Folder Architecture — Your Operating System

Why Folders Are Not Just Organization

Four independent sources — Paul, Brooke Wright, Jack Roberts, and Jenny at Anthropic — have each said the same thing in different ways: the single biggest predictor of Co-Work output quality is folder structure. Not the model. Not the prompt. Not the connectors. The folder you are working in when you run the task.

Here is why. When you activate a folder in Co-Work, it reads the claude.md and memory files stored in that folder before it does anything else. Those files define the role Co-Work is playing, the output format it should use, and the constraints it should follow — all before you type a single word. If the folder has no claude.md, Co-Work arrives blank. If it has a good one, Co-Work arrives pre-briefed.

Most people using Co-Work skip this entirely. They point Co-Work at an existing folder — often their Documents folder or their Desktop — and run tasks. They get generic, unfocused output. They blame the model. The actual problem is that they gave Co-Work no context about what it is supposed to be doing in that folder.

The Three Levels of Instructions

Co-Work instructions work in three layers, and understanding which layer serves which purpose is essential for building a setup that actually scales.

Layer 1: Global Instructions

These apply to every session, every folder, every task. They are the constants: your identity, your safety guardrails, your universal output preferences. Set them up in Module 2 and rarely revisit them. Adding task-specific or folder-specific content here is a mistake — it pollutes every task with context that is only relevant sometimes.

Layer 2: Folder-Level claude.md

This is the layer most people miss. Each subfolder in your Co-Work structure should have its own claude.md file defining the role Co-Work plays when it is working in that folder. A claude.md in your Content folder tells Co-Work it is a content strategist. The same file in your Operations folder tells Co-Work it is an operations coordinator. Same model, same session — entirely different context and output quality.

Layer 3: Project Instructions

These apply within Co-Work Projects — a more structured container covered in Module 9. For now, understand that project instructions are the most specific layer: they override folder instructions for a specific, time-bounded project. Most Tier A users will not need this until Module 9.

Building the Parent/Subfolder Hierarchy

You already created the root folder ~/CoWork/ in Module 2. Now you are going to build the subfolder architecture that makes it useful. Each subfolder represents one distinct domain of work — one set of context, one role, one output format.

Here are four subfolders that cover most knowledge work scenarios. Adapt them to your actual work:

• ~/CoWork/Content/ — writing, publishing, content creation, social media drafts, newsletters
• ~/CoWork/Operations/ — email triage, scheduling, admin tasks, file organization
• ~/CoWork/Finance/ — invoices, expense review, financial summaries (note: do not send regulated financial data through Co-Work until you have read Module 16)
• ~/CoWork/Research/ — user research, competitive analysis, briefings, summaries

The rule for deciding when to create a new subfolder is simple: if a task could reasonably belong in two existing folders, you need a third folder. Ambiguity in folder assignment means ambiguous context for Co-Work.

Writing an Effective claude.md

A good claude.md file is short, specific, and actionable. The goal is to give Co-Work exactly what it needs to play its role in this folder — not to document everything about your business (that is what the Business Brain is for).

Here is an example for a Content subfolder:

You are a content strategist for [Business Name].

When working in this folder, your output should match the brand voice
described in ~/CoWork/Business/business-brain.md.

Output format: Markdown unless otherwise specified. Always include a
one-paragraph summary at the top of longer documents.

Default tone: conversational and direct. Avoid corporate jargon.

Constraints:
- Never publish directly to any platform. Always save draft output to
  the Output/ subfolder for review.
- Flag any claim that needs a source with [NEEDS SOURCE] inline.
- If asked to write for a specific audience, confirm the audience before
  starting if not already specified.

Notice what this does not include: safety guardrails (those are global), full business context (that is in the Business Brain), or platform-specific instructions for tasks you have not defined yet. Keep folder files focused on the role, the format, and the constraints specific to that folder's domain.

Length guideline: 100 to 300 words per claude.md. Longer is not better. If you are writing more than 300 words, you are probably including content that belongs elsewhere.

Memory Files: Persistent Context Across Sessions

Co-Work's memory is folder-scoped. When you tell Co-Work to remember something while working in a folder, it saves that information to a memory file in that folder. When you return to the same folder in a new session, Co-Work reads the memory file and picks up where it left off.

Create a memory.md file in each subfolder. It can start almost empty — just a seed line to establish context:

# Memory — Content Folder
Created: [date]. Purpose: Content creation and brand communication.

## Running notes
(Co-Work will add to this section as you work together)

To save something to memory: "Save this to memory: our newsletter goes out every Tuesday." Co-Work writes it to the memory file. To review what is stored: "What do you know about our content workflow from memory?"

Memory accumulates over time. After a few months of active use, your memory files will contain a dense record of preferences, decisions, and context that makes Co-Work dramatically more accurate — without you re-entering any of it.

The Builder Skill: Let Co-Work Create Subfolders for You

Once you have the basic hierarchy in place, you can use Co-Work to create new subfolders. This is a community practice shared by Paul — not an official feature, but a useful pattern that eliminates the friction of manual setup.