Claude Cowork Tutorial: Build a Personal AI Assistant

You don’t need to write code to have an AI agent working for you.

I wrote the complete Claude Code tutorial to help developers set up the most powerful coding agent available. But I kept hearing the same thing from non-developer friends: “I want this, but I don’t live in the terminal.” Marketers, ops people, founders, consultants. All watching Claude Code from the outside, wishing they could tap into the same agentic power.

Cowork is the answer. It brings the same agentic architecture to knowledge work instead of software development. You tell it what to do, it makes a plan, breaks the work into subtasks, uses tools, and delivers finished outputs such as reports, spreadsheets, presentations, and organized files.

And just like Claude Code changed how I write software, Cowork has changed how I handle everything else. Research, reports, file organization, recurring tasks, pulling data from tools. Stuff I used to spend hours on now takes minutes.

By the end of this guide, you’ll have a personal AI assistant that knows how you work, connects to your tools, runs tasks on a schedule, and follows you between desktop, web, and mobile.

Let’s set it up.

What is Cowork (and What Makes It Different from Chat)

Every breath you take, every file you make, I’ll be organizing you.

If you’ve used Claude’s chat interface, you know the flow. You type a question, Claude answers. You paste in some text, Claude summarizes it. It’s a conversation. You bring the inputs, Claude gives you outputs, and everything lives in that chat window.

Cowork flips that model. Instead of chatting with Claude, you’re delegating to it. You describe a task, Claude creates an execution plan, shows its progress, and then goes to work. It can use cloud files, connected services, or local folders you explicitly share through Claude Desktop. It spawns sub-agents to handle subtasks in parallel. When it is done, the finished output is attached to the session or saved into the working location you chose.

Here’s the key difference in practice:

	Chat	Cowork
You say	”Summarize this document” (paste text)	“Summarize every PDF in my /Reports folder and create a comparison spreadsheet”
Claude does	Returns text in the chat window	Reads 15 PDFs, spawns sub-agents in parallel, creates an Excel file with formulas
Output lives	In the conversation	In the session, as downloadable files, or in a connected local folder
Duration	Seconds	Minutes (but unattended)
Tool access	Chat tools and connectors	A task environment with files, connectors, scheduling, parallel work, and computer control

The agentic architecture underneath is the same thing that makes Claude Code so effective. Claude doesn’t just respond. It plans, decomposes complex tasks, uses tools, and iterates on its own work. The difference is that Cowork’s tools are built for knowledge work. Instead of running bash commands and editing code files, it’s creating spreadsheets, drafting presentations, pulling CRM data, and organizing your file system.

If you want to understand the deeper mechanics of how agents plan and execute, my guide to AI agents covers the architecture. But you don’t need to understand any of that to use Cowork. That’s the whole point.

Getting Started: Setup and Your First Task

Requirements

You need a paid Claude plan. Cowork is available on Pro, Max, Team, and Enterprise plans. Install Claude Desktop if you want Cowork to work with local folders, your browser, or desktop applications. For cloud-only tasks using uploaded files and connectors, you can also work from the web or mobile app.

Some features have narrower plan availability while they are in beta. Computer use, for example, is currently a research preview for Pro and Max users. Check Anthropic’s current plan page before upgrading for one specific feature.

Opening Cowork

Claude now presents Chat and Cowork from a shared home. Choose Cowork in the message box, or open Cowork from the Claude app navigation. On a new account, run /setup-cowork. Claude asks about your role, recommends useful connectors and plugins, and helps establish your working preferences.

The interface still feels conversational, but the unit of work is a task. Claude plans, uses tools, and shows its progress while it works. Your approval mode controls whether actions pause for review or continue automatically.

Folder Permissions

Local folders are optional, but they are the easiest way to let Cowork work with files already on your computer. Cowork only reads and writes within folders you explicitly connect. This is an important security boundary, so think about it carefully.

Choose Work in a folder on desktop and select a directory. A cloud session reaches that folder through Claude Desktop, so the app must remain open whenever Claude needs the local files.

