Claude Cowork Setup:
3 Mistakes That Waste Your Tokens

Last updated September 18, 2026 to reflect Anthropic's unified Claude experience, current instruction settings, and cloud scheduled tasks.

Last week I read a beginner guide to building a Claude Cowork setup in 10 minutes. It recommended keeping the main instruction file short. I thought mine was probably fine.

It was not fine.

I opened the CLAUDE.md file that ran my system for the first time in four months. Every folder path in it was wrong. The file was 306 lines long, it pointed Claude toward voice files it could not reach, and one instruction would have loaded roughly 55,000 tokens before almost every writing task if the path had worked.

My setup is not small. It spans four work areas, about 20 custom skills, and 6,634 files in one content folder. That is exactly why the failure stayed quiet. The larger a system gets, the easier it becomes to assume the map still matches the ground.

The audit reduced the root file from 306 lines to 248 and verified 16 out of 16 paths. The three mistakes were simple: a bloated root file, the same rule stored in two places, and no maintenance ritual.

What is a Claude Cowork setup?

A Claude Cowork setup is the combination of account-wide instructions, project or folder instructions, connected tools, skills, and recurring tasks that tell Claude how to work with you. In a local folder workflow, a CLAUDE.md file can act as the project brief Claude reads before doing the work.

Anthropic is currently rolling Cowork and chat into one Claude experience. Some accounts still show a Cowork tab. Others start every conversation from the same message box. The structure still matters:

• Instructions for Claude apply across conversations. In the new experience, they live under Settings, then General.
• Folder or project instructions contain the context that belongs only to that body of work.
• Skills hold repeatable methods, examples, and reference material for a specific task.
• Scheduled tasks run a saved instruction on a cadence. Cloud tasks can run while your computer is off, but tasks that need local files still require access through Claude Desktop.

Think of account-wide instructions as the employment contract and folder instructions as the project brief. A rule belongs at the highest level where it is always true, and nowhere else.

Why do broken paths in CLAUDE.md fail silently?

A broken path often fails without a visible warning because Claude treats the instruction file as a map, then tries to complete the task with whatever it can reach. It may skip the missing reference, infer the intended behavior, or create output in a new location.

I found three versions of that failure:

The wrong root

My file said the brisk. brand folder lived at one path. The real folder lived several levels deeper. The path had been wrong since the day I wrote it, but no task stopped to announce the mismatch.

The empty voice folder

The instruction said to read my voice profile before any brisk. writing. The named folder existed, but it contained fonts. My actual voice files lived somewhere else. For four months I assumed every draft used those rules. It did not.

The phantom folder

One session followed the broken map and created a folder where the instruction said it should exist. That left four orphaned carousel drafts in a second content folder while the real folder held 6,634 files.

Add this guardrail to every folder map:

If a path in this file does not exist, stop and tell me.
Do not create a replacement folder unless I approve it.

That two-line rule turns a quiet failure into a visible decision.

Mistake 1: Why does a bloated root file waste context?

A bloated root file makes Claude carry instructions that do not apply to the current job. The problem is not only token cost. Irrelevant rules compete with the instructions that actually matter.

My 306-line root file included about 60 lines of coding rules. I use Cowork mainly for newsletters, carousels, research, and client work. Claude was reading rules about refactoring functions before helping with a caption.

The fix took two minutes. I moved the coding rules into their own file and replaced them with one conditional pointer:

For coding tasks only, read: [verified path]/coding-rules.md
Do not load this file for writing, research, or planning tasks.

Use the root file for information that is both stable and broadly relevant:

• who you are and what you do
• the folder map
• non-negotiable safety rules
• how Claude should communicate with you
• pointers to specialized skills or reference files

If a rule matters only for one kind of task, move it closer to that task.

Mistake 2: Why should each rule have one source of truth?

Duplicated instructions drift because people update one copy and forget the other. Claude then receives two versions of the same rule and has to guess which one is current.

My voice rules appeared inside CLAUDE.md, while the same file also told Claude to read a separate voice profile. Same rules, two homes.

The fix is simple: choose one source of truth. Everywhere else gets a pointer, not a copy.

Before:

Voice: read two large reference files and apply them to all brisk. content.
[The root file also repeats a shorter version of those voice rules.]

After:

The content-writing skill is the source of truth for brisk. voice.
Use it for brisk. writing tasks. Do not restate its rules here.

Open the full voice reference only when I ask for a voice audit.

This makes the system easier to maintain and gives Claude a clear precedence rule.

Mistake 3: How do large reference files burn tokens?

Large reference files burn tokens when an instruction tells Claude to load them before every task, even when the task needs only a small rule set. The expensive part is the unconditional loading rule.

My two voice references totaled 223KB of plain text. I estimated that at roughly 55,000 tokens, although the real count depends on the contents and tokenizer. The instruction said to read both files for any brisk. content task.

