Cowork vs Chat — The Core Distinction
| Feature | Standard Chat | Cowork |
|---|---|---|
| Interaction model | Prompt-response cycle | Agentic task execution |
| File access | Manual upload (30MB limit, 20 files) | Direct local file system access |
| Output delivery | Download links in conversation | Written directly to your file system |
| Execution | Single response per prompt | Multi-step pipeline with plan review |
| Parallelism | None | Sub-agents process batches concurrently |
| Scheduling | Not available | Recurring tasks via /schedule |
| Persistence | Stateless between sessions | Projects with memory and context files |
Decision rule: If you are describing a desired outcome involving files, use Cowork. If you are asking a quick question, use Chat.
The Five-Stage Execution Pipeline
- Analyse — Claude examines your request and the available resources in the working folder
- Plan — A visible execution plan appears showing every intended action
- Decompose — Complex work is broken into subtasks (potentially parallel)
- Execute — Work runs in the sandboxed VM, with sub-agents if applicable
- Deliver — Outputs are written directly to your file system
Common trap: Closing the desktop app or sleeping the computer stops execution immediately. There is no cloud fallback — the VM runs locally.
The Sandbox & Security Model
- Folder scoping: Claude can only access the folder you explicitly grant — cannot traverse upward or reach other directories
- Read/write: Free within the granted folder — no permission prompt needed
- Deletion: Requires explicit "Allow" click via system-level prompt
- Overwrites and renames: No permission prompt — deletion protection does not cover these
- Computer Use: Operates outside the VM on your actual desktop — separate per-app permissions
- Network egress: Configurable allowlist, but web search bypasses egress restrictions
The hierarchy of protection (know this cold):
| Layer | What It Controls |
|---|---|
| Folder scoping | What Claude can see |
| Deletion protection | What Claude can permanently remove |
| Plan review | What Claude intends to do (your safety net) |
| Computer Use permissions | Which desktop apps Claude can interact with |
Sub-Agents & Parallel Execution
- Trigger phrases: "for each one," "analyse all," "process these," "in parallel"
- No special syntax needed — sub-agents trigger automatically from natural language
- Independent tasks only — parallelism fails when Step 2 needs Step 1's output
- Token cost increases — each sub-agent has its own reasoning stream (more tokens, not fewer)
- All local — sub-agents run in the VM on your device, not in the cloud
- Error multiplication — a wrong naming pattern gets applied to all files simultaneously
When to parallelise: Batch file processing, multi-document research, data extraction from many sources. When NOT to parallelise: Tasks with sequential dependencies, iterative refinement, or outputs that build on each other.
Effective Task Delegation
| Instead of... | Do this... |
|---|---|
| Step-by-step micromanaging | Describe the desired end-state |
| "Fix my files" (too vague) | Specify naming conventions, folder structure, output format |
| Repeating context every session | Use context files (about-me.md, brand-voice.md) |
| Using Opus for everything | Match model to task: Haiku for simple, Sonnet for standard, Opus for deep reasoning |
| Skipping plan review | Always read the execution plan before clicking Allow |
The golden rule: You describe what "done" looks like. Claude figures out how to get there.
Common Exam Traps — Domain 1
| Trap | Correct Answer |
|---|---|
| "Cowork processes files in the cloud" | Everything runs locally in a sandboxed VM on your device |
| "Tasks continue when you close the laptop" | The computer must stay awake and the app must remain open |
| "Deletion protection covers all file changes" | It only covers permanent deletion — edits and overwrites have no prompt |
| "Sub-agents reduce token consumption" | They consume more tokens due to separate reasoning streams |
| "You need /parallel syntax to trigger sub-agents" | Natural language triggers them automatically |
| "Standalone sessions remember previous sessions" | Memory requires Projects; standalone sessions start fresh |
| "Cowork activity appears in Enterprise audit logs" | Cowork is excluded from audit logs on all plan tiers |