My recommendation:

• Share: Your Documents folder, a dedicated workspace folder, your Downloads folder (great for cleanup tasks)
• Don’t share: Your entire home directory, your .ssh folder, anything with credentials or secrets

A good pattern is to create a dedicated ~/Claude-Workspace folder and point Cowork there. This keeps things clean and gives you a single place where all of Claude’s outputs land.

Your First Real Task

Everyone’s Downloads folder is a mess. Give Cowork this:

Organize my Downloads folder. Create subfolders by file type (Documents, Images, Spreadsheets, Archives, Installers, Other). Move each file into the appropriate subfolder. For any file older than 90 days, move it to an “Archive” subfolder instead. Give me a summary of what you moved when you’re done.

Claude will show you a plan:

1. Scan the Downloads folder and catalog all files with their types and dates
2. Create the subfolder structure
3. Categorize each file by extension and age
4. Move files to appropriate subfolders
5. Generate a summary report

You review the plan, approve it, and Claude executes. A few minutes later, your Downloads folder is organized and there’s a summary telling you exactly what moved where.

That’s Cowork in a nutshell. You describe the outcome, Claude figures out the steps, you approve, it delivers.

Pro tip: Run this on a copy of your Downloads folder first. Claude will ask before deleting anything, but better safe than sorry until you’ve built up trust.

Building Your Assistant’s Memory: Instructions and Context Files

Here’s where Cowork goes from “useful tool” to “personal assistant.” The difference is context. A tool does what you tell it. An assistant knows who you are, how you work, and what you need before you ask.

If you’ve read my post on context engineering, you already know how important it is to give AI models the right context. Cowork gives you three useful layers: global instructions, Project knowledge, and folder instructions for local work.

A Context Folder You Control

There is no magic global ~/Claude-Workspace/context/ directory that Cowork automatically loads everywhere. That was how I organized my early setup, but the current product has explicit places for persistent instructions.

• Put short, universal rules in Settings > Cowork > Global instructions.
• Put durable background documents into the relevant Claude Project.
• When working in a local folder, add folder instructions and keep reference files such as CONTEXT.md beside the work.

I still like a context/ folder because it is portable and easy to inspect. The difference is that I attach it to a Project or explicitly tell Cowork to read it. Here are the files I recommend creating:

Here are the files I recommend creating:

about-me.md

This is who you are. Your role, your company, your responsibilities, anything Claude needs to know to work effectively with you.

# About Me

- Name: [Your name]
- Role: VP of Marketing at [Company]
- Team: 4 direct reports (content, demand gen, product marketing, design)
- Key tools: HubSpot (CRM), Notion (wiki), Google Workspace, Slack
- Weekly leadership sync every Monday
- Current priorities: Q2 product launch, website redesign, annual planning

## Working style
- I prefer bullet points over paragraphs in internal docs
- I like data-driven arguments with specific numbers
- Meeting notes should capture decisions and action items, not discussion

brand-voice.md

If Claude is going to write for you, it needs to know how you sound. This is especially useful for email drafts, social posts, and any external-facing content.

# Brand Voice

## Tone
- Professional but not formal. We say "you" not "one."
- Confident and direct. Lead with the recommendation, then the reasoning.
- No jargon. If a concept needs explaining, explain it.

## Email style
- Keep emails under 5 sentences when possible
- Always include a clear ask or next step at the end
- Sign off with just my first name, no "Best regards" or "Cheers"

## Writing samples
[Include 2-3 short examples of your actual writing so Claude can
match your style. This makes a massive difference.]

working-preferences.md

Your formatting preferences, naming conventions, and operational patterns.

# Working Preferences

## File naming
- Use YYYY-MM-DD prefix for dated documents
  (e.g., 2026-08-25-quarterly-review.xlsx)
- Use lowercase-with-hyphens for everything else
- No spaces in filenames

## Document formatting
- Spreadsheets: always include a summary tab with key metrics at top
- Presentations: max 5 bullet points per slide, use company template
  in /Templates
- Reports: include an executive summary on page 1

## Recurring work
- Weekly team update is due every Monday by 9am
- Monthly metrics deck pulls from HubSpot and Google Analytics
- Quarterly business review uses the template in
  /Templates/QBR-template.pptx

Global Instructions vs. Project and Folder Context

Cowork has two layers of persistent instructions, and the distinction matters.

