The same task in Chat vs Cowork

The clearest way to understand Cowork is to watch the same task done twice — once in Chat, once in Cowork. Expense reports from receipt photos is the canonical demo. Anthropic uses it. Every power user who teaches Cowork uses it. Use it yourself in Module 1's exercise.

The same task, two surfaces

In Chat: You upload one receipt photo. Ask Claude to extract vendor name, date, amount, and category. Claude gives you a paragraph or a small table. You copy it into a spreadsheet. You upload the next receipt. Repeat 49 more times. You are the orchestrator. Claude is answering questions.

In Cowork: You mount the folder containing all 50 receipts. You type one prompt:

"Process every receipt image in this folder. For each one, extract vendor name, date, amount, and category. Create an Excel spreadsheet called expense-report.xlsx with all receipts in one table, sorted by date, with a total row and category subtotals at the bottom."

Cowork reads all 50 receipts. It creates the spreadsheet. The spreadsheet has working formulas. You open it, review it, and forward it. Total prompts: one.

This is the difference. Chat produces text you act on. Cowork produces deliverables you ship.

The #1 mistake: over-specifying steps

Here's where most new Cowork users go wrong. They understand that Cowork is more powerful, so they try to control it more tightly. They specify every step. And that's exactly backwards.

Compare these two prompts for matching a new presentation to a brand template:

Bad:

"Open the folder, find the .pptx file, extract the color hex codes using python-pptx, create a new presentation object, apply those colors to the slide master, then recreate each slide using the extracted layout structure, and make sure the font sizes match the original..."

Good:

"Build a presentation that matches the design of abp-presentation-template-light.pptx in the same folder."

The bad prompt forces Claude into a specific path — your path. If that path has a flaw, Claude will follow it off the cliff. The good prompt describes the outcome and names the reference file. Claude picks the approach. It might use python-pptx, or it might use a different method that works better. It doesn't matter — you get the output you described.

The principle: when you over-specify steps, you hobble the agent's ability to pick the right tools. Describe the outcome and the success criteria. Let Claude figure out the path.

This is the hardest mental shift for people coming from Chat. In Chat, more detail usually helps. In Cowork, too much procedural detail gets in the way.

A preview of the task loop

When you submit a task, Cowork doesn't immediately start executing. It shows you a plan first. You'll see something like:

1. Mount and scan the folder contents
2. Read each receipt image
3. Extract the required fields
4. Build the spreadsheet with the specified structure

You can approve this plan as-is, or steer it. "Merge the restaurant and coffee categories into 'Meals.'" "Don't process anything in the archive/ subfolder." Once you approve, Cowork runs — and you can still steer mid-execution if something looks off.

This plan-before-execute pattern is one of Cowork's safety features and one of its biggest advantages over Chat. You catch misunderstandings before they become wasted work. Module 3 covers the full task loop in detail.