The task loop: Describe → Plan → Execute → Review

Every Cowork session follows the same four-stage loop. Once you see it clearly, you stop being surprised by what Cowork does — and you get much better at steering it.

The stages are: Describe Goal → Claude Plans → Claude Executes → Finished Files. Simple to say. The leverage is in understanding what happens at each stage and where your intervention actually matters.

Stage 1: Describe goal

You write the task. The quality of your description shapes everything that follows, which is why Module 3's remaining lessons focus entirely on how to write it. For now: your job at this stage is to hand off an outcome, not a procedure.

Stage 2: Claude plans

This is the most important stage, and the one most people rush past. Cowork doesn't start executing immediately. It analyzes your request, decomposes it into subtasks, and presents a plan before touching any files.

This is where you steer. Read the plan. Does it understand the scope correctly? Is it working in the right folder? Does it have the right output format in mind? If something is off, correct it here — before a single file is modified. A thirty-second review at the plan stage saves thirty minutes of unwanted output.

If the plan looks right, approve it. Then get out of the way.

Stage 3: Claude executes

Four things happen under the hood while Cowork works:

Plans into subtasks. Cowork breaks your goal into discrete steps and executes them in sequence — or in parallel, when steps are independent of each other.

Runs code in an isolated VM. Shell commands and Python scripts execute inside a virtual machine on your machine. The code is isolated from your operating system. But file operations — reads and writes — happen on your real files in the mounted folder. The VM is a safety boundary for code; it is not a sandbox for files.

Coordinates sub-agents. For complex work with independent workstreams — fifty PDFs to process, ten competitor sites to research — Cowork spawns parallel sub-agents. You'll see multiple workstreams reporting progress simultaneously. You don't trigger this manually; Claude decides when it's useful.

Streams progress. You see everything as it happens: reasoning, file reads, tool calls, intermediate results. This isn't decorative. It's your window to catch a wrong turn early. If you see Cowork working in the wrong folder or processing the wrong file type, you can intervene without waiting for it to finish.

Stage 4: Finished files

Cowork deposits outputs into your designated folder — usually out/. Open them, review them, decide what to iterate on. If something needs adjustment, you have two choices: continue in the same session with a correction, or clear the session and start fresh with a refined prompt. The next lesson covers when to do which.

Safety behaviors built into the loop

Three behaviors are worth knowing before you run your first real task.

Deletion protection. Cowork always requires explicit, per-operation permission before permanently deleting any file. It will not silently delete things. If a task requires deletion, it will ask.

Code execution isolation. All code runs inside the VM. A script that crashes, loops, or behaves unexpectedly doesn't touch your operating system — only the files in your mounted folder.

Permission prompts for connectors and Computer Use. When Cowork needs to connect to an external service or take over a browser, it asks permission scoped to that specific app or site. It cannot silently access services you haven't approved.