Global instructions (Settings > Cowork > Edit Global Instructions) apply to every single session. Keep these short and behavioral: “Always ask before deleting files.” “Use metric units.” “Respond in English.”

Project and folder context hold the detailed, reference-heavy information: your role, your voice, project goals, source material, and preferences. Keep these close to the work instead of stuffing everything into global instructions.

Think of global instructions as rules and context files as knowledge. Rules are always enforced. Knowledge is consulted when relevant.

The Payoff

Here’s what context files look like in practice. Without context files, you type “Draft an email to Sarah about the project timeline” and Claude writes a generic, formal email with placeholder content.

With context files loaded, the exact same prompt produces an email in your voice, signed off with just your first name (because brand-voice.md says so), kept under 5 sentences (your preference), referencing the actual Q2 launch timeline from about-me.md.

Same prompt, dramatically different output. Context is everything.

Projects: Organizing Your AI’s Workspaces

Once you start using Cowork for more than one area of your life, you’ll want Projects. A Project is a workspace with its own instructions, knowledge files, and conversations.

Why this matters: without Projects, everything bleeds together. Your marketing context leaks into your personal finance tasks. Your client work history mixes with your internal planning.

I run three Projects:

1. Work — Company tasks with work context files, HubSpot and Slack connectors
2. Content — Blog writing, newsletter prep, social media with brand voice context
3. Personal — File organization, travel planning, personal finance with separate preferences

Each Project supplements your global instructions with domain-specific knowledge. Claude knows your universal preferences, while the Project supplies the facts it needs for that body of work.

Create a Project from Claude’s sidebar, add instructions and files, then start Cowork inside it. One limitation matters: Cowork does not automatically rewrite the files stored as Project knowledge. Save important outputs back into the Project yourself. If a Project is tied to a local folder, its local-file sessions are desktop-only and need Claude Desktop online.

Pro tip: Put behavior in global instructions and facts in the relevant Project. This keeps your universal prompt short and reduces context collisions.

Connectors: Plugging Into Your Tools

Context files give Claude knowledge about you. Connectors give Claude access to your tools. This is where things get really powerful.

Cowork supports connectors across the tools most knowledge workers already use. The catalog changes frequently, so I would not optimize around a fixed connector count.

Category	Connectors
Productivity	Microsoft 365, Google Drive, Notion
CRM	HubSpot, Salesforce, Close
Project Management	Jira, Asana, Linear
Data	Snowflake, BigQuery, Databricks
Communication	Slack, Gmail, Zoom
Other	DocuSign, Apollo, Clay, WordPress

Setting Up Connectors

Open Customize, choose Connectors, find the service you want, and authenticate. Once a connector is active, reference the tool naturally in your task prompt and tell Claude exactly what data it may read or change.

Gmail + Google Calendar: The Daily Planner

This is the first combo I’d recommend setting up. Connect Gmail and Google Calendar, then try this:

“Check my calendar for today. Look at each meeting and pull any relevant email threads with the attendees from the past two weeks. For each meeting, create a prep doc with: who’s attending, what we last discussed over email, and the meeting agenda if one exists. Then look at the gaps between meetings and suggest time blocks for my top 3 priority tasks this week. Save the full day plan as /Daily/2026-08-25-day-plan.md.”

Claude reads your calendar, cross-references your email history, and produces a day plan with meeting prep baked in. I used to spend 20 minutes every morning doing this manually. Now it’s done before I sit down (more on scheduling this later).

HubSpot + Slack: Sales Intelligence

If you’re in a revenue role, this combo is gold. Connect HubSpot and Slack, then:

“Pull all deals in the ‘Negotiation’ stage from HubSpot. For each one, check the #sales Slack channel for any mentions of that company name in the past 30 days. Create a spreadsheet with columns for deal name, amount, days in stage, last activity date, and a summary of any Slack discussions. Flag any deal that’s been in Negotiation for more than 14 days with no Slack activity as ‘at risk’. Save as /Sales/2026-08-25-pipeline-review.xlsx.”

