CLAUDE.md in Cowork: Check Which Instructions Apply

For CLAUDE.md in Cowork, use the instruction mechanisms documented for your Cowork task and explicitly verify any file you expect it to read. Do not assume Claude Code’s file-loading behavior applies simply because both products use Claude. Tools like fenbs can supply shared project rules and progress explicitly, while the app’s own instructions guide how the task handles them.

The practical question is whether this task received this instruction. A project can contain a perfectly good file that the task never opens. Before asking Claude to organize important files or prepare a client deliverable, check the instruction settings, the connected folder, and a small sample of its behavior.

Which instruction mechanisms does Cowork document?

Anthropic’s Cowork setup guide↗ (opens in a new tab) describes global instructions and folder instructions. Its Cowork project guide↗ (opens in a new tab) also describes project instructions. Use the settings and project controls available in your app; labels and layouts can differ between versions.

The table is a setup checklist, not a claim that all mechanisms resolve conflicts in the same way. Avoid conflicting text in the first place. If a general preference says to summarize and a project requires complete quotations, make the task’s expected output clear before work starts.

Why is a familiar filename not sufficient?

Claude Code documents CLAUDE.md as a persistent instruction file in its memory documentation↗ (opens in a new tab). That is evidence about Claude Code. It does not establish that every task in every Claude interface searches the same directories or applies the same scope rules.

A file can also enter a task in several different ways: you attach it, select a folder containing it, ask Claude to open it, or configure a product feature that supplies its contents. Those routes are not interchangeable. Successful use after an explicit read does not prove automatic discovery on the next task.

If you already maintain a repository instruction file, the CLAUDE.md guide explains that file’s role. For Cowork, begin with the instructions visible in your current setup and test any additional assumption separately.

How can you check what a task actually uses?

1. Open the intended project or task and inspect its instruction settings. Confirm that the instructions belong to this project rather than another workspace.
2. Connect only the folder or references needed for a small test. Use harmless sample material instead of sensitive client work.
3. Ask Claude to read the named instruction file before acting and identify the relevant requirements. Have it report inaccessible files instead of guessing their contents.
4. Request a tiny output that visibly exercises one requirement, such as a specific heading structure or a required approval note.
5. Start a separate task to test persistence. If you want to test automatic file discovery, do not also tell that task to open the file; otherwise the result cannot distinguish the two routes.

Do not rely solely on “yes, I read it.” Compare the response with a distinctive, harmless instruction whose effect you can inspect. A rule that says “write clearly” is difficult to test. A requirement to separate confirmed facts from open questions gives you something concrete to review.

Record a useful result precisely: “This task read the named file and produced the required sections.” That is stronger than “Cowork supports our setup,” which hides the conditions. For teams, fenbs AI context provides a place to keep the verified workflow and the current work it serves.

What should you write in a small project instruction?

State the purpose, the required sources, the output, and the point where a person must review the result. Keep changing task progress outside that instruction. A weekly reporting project needs a stable definition of a report, while this week’s missing spreadsheet belongs to the current task.

Purpose: prepare an internal operations summary.
Read the source files named in the task before drafting.
Separate confirmed facts, calculations, and open questions.
Do not invent missing values or approvals.
Save a new draft; preserve the source files.
Report the sources used and checks performed.
Ask a person to approve the draft before it is shared.

Adapt the example to the actual work and access you grant. Instructions describe the behavior you want; they do not replace connector permissions, file access controls, or review of consequential actions. A sentence asking for approval is useful guidance, but you should also configure the available permission controls appropriately.

For more task ideas, see Cowork examples. Keep the setup test small even when the eventual job includes many documents or tools.

What usually causes an instruction mismatch?

First check the ordinary causes: the task is in a different project, the folder is not connected, the named file moved, or a copied instruction is out of date. Then inspect whether the instruction is specific enough. A broad request for “the usual report” can be interpreted several reasonable ways.

Next, look for conflicting copies. If the project settings require one format and a reference file describes another, ask a person which is current. Update the authoritative instruction and remove obsolete guidance where appropriate. Repeating the newer rule in several more places only makes the next conflict harder to resolve.

Different interfaces also deserve separate checks. Our Cowork and Claude Code comparison covers their broader use cases. A successful repository workflow should be tested again when you move the same work into a different task environment.

How to do this with fenbs

1. Create a free account and add one active project.
2. Record the decisions people have approved in Decisions and rules. Add current tasks and the next step rather than copying complete conversations.
3. Use Integrations to connect a supported Claude interface or another supported assistant. Check the setup for the interface you use instead of assuming every mode exposes the same tools.
4. Ask the connected assistant to read the project context and explain the relevant rules before producing a small test deliverable.
5. Save the outcome and Where we left off, including any missing file access or unfinished review. Use that record at the next session.

fenbs does not make Cowork load CLAUDE.md and does not replace Cowork’s settings. If all you need is a personal formatting preference in one app, its built-in instructions may be sufficient. fenbs helps when the project’s decisions and progress must remain usable across supported assistant connections.

Keep the final check practical: can the next task identify the approved rules, read the necessary sources, and say what remains uncertain? The AI prompts hub collects more guidance on instructions that support that workflow.