The Claude Cowork Playbook

Claude Chat vs Claude Cowork: the three differences that matter

Difference 1: File limits

Difference 2: Output format

Difference 3: How to prompt it

Context window

Cowork has a significantly larger context window than Chat. You can do substantially more work before hitting the compacting function - where Claude summarises previous context to free up space and potentially loses details in the process.

First-time setup

Step 1: Settings - General tab

Step 2: Settings - Cowork tab

Step 3: Settings - Capabilities

Step 4: Create a working folder

What it does

Cowork can create, edit, move, and organise files directly on your machine. No copy-pasting outputs. The file appears in your folder.

Use cases

Expense report from 100+ receipts

Split a 400MB+ PDF into chapters

Convert static image slides into editable PowerPoint

What it does

Cowork saves memory to actual files on your machine (CLAUDE.md and memory.md). It can remember every decision, preference, and correction you give it indefinitely - with no hard limit on how much it holds.

Chat memory is stored online and has a hard cap on how much it can retain.

How the memory files work

Whenever you tell Cowork to remember something, it writes to CLAUDE.md and memory.md in your working folder. The more it writes to those files over time, the more accurately it works the way you want.

Keep these two files in your folder. Do not delete them.

How to build memory in practice

Teaching Cowork your preferences step by step

What Cowork can recall across sessions

What it does

Connectors let Cowork reach directly into external tools - Gmail, Google Drive, Google Calendar, Notion - to read from them and work inside them. By default Cowork can only see your local folder.

How to set up connectors

→ Click the Customize tab on the left

→ Go to Connectors

→ Click the + icon

→ Browse and search for the tool you want

Minimum recommended connectors to set up: Gmail, Google Calendar, Google Drive, Notion.

If your tool is not listed, you can add a custom connector via MCP.

Use cases

Extract your tone of voice from your own emails

Cross-reference two tools to catch what fell through the cracks

What skills are

A skill is a saved, repeatable workflow. You build one by going through a task with Cowork - doing the back and forth, giving feedback, refining the output - then asking Cowork to turn that entire workflow into a skill it can run again from a single instruction.

A simple one-step skill is useful. But a 10-step workflow you run every week becomes a skill that saves hours over time.

How to create a skill

The right way to build a skill

Enabling skill creation

Example: weekly report skill

Scenario: Three teams send weekly updates in different formats. The task is combining them into one clean update for leadership.

Step 1: Do the task

Step 2: Refine it

Step 3: Create the skill

Updating and backing up skills

→ To update a skill: tell Cowork which skill to change and what to change. After it recreates the skill, click copy to overwrite the previous version.

→ Back up skill .md files to Google Drive. Skills do not transfer if you move to a new machine.

What it does

Cowork projects work the same as Claude Chat projects - a dedicated space with its own instructions, knowledge files, and context - but they come with all the capabilities already covered: local file access, persistent memory, connectors, and skills.

The one difference worth knowing

In a Chat project, when Claude learns something new about your style and you want to save it permanently, Claude can generate an updated knowledge file for you - but you still have to manually add it to the project and delete the old version yourself.

In a Cowork project, you just say: "Save this as a principle under the [project name] project." Cowork writes it directly to the instruction file. No manual steps.

When to use Cowork projects vs Chat projects

Use a Cowork project when the work involves files, external tools, repeatable workflows, or anything where you want memory that improves automatically over time.

Use a Chat project when the work is purely conversational - thinking through ideas, reviewing text, asking questions - and does not require file output or tool access.

What it does

If you have the Claude browser extension installed, Cowork can hand off tasks to it - filling in forms, navigating pages, clicking through interfaces.

Current limitations - do not rely on it yet

Three practical issues make the browser extension unreliable right now:

Slow

Unreliable

High token burn

The related web search limitation

Both Chat and Cowork can search the web. But Chat gives you direct control - you can force enable or disable web search, and toggle it mid-conversation.

Cowork does not let you force a web search. When it needs web data, it often falls back on the browser extension. Given how unreliable the extension is right now, this is a real gap.

For research-heavy tasks, Chat currently has the edge.

What it does

Scheduled tasks let you set Cowork to run a workflow automatically at a set time - daily, weekly, or on a custom schedule. Gemini and ChatGPT have had this feature, but the reason Cowork's version works reliably where others do not is because of everything already covered: it has access to your local files, persistent memory of your preferences, and direct connections to your tools.

How to build a scheduled task that actually works

A scheduled task is only as good as the context it runs on. Three layers of setup make the difference between a useful output and a generic one:

Layer 1: Save your workflow as a rules document

Layer 2: Connect the tools it needs

Layer 3: Train it through feedback

Example: daily email triage at 6am

What the finished task does

Folder structure for this task

Setting up the scheduled task

→ Open Scheduled Tasks in Cowork

→ Click to create a new task

→ Write the task instruction using outcome-first language - describe the end result you want

→ Point it at the right folder and connectors

→ Set the schedule (daily, weekly, specific time)

→ Run it manually first to confirm the output is correct before activating the schedule

Using Prosp.ai with Claude to book 15 calls/week from LinkedIn