Claude pulls live CRM data, cross-references it with your team’s Slack conversations, and produces a pipeline review that would take an analyst an hour to assemble.

Notion + Jira: Project Status

For project managers running work across Notion (for docs) and Jira (for tickets):

“Check the ‘Q2 Launch’ Jira project for all tickets due this week. Cross-reference with the Q2 Launch page in Notion to see if any of the documented requirements don’t have a corresponding Jira ticket. Create a status report with: tickets completed this week, tickets at risk of missing their due date, and any gaps between the Notion spec and Jira coverage. Save to /Projects/Q2-Launch/2026-08-25-status.md.”

The real power here is the cross-referencing. Each tool on its own gives you a partial picture. Claude connects them and finds the gaps.

More Connector Combinations

A few other setups I’ve seen people use effectively:

• Gmail + Close CRM: Auto-draft follow-up emails for leads that haven’t been contacted in 7+ days, personalized based on their CRM record
• Google Drive + Slack: Monitor a shared Drive folder for new files and post a summary to a Slack channel when something’s added
• BigQuery + Google Sheets: Pull weekly metrics from your data warehouse and populate a pre-formatted Google Sheet for stakeholders
• Zoom + Notion: After a recorded Zoom meeting, pull the transcript and create a Notion page with meeting notes, action items, and decisions

The Cloud Caveat

One important thing to understand: connectors route data through Anthropic’s cloud, not your local network. When Claude pulls data from HubSpot or Slack, that information is processed on Anthropic’s infrastructure. For most people this is fine. But if you’re working with highly sensitive data or have strict compliance requirements, talk to your security team first.

If a service lives behind your corporate firewall, a cloud connector cannot reach it directly. Use exported local files, an approved local connector through Claude Desktop, or work with your IT team on a safe integration.

The Power of Sub-Agents

One of Cowork’s most impressive capabilities is how it handles parallelizable work. When you give it a task that can be broken into independent subtasks, Claude automatically spawns sub-agents to handle them concurrently.

This is the same architecture that makes Claude Code’s agent teams so effective. The difference is you don’t need to configure anything. Cowork figures out what can be parallelized on its own.

Here’s a real example. Say you have 12 customer interview transcripts and you need a synthesis:

Read all the interview transcripts in /Research/Q1-Interviews/. For each one, extract the top 3 pain points, any feature requests, and notable quotes. Then create a synthesis document that groups findings across all interviews by theme, ranks themes by frequency, and includes supporting quotes.

Without sub-agents, Claude would read transcript 1, extract findings, read transcript 2, extract findings, and so on. Sequential. Slow.

With sub-agents, Claude spawns 12 parallel workers. Each one processes a single transcript simultaneously. Then a coordinator agent synthesizes all the individual outputs into the final document. A task that would take 30+ minutes sequentially finishes in about 5 minutes.

You don’t control this directly. Claude decides when to parallelize based on the task structure. But you can encourage it by framing tasks as “do X for each of these items” rather than “go through these one by one.”

Creating Documents Like a Pro

Cowork doesn’t just output text in a chat window. It creates real, usable files on your filesystem. This is one of the features I use the most.

What Claude Can Create

• Excel files with working formulas, conditional formatting, and multiple tabs
• PowerPoint presentations with slide layouts and content hierarchy
• Formatted reports in markdown, HTML, or PDF
• CSV files for data imports/exports
• Any structured text file (JSON, YAML, configs, you name it)

Walkthrough: Quarterly Business Review Deck

Let’s walk through something practical. You need a QBR deck and you have raw data in a spreadsheet.

I need to create a Q1 quarterly business review presentation. Use the data in /Work/Q1-Data/metrics.xlsx and the template structure in /Templates/QBR-template.pptx.

The deck should include:

1. Executive summary slide with headline metrics (revenue, growth, retention)
2. Revenue breakdown by segment with quarter-over-quarter comparison
3. Customer metrics: new logos, churn, NPS trend
4. Product highlights: top 3 shipped features with adoption numbers
5. Q2 priorities (use the goals doc in /Work/Q2-Planning/goals.md)

Keep each slide to 5 bullet points max. Use the company color scheme from the template.

