Claude Cowork — Quick Guide

#claude #cowork #agentic #automation #delegation

Claude Cowork

Delegation over Conversation — describe outcome, get finished file.

Cowork is the same agent as Claude Code, built for knowledge workers who don't use command line. Instead of chat replies, it produces real files (.pptx, .docx, .xlsx) saved directly to your drive.

Chat vs Cowork — When to Switch

Use Chat when	Use Cowork when
Quick Q&A, drafting text	Task needs to read/write actual files
One-turn exchange	Multi-step, runs 5–30 min
No file output needed	Output is a real file on your machine
	Connects to Gmail, Slack, Drive, browser

How to Implement

Open Claude Desktop → click the Cowork tab
Click "Work in a folder" → select your working folder → "Always allow"
Write your brief using the I-T-O formula
Answer Claude's clarifying questions → approve the plan → let it run
Monitor the progress panel — steer mid-run if needed
Open the finished file, run your 3-step review

The I-T-O Prompt Formula

Every Cowork brief needs 3 things:

I — Input: Where is the source data? (folder, file name, Slack channel)
T — Transformation: What should Claude do with it?
O — Output: What does "done" look like? (format, filename, save location)

❌  "Clean up my files"

✅  Input:          My /Downloads folder
    Transformation: Sort files by type into dated subfolders
    Output:         Organized subfolders, saved in /Downloads

Key rule: Describe the outcome — don't list steps. When you micromanage steps, you block Claude's planning ability and lose the leverage you're paying for.

Scheduled Tasks — Full Automation

Combine Skills + Schedules to build a background automation engine:

Type /schedule in any Cowork chat
Set cadence: hourly / daily / weekly / monthly
Define folder, output format, approval gate

Example:
"Every Friday 4pm: pull Slack #sales messages from the week,
extract key deal updates, draft summary email to leadership.
Draft only — don't send."

Critical limitation: Cowork only runs when your computer is on and the Claude Desktop app is open. Laptop asleep = task doesn't run.

Effectiveness

You stay in control at 3 points: review the plan before, steer during, review output after
Always default to "Draft only — don't send" for emails and actions with real-world impact
3-step review habit after every run: (1) open the file, (2) check one number, (3) follow one logic thread
Output looks polished even when wrong — polished ≠ correct

Source: Claude Course 2026 — NotebookLM