That meant every LinkedIn hook, Instagram caption, and newsletter draft could start by consuming a large part of the available context. The broken path was accidentally protecting my usage.

Do not solve this by deleting useful reference material. Separate the layers:

• Keep the small operating rules inside the skill or project instructions.
• Keep examples and deep reference material in separate files.
• Load the deep files only for audits, edge cases, or high-stakes work.
• Name the condition that triggers each file.

The brisk. rule is: small instructions run the work; large references check the work.

How can you audit a Claude Cowork setup in two minutes?

Audit the map, the duplicates, and the loading conditions. You do not need a technical tool to find most setup failures.

1. Open your account-wide instructions and the instruction file for your main folder or project.
2. List every path referenced in those instructions.
3. Check that every path exists and points to the intended file.
4. Search for rules repeated in more than one place.
5. Find every instruction that says "always read," "load first," or "apply to every task."
6. Move task-specific rules out of the root file.
7. Add a date and a monthly review reminder.

You can also give Claude the folder list and instruction file, then use this prompt:

Audit this instruction file against the folder list below.

1. Check whether every referenced path exists.
2. Flag rules that appear in more than one place.
3. Flag files that are loaded for every task but should be conditional.
4. Do not edit or create anything yet.
5. Return a table with: issue, exact line, risk, recommended fix.

[paste the instruction file]
[paste the folder list]

Review the table before allowing changes. An audit should expose the system before it rewrites the system.

How should a beginner build a clean Claude setup?

Start with one project, two instruction layers, and one recurring audit. A small setup that you understand is more useful than a complex system copied from somebody else.

1. Choose one project or trusted folder

Give Claude access only to the files required for that work. Anthropic warns that Cowork can act on files and connected services, so begin with a contained, low-risk workspace.

2. Write account-wide instructions

Add information that should apply everywhere: your role, preferred tone, output format, and safety boundaries. In the unified Claude experience, use Settings, then General, then Instructions for Claude. Accounts still showing Cowork may see Global instructions under Cowork settings.

3. Write a short project brief

Keep the root CLAUDE.md or folder instructions focused:

# Project instructions

## Purpose
[What this project is for and who the work serves.]

## Folder map
[Exact verified paths for the files Claude may need.]

## How to work
Be direct. Ask when the brief is unclear. Do not invent facts.
If a referenced path does not exist, stop and tell me.

## Definition of done
[What a completed task must include.]

Twenty correct lines beat 300 stale ones.

4. Add skills only for repeated work

A skill should solve one repeatable task, such as writing a newsletter, cleaning a spreadsheet, or preparing a client brief. Anthropic's current guidance says the best skills stay focused and include clear instructions and examples where helpful. The brisk. skills directory shows practical examples, and this guide explains three ways to build a Claude skill.

5. Connect tools when a real task needs them

Start with one or two connectors. Add another when you can name the job it will do. Eight unused connectors create more permissions and more surface area to maintain.

6. Use project memory deliberately

Claude Projects now support their own memory, instructions, files, and scheduled tasks. Use project memory for context that should survive across related tasks. Keep durable operating rules in instructions so you can inspect and edit them directly.

7. Schedule one low-risk task

Start with a summary or report that you can review. Cloud scheduled tasks can run with no device online. If the task depends on a local folder, browser, or desktop app, Claude Desktop must be able to reach that resource.

8. Put the audit on the calendar

Once a month, verify paths, remove stale rules, review permissions, and pause scheduled tasks you no longer use. The setup takes a morning. Maintenance keeps it trustworthy.

What can a Claude Cowork audit still miss?

A manual audit cannot prove that every instruction produces the intended behavior. It can verify structure, find broken paths, and expose conflicts, but you still need to test the system with real tasks.

Check these limits:

• A valid path can still point to the wrong file.
• Two rules can conflict without using the same words.
• A shorter file can still contain irrelevant instructions.
• Cloud tasks may lose access to local resources when Claude Desktop is closed.
• Connectors and browser actions can expose sensitive data or follow malicious instructions.

After the structural audit, run one representative task from each workflow. Check what Claude opened, what it changed, and where it saved the result.

Sources and further reading

• Anthropic: Get started with Claude Cowork, current setup, instructions, projects, and scheduling
• Anthropic: Claude Cowork and chat are one Claude, the September 2026 interface change
• Anthropic: Schedule recurring tasks in Claude Cowork, cloud and local task behavior
• Anthropic: Use Claude Cowork safely, permissions, prompt injection, and scheduled-task precautions
• Anthropic: How to create custom skills, skill structure and scope

Open your main instruction file. Check the date, then check every path. That is the brisk. maintenance rule: if Claude depends on the map, verify the map before adding more automation.