Claude reads the data, reads the template, reads the goals doc, and produces a complete deck. Will it be perfect on the first try? Probably not. But it gives you an 80% complete draft that you polish in 10 minutes instead of building from scratch in 2 hours.

The key is being specific about what you want. “Make me a QBR deck” gives Claude too little to work with. The prompt above specifies data sources, structure, constraints, and output format. This is context engineering in practice.

Receipts to Expense Reports

Another one I use constantly:

Look at all the receipt images in /Expenses/August/. Extract the vendor name, date, amount, and category for each one. Create an Excel spreadsheet with columns for each field, a Total row at the bottom with a SUM formula, and a pivot summary by category. Name it 2026-08-expenses.xlsx.

Claude opens each image, uses vision to read the receipt, extracts structured data, and creates a real .xlsx file with working formulas. Not a CSV you need to import and fix. A proper spreadsheet.

Scheduling: Your Assistant on Autopilot

Everything we’ve covered so far requires you to actively give Claude a task. Scheduling flips that. You define recurring tasks and Claude delivers results without you lifting a finger.

This is where the “personal assistant” vision really clicks. A good assistant doesn’t wait for you to ask for your morning briefing. They just have it ready when you sit down.

Setting Up Scheduled Tasks

Open Scheduled in the sidebar and create a task, or type /schedule in a Cowork session. Scheduled tasks run as separate cloud sessions and can use your connectors, skills, installed plugins, and files saved to your Claude account.

Here are the schedules I run:

Morning Briefing (Daily, 7:30am)

Check my Gmail for urgent emails that arrived overnight. Check my Google Calendar for today’s meetings. Check the #general and #product Slack channels for anything that needs my attention. Create a briefing with sections for Urgent, Calendar, and FYI. Attach it to the completed task as a markdown file.

Every morning when I sit down at my desk, there’s a briefing waiting for me. No inbox trawling, no Slack scrolling. Just the stuff that matters, organized and ready.

Weekly Metrics Pull (Monday, 8am)

Pull this week’s key metrics from HubSpot: new leads, MQLs, SQLs, opportunities created, deals closed, and revenue booked. Compare them with last week and the same week last year. Create an attached spreadsheet with a summary tab and a raw-data tab.

This saves me 30+ minutes every Monday morning. The data is pulled, formatted, and compared before I even think about it.

Meeting Prep (Daily, 8am)

Check my calendar for the next meeting with external participants. For each external attendee, do a web search for their LinkedIn profile and check for any recent email threads. Create a prep doc with attendee background, any open items from previous conversations, and suggested talking points based on the meeting agenda.

Friday Report (Friday, 4pm)

Review my completed Asana tasks and connected weekly metrics from this week. Create an attached weekly summary that covers: what got done, key metrics changes, blockers or escalations, and priorities for next week.

How to Create a Schedule

To set up any of these, open Cowork and type /schedule. You’ll get prompted for:

• Name: A descriptive name for the task (e.g., “Morning Briefing”)
• Frequency: daily, weekly, or a custom cron expression
• Time: When it should run
• Task description: The prompt Claude will execute

Your schedules show up in the Cowork sidebar, and you can edit, pause, or delete them anytime.

The Important Limitation

Scheduled tasks now run remotely, so they continue even when your computer is asleep or Claude Desktop is closed. That fixes the biggest limitation in the April version of this guide.

The tradeoff is that a scheduled task cannot be tied to a folder on your computer. Build remote schedules around connectors and files stored in your Claude account. If a workflow must read a local spreadsheet or drive a desktop app, start it as a regular Cowork task while Claude Desktop is online.

Web and Mobile: Continue the Same Task Anywhere

Scheduling handles recurring work. For ad hoc tasks, Cowork sessions now follow your Claude account across desktop, web, and mobile. You can start research at your desk, check progress from your phone, answer a question, and open the finished output later from another device.

