06 · Configuration & extensibility

How you teach each product about your organisation. Cowork's Projects vs Claude Code's CLAUDE.md and .claude/ — plus the shared skills / plugins / MCP layer, and the one gotcha that catches every developer.

The gotcha, first

Verbatim — Cowork overview

"Cowork loads the connectors, skills, and plugins enabled for your claude.ai account, synced at session start, and doesn't read the Claude Code CLI's ~/.claude directory on your machine. To use a skill or plugin that exists only in ~/.claude, add it in Customize."

Consequences to state explicitly in class:

• Cowork configuration is account-scoped and cloud-synced; Claude Code configuration is machine- and repo-scoped.
• A skill written for one does not automatically appear in the other. Sharing means publishing it as a plugin (or re-adding it via Customize).
• Cowork picks up changes at session start — editing a skill mid-session won't take effect until you start a new one. The same is true of Claude Code's CLAUDE.md within a session.

Cowork: the Project as the configuration unit

A Cowork project "collects everything Claude needs for a recurring area of work." Contents, verbatim from the docs:

Creating one

Three starting points: Start from scratch (empty project, new folder), Import a project (bring in an existing claude.ai project), or Use an existing folder.

Cowork project vs claude.ai project — a distinction people conflate

You can link a claude.ai project into a Cowork project so sessions draw on its knowledge. "Linking doesn't merge them; the claude.ai project stays where it is."

Other documented behaviours

• Files Claude creates land in the project's folders; what Claude learns is saved to project memory.
• Dragging files in copies them into the project's first folder; dragging folders in mounts them as additional project folders.
• Claude reads individual files up to 50 MB.
• Archiving removes the project and deletes its metadata (name, instructions, links, memory) but "does not touch the local folders you attached; your files stay exactly where they are on disk."
• Dispatch can route background tasks into a project so long-running work inherits the same folders, instructions and memory.

Claude Code: CLAUDE.md and the .claude/ directory

Where CLAUDE.md lives, in load order

How they load

• Claude Code walks up the directory tree from the working directory, loading every CLAUDE.md and CLAUDE.local.md it finds.
• All discovered files are concatenated, not overridden. Root-down ordering, so instructions closest to where you launched are read last. Within a directory, CLAUDE.local.md comes after CLAUDE.md.
• Files in subdirectories load on demand, when Claude reads files there.
• @path/to/file imports are expanded at launch, up to four hops deep. External imports (resolving outside the working directory) trigger a one-time approval dialog.
• Managed-policy CLAUDE.md cannot be excluded. Everything else can, via claudeMdExcludes globs.

The size discipline Anthropic prescribes

Rules — path-scoped instructions

.claude/rules/*.md files load every session, unless they carry paths: frontmatter, in which case they load only when Claude touches matching files:

---
paths:
  - "src/api/**/*.ts"
---

# API Development Rules
- All API endpoints must include input validation
- Use the standard error response format

Rules support glob patterns, brace expansion (budget: 1,000 expanded patterns / 4 MiB per rule), symlinks for sharing across projects, and a user-level variant at ~/.claude/rules/.

Auto memory

• On by default. Toggle in /memory, or autoMemoryEnabled: false, or CLAUDE_CODE_DISABLE_AUTO_MEMORY=1.
• Stored per repository at ~/.claude/projects/<project>/memory/ — shared across worktrees, machine-local, never synced across machines.
• MEMORY.md is a concise index; only its first 200 lines or 25 KB load at session start. Topic files load on demand.
• Plain markdown — auditable, editable, deletable.

AGENTS.md interoperability

Claude Code reads CLAUDE.md, not AGENTS.md. If a repo already has one, create a CLAUDE.md that imports it:

@AGENTS.md

## Claude Code
Use plan mode for changes under `src/billing/`.

/init also reads Cursor rules, Copilot instructions, and (with CLAUDE_CODE_NEW_INIT=1) AGENTS.md, Devin, Windsurf and Cline rules.

The shared extension layer

Skills

Markdown files containing knowledge, workflows or instructions. Invocable as /name, or loaded automatically by the model when relevant. Both products support them.

• Cowork Managed through Customize in the sidebar; account-scoped.
• Claude Code Discovered from managed / user / project / plugin scopes with documented precedence (managed > user > project). Can run in isolated context (context: fork), be hidden from the model (disable-model-invocation: true), or be preloaded into a subagent.

Plugins — the interoperability layer

Verbatim — Cowork plugins guide

"Plugins are available in Cowork and Code. They aren't used in Chat."

A plugin manifest can declare any combination of:

Distribution

• Anthropic's official catalogue is the default marketplace; you can add others by URL.
• A git repository can serve as a marketplace — GitHub (including Enterprise) supported; public GitLab and Bitbucket repos also work. Accepts https://github.com/owner/repo or the owner/repo shorthand.
• Install from a file by uploading the package.
• On Team and Enterprise, admins can mark plugins Required — they install automatically, show "This plugin is required by your organization," and cannot be removed by the user.
• Cowork detects locally edited plugin files and warns before an update would overwrite them.

Documented limits (both products)

MCP / connectors

Model Context Protocol is the open standard connecting Claude to external systems — Google Drive, Jira, Slack, Linear, Notion, databases, internal APIs. Both products consume it.

• Cowork Connectors managed under Customize; loaded from the claude.ai account at session start. Local MCP servers require the desktop app open. Admin setting: "Allow 'Always allow' for connector tools" (off by default) controls whether users can skip per-task approval for write-capable connectors.
• Claude Code Scope hierarchy local > project > user; tool search defers schemas so idle tools cost almost no context; automatic reconnection; claude mcp login; managed allowlists/denylists via managed-mcp.

Hooks — the asymmetry

Hooks are Claude Code's enforcement primitive, and the documentation is unusually direct about why they matter:

Verbatim — Extend Claude Code

"Put guardrails in hooks. An instruction like 'never edit .env' in CLAUDE.md or a skill is a request, not a guarantee. A PreToolUse hook that blocks the edit is enforcement. If a rule must hold every time, make it a hook rather than a prompt instruction."

In Cowork, hooks exist only as a component that an installed plugin can provide. There is no documented user-authored hook workflow. This is the sharpest governance gap between the two products and is worth its own slide for any risk or compliance stakeholder.

Choosing the right mechanism — Anthropic's own guidance

Claude Code's docs publish a "build your setup over time" table that transfers well to Cowork with light translation:

Context cost — a concept worth teaching in both products