ~/.claude/CLAUDE.md that point outside the session's working folder are skipped, and the rest of the file loads.
What we could not find is an official page that promises: "Cowork reads the CLAUDE.md in the folder you select, at the start of every task." None of the Cowork pages we checked use that file name. Until such a page exists, treat automatic loading of a folder's CLAUDE.md in Cowork as possible, not promised.
What happened in our testing
We build a rules system for Claude Cowork and Claude Code, so we test exactly this. Here is one real case from October 2026, from a test folder with no client data in it.
The folder had a short CLAUDE.md that pulled in the full rule files with @ imports (the import syntax described in Claude Code's documentation).
- 7 October: a new Cowork task on a Windows PC. The rules loaded. The AI ran the start-up check that the rules ask for, and that check left its trace in a file on disk.
- 8 October: the same folder, a new task, a short greeting as the first message ("Hi, let's start. What's new?"). The AI said that CLAUDE.md only pointed to other rule files and that it hadn't read them. Later in the same task, it reported that the session listed CLAUDE.md as no longer among its instructions, although the file was still on disk. Once we asked it to read the rules, it did, and the start-up check ran.
We noticed one more thing in both tests: the account's own Instructions for Claude showed up inside the tasks, just as the help article says they should. The account-wide layer did reach Cowork.
Two tests are not a study, and we don't know why the second one failed. But it was enough for us to stop relying on automatic loading in Cowork and to add a check that tells us, in the very first reply, whether the rules are in.
Claude Cowork instructions vs. CLAUDE.md: the table
| Layer | Where you set it | What the docs say it reaches | What to put there |
|---|---|---|---|
| Instructions for Claude / Cowork global instructions | Settings > General (new experience) or Settings > Cowork > Global instructions | Every Cowork session; in the new experience, every conversation | Who you are, your language and tone, three to five rules that must hold everywhere, and one line telling Claude to read the folder's CLAUDE.md |
| Folder instructions (Cowork) | The Customize panel, when you select a local folder | Project-specific context for that folder; Claude can update them itself | Short context: what the folder is for, what kind of work it holds, where outputs go |
| Cowork project instructions | The project page | Every session in that project | Guidance for one recurring area of work |
CLAUDE.md in a project folder | A plain file in the folder | Claude Code: loaded at the start of every session in that folder | Detailed rules and workflows for the project, under 200 lines |
~/.claude/CLAUDE.md | A plain file in your home folder | Claude Code: all your projects | Your personal rules for Claude Code |
| Memory | Written by Claude, not by you | Claude Code: auto memory notes; Cowork: memory from your chats (cloud sessions) and project memory | Nothing you rely on as a rule. Read it now and then, and correct it |
Five rules for splitting your instructions
- Keep the global layer short. In the new experience, it reaches every conversation, including chats where your project rules make no sense. Write project rules there as conditions: "If you are working in a folder…".
- Give each rule one home. Claude Code's documentation warns that if two instructions contradict each other, Claude may pick one arbitrarily. If a rule sits in both global instructions and CLAUDE.md, sooner or later one copy gets edited and the other doesn't.
- Don't keep fixed rules where Claude can rewrite them. Claude can update folder instructions during a session. That's handy for context, but rules you want to stay fixed belong in a file you control and back up, such as CLAUDE.md.
- Keep CLAUDE.md under 200 lines. That is Anthropic's own target for Claude Code: longer files consume more context and reduce adherence. Imports help you organize a long file, but they don't make the load smaller.
- Ask for a receipt that the rules loaded. You can't see what's in Claude's context. You can see its first reply. Give your rules a version number that changes with every edit, and have Claude confirm at the start which rules and which version it read. If the confirmation is missing or shows an old version, stop and have Claude read the file before any work starts. This is not a guarantee, because Claude can skip the confirmation even when the file has loaded. But noticing a missing confirmation takes seconds, while work done without your rules may have to be redone.
Loaded is not the same as followed
Getting the rules into Claude's context is only half the job. Claude Code's documentation says plainly that Claude treats CLAUDE.md as context, not enforced configuration, and that there's no guarantee of strict compliance. For steps that must happen every time, it recommends hooks: commands that run at fixed points, whatever Claude decides.
If your rules load but still get skipped, read why Claude ignores your CLAUDE.md and what makes rules stick. If Claude reports work as finished when it isn't, see how to check when your AI says "done". And before Claude works in a folder with files you care about, read can Claude Cowork delete your files?
Claims Need Receipts adds rules, project memory kept in files and guard scripts that check the AI's claims to Claude Cowork and Claude Code, for people who aren't programmers. If you already have a CLAUDE.md, it stays: the installer shows you a dry run first and adds only one marked block.
FAQ
Does Claude Cowork read CLAUDE.md?
The Cowork help pages from Anthropic that we checked on 9 October 2026 don't mention CLAUDE.md; they describe global, folder and project instructions. Claude Code's documentation says CLAUDE.md loads at the start of every Claude Code session. In our Cowork tests, a folder's CLAUDE.md loaded one day and not the next, so ask for a "rules loaded" line.
Where are Claude Cowork global instructions?
According to Anthropic's help article, open Settings > Cowork and click "Edit" next to Global instructions. In the new Claude experience, they are part of Instructions for Claude in Settings > General, and they apply to every conversation, not only to Cowork. Keep them short, and write rules meant only for project work as conditions.
Why is Claude not following my project instructions?
First check that they loaded at all: ask for a version line in the first reply. If they loaded and Claude still skips them, know that the documentation promises no strict compliance: Claude treats CLAUDE.md as context, not enforced configuration. Short, specific rules help, and a check outside the AI catches what slips through.
What is a guard script, and do I need to write one?
No. A guard script is a small program that checks files on your disk (a date, a number, a log entry) and prints a warning when a receipt is missing. In Claims Need Receipts you don't write or edit any of them: your AI reads the guide, asks you seven questions, shows what it will add and installs everything only after you say yes. The one requirement is Python 3.10 or newer on your computer; if it's missing, your AI walks you through installing it.
Sources
All accessed 9 October 2026.
- Anthropic Help Center: Get started with Claude Cowork
- Anthropic Help Center: Understanding Claude's personalization features
- Claude documentation: Cowork overview
- Claude documentation: Organize work with projects (Cowork)
- Claude Cowork product guide: Getting started
- Claude Code documentation: How Claude remembers your project
Last updated: 9 October 2026
Claude is a trademark of Anthropic. Claims Need Receipts is an independent product and is not affiliated with, endorsed or sponsored by Anthropic.