How to Use It

1. Start a cloud Cowork task on any supported surface.
2. Open the same session from desktop, web, or mobile.
3. Check progress, answer Claude’s questions, or redirect the task.
4. Open or download the finished files from the session.

How It Works

Some examples of how I use it:

• In a meeting: “Turn these notes into decisions and action items, then create the tasks in Asana”
• On a walk: “Research the top 5 competitors in [space] and create a comparison matrix in /Research/”
• Commuting: “Draft a follow-up email to [person] referencing the proposal doc in /Deals/”

You get a notification when Claude finishes or needs input. Cloud-only work keeps running with your laptop closed. Local files, local connectors, browser use, and computer use are different: those capabilities still need the desktop app online because it acts as the bridge to your machine.

Plugins: Domain-Specific Superpowers

If context files are Claude’s memory and connectors are its hands, plugins are its specialized training. A plugin bundles together skills, connectors, and sub-agent configurations tailored for a specific domain.

Built-In Plugins

Cowork comes with pre-built plugins for common business functions:

• Sales: Pipeline analysis, deal forecasting, outreach drafting, competitive intel
• Finance: Expense reporting, budget tracking, financial modeling, invoice processing
• Legal: Contract review, compliance checking, NDA drafting
• Marketing: Campaign analysis, content calendar management, competitor tracking
• HR: Job description writing, candidate screening, onboarding checklists
• Data Analysis: Statistical analysis, visualization, data cleaning

Open Customize > Plugins > Browse plugins. Each plugin documents what it can do and which connectors it needs. Plugins are available on paid plans, and the bundled skills work in both Chat and Cowork. Hooks and sub-agents run only in Cowork.

Building Custom Plugins

If the built-in plugins do not cover your use case, use the Plugin Create plugin or customize an installed plugin. A plugin can bundle skills, connectors, hooks, and sub-agents into one reusable package.

Here’s an example. Say you regularly create client deliverables in a specific format:

# Client Deliverable Creator

When asked to create a client deliverable:

1. Use the template in /Templates/client-deliverable.docx
2. Pull the client's company info from HubSpot
3. Include our standard disclaimer text from
   /Templates/legal-disclaimer.md
4. Name the output file:
   YYYY-MM-DD-[client-name]-deliverable.docx
5. Save to /Deliverables/[client-name]/

Always include a cover page with our logo and the client's logo
(if available in their HubSpot record).

This is the same skill system that powers Claude Code’s custom slash commands. And if you’re on a Team or Enterprise plan, admins can create and distribute plugins across the organization. Great for standardizing how teams work with Claude.

The Plugin Marketplace

Anthropic includes plugin marketplaces inside Customize, and you can add approved marketplaces from GitHub repositories. Before building something custom, search the installed catalogs. You often find a useful starting point that you can customize.

Computer Use: The Experimental Option

Computer Use gives Claude direct control of your mouse, keyboard, and screen. It can navigate browser tabs, fill out forms, click buttons, and interact with desktop applications.

I want to be upfront: this is still a research preview. It works best for simple, predictable workflows and can struggle with nuanced, multi-step interfaces.

When to Use It

Computer Use makes sense when:

• The application has no API or connector (legacy internal tools, certain web apps)
• You need to interact with a UI that can’t be automated any other way
• The task is simple and repetitive (form filling, data entry across apps)

Don’t use it when:

• A connector exists for the service (connectors are faster and more reliable)
• File-based workflows would work (always prefer reading/writing files over UI manipulation)
• The task requires complex navigation or multi-step decision making

Setting It Up

Enable computer use in Claude Desktop > Settings > General. When Claude needs an application, it asks for access to that app. Your computer must remain awake and Claude Desktop must remain open while the task is controlling the screen.

A practical example:

Open the expense report form in our internal tool (it’s the browser tab titled “Expense Reports”). Fill in today’s date, the vendor name from the receipt image in /Expenses/receipt-2026-08-25.png, the amount, and submit.

For simple, repetitive tasks like this, Computer Use saves a lot of tedious clicking. Just don’t expect it to navigate a complex multi-step workflow reliably. Not yet. It’ll get better, but today it’s best suited for straightforward UI interactions.

One important caveat: computer use has no sandbox between Claude and the applications you approve. Claude takes screenshots, clicks, and types as your logged-in user. Block sensitive apps and monitor the first runs closely.

The Lesson I Brought Over From Jarvis

The best idea from my original Cowork guide came from building Jarvis, my own personal AI assistant: almost every repetitive piece of knowledge work can be expressed as a workflow.

Reading email means fetching messages, classifying them, and drafting replies. Preparing for a meeting means reading the calendar event, finding recent correspondence, and assembling a brief. Processing expenses means reading receipts, extracting fields, and writing rows into a spreadsheet. You do not need to write those scripts anymore. You need to describe the outcome and give Claude the right access.

A good test is: am I doing the same thing I did last week, just with different data? If the answer is yes, Cowork can probably take a meaningful part of it.

Package Repeated Work Into a Workflow

Do not rewrite your best prompt every morning. Save it as a skill, plugin, Project instruction, or reference file. Here is the email-triage workflow from my original guide, updated for the current connector model:

# Email Triage Workflow

1. Read unread Gmail messages from the last 24 hours.
2. Categorize each message:
   - URGENT: needs a response today
   - THIS WEEK: needs a response within seven days
   - FYI: no response needed
   - ARCHIVE: no action and safe to archive
3. Draft replies for URGENT messages, but do not send them.
4. Create a summary with sender, subject, category, and next action.
5. Flag any message where the right action is ambiguous.

Run it manually until the classification is reliable. Then package it into a plugin or schedule the read-only parts. Keep sending and archiving behind approval until you trust the workflow.

Start With Copies, Not Originals

This safety rule is also worth preserving from the old article. For your first file workflow, duplicate the input into a sandbox/ folder. Let Cowork organize or transform the copy, compare the result with the original, and only then move the workflow onto real data. Backups are cheaper than confidence.

Building Your Personal Assistant Stack

We’ve covered all the individual features. Now let’s put them together into a system. Here’s how I’ve configured Cowork as my daily assistant.

The Master Context File

In addition to the individual context files (about-me.md, brand-voice.md, working-preferences.md), I keep a master file called assistant-instructions.md that ties everything together:

# Assistant Instructions

You are my personal work assistant. Your job is to help me
be more effective by handling operational tasks, preparing
information I need, and creating documents in my preferred
formats.

## Core principles
- Always check if a template exists in /Templates/ before
  creating documents from scratch
- Save all outputs to the appropriate project folder, never
  to Desktop
- Use date-prefixed filenames (YYYY-MM-DD) for anything
  time-sensitive
- When pulling data from connectors, include a "data freshness"
  note showing when the data was pulled
- If a task is ambiguous, create the output but leave a
  [TODO] marker where I need to make a decision

## Approval rules
- Never delete files without asking first
- Never send emails or messages on my behalf without showing
  me a draft first
- Never modify files in /Archive/ (locked for compliance)

## Defaults when I'm not specific
- Spreadsheet format: .xlsx with a summary tab
- Document format: markdown
- Presentation format: .pptx using company template
- If I say "report," I mean a markdown doc with executive
  summary, key findings, and recommendations

A Day in the Life

Here’s what a typical workday looks like with this full setup running. I’m going to walk through a real Wednesday in detail so you can see exactly how the pieces connect.

7:30am — The Morning Briefing Arrives

My “Morning Briefing” schedule fires automatically. I set this up using /schedule in Cowork with a daily frequency at 7:30am. The task prompt I configured looks like this:

Check my Gmail for emails received since 6pm yesterday.
Categorize them as URGENT (from direct reports or flagged),
ACTION NEEDED (requires a response), or FYI (informational).
For urgent ones, draft a short response.

Check my Google Calendar for today's meetings. For any meeting
with external attendees, note who's attending and pull the
most recent email thread with them.

Check #general, #product, and #sales in Slack for any messages
with more than 3 reactions or any thread longer than 10 messages.

Create a markdown file attached to this task with sections for
Urgent, Today's Calendar, Slack Highlights, and Action Items.

When I open my laptop at 8am, there’s a markdown file waiting. Today’s briefing tells me: two urgent emails from a client about a timeline change, a board prep meeting at 2pm (with the last email thread about the deck attached), and a Slack thread in #product where users are reporting a bug in the latest release.

I didn’t open Gmail, Slack, or my calendar. I read one document and I know exactly what my day looks like and what needs attention first.

8:00am — Monday Metrics (Runs Mondays Only)

On Mondays, a second schedule fires at 8am. This one pulls key metrics from HubSpot, compares them to last week and the same week last year, and creates a formatted Excel file. My team standup is at 8:30 and the numbers are ready before I’ve finished my coffee.

The schedule prompt references my working-preferences.md context file, so the spreadsheet always has a summary tab on top, uses the date-prefix naming convention, and includes the “data freshness” timestamp I specified in my assistant instructions.

9:15am — Meeting Prep (Auto-Generated)

I have a meeting-prep schedule that runs each morning. Claude checks the day’s external meetings, researches the attendees, pulls our recent email history, and creates a one-page brief for each call.

Today’s 10am call is with a potential partner. The prep doc shows me their VP of BD’s background (15 years in enterprise SaaS, previously at Salesforce), a summary of our last three email exchanges, and the fact that we haven’t spoken in 22 days. I walk into the call prepared without spending a minute on research.

10:30am: Continue From My Phone

Between meetings, I’m grabbing coffee and remember I need a competitor comparison for an upcoming strategy session. I open the same Cowork session on my phone and type:

“Create a comparison deck of our product vs Acme Corp and GlobalTech using the feature matrix in /Research/competitive-landscape.xlsx. Include pricing, key differentiators, and where we win vs. lose. Save to /Projects/Strategy/competitor-comparison.pptx.”

I put my phone away. Fifteen minutes later I get a push notification: task complete. When I’m back at my desk, the deck is sitting in my Projects folder.

1:00pm — Batch Processing with Sub-Agents

After lunch, I need to synthesize 8 customer interview transcripts from last week’s research sprint. I open Cowork and give it the folder:

“Read all transcripts in /Research/Q1-Interviews/. Extract the top 3 pain points, feature requests, and notable quotes from each. Then synthesize across all interviews: group by theme, rank by frequency, include supporting quotes. Save the individual extracts and the synthesis to /Research/Q1-Interviews/analysis/.”

Claude spawns 8 sub-agents in parallel. Each one processes a transcript simultaneously. A coordinator agent waits for all of them to finish, then synthesizes the individual outputs into a themed report. The whole thing takes about 5 minutes. Doing this manually would have been my entire afternoon.

3:00pm — Quick Follow-Up Draft

“Draft a follow-up email to Sarah Chen from the 10am call. Reference the partnership proposal we discussed, mention I’ll send over the technical integration doc by Friday, and suggest next Tuesday at 2pm for a follow-up. Keep it short.”

Claude drafts it in my voice (thanks to brand-voice.md). Three sentences, ends with a clear next step, signs off with just my first name. I change one word and hit send.

4:00pm Friday — The Weekly Wrap

On Fridays at 4pm, the weekly report schedule fires. It pulls completed tasks from Asana, queries the week’s metrics through connectors, and compiles a summary with four sections: what got done, key metrics changes, blockers, and next week’s priorities.

I send this to my team every Friday. It used to take me 45 minutes to write. Now I spend 5 minutes reviewing Claude’s draft and adding any color commentary that needs a human touch.

The Compound Effect

After running this setup for a few months, the time savings compound in a way that’s hard to appreciate until you experience it. It’s not just the individual tasks. It’s the mental overhead that disappears. I don’t worry about whether I’ve missed an important email or forgotten to prep for a meeting. I don’t spend my mornings context-switching between Gmail, Slack, and my calendar trying to figure out what matters.

The scheduled tasks handle the information gathering. Cross-device sessions handle ad hoc requests. Project instructions make sure everything comes back formatted my way. I make decisions and do the work that actually requires my brain.

Security and Privacy

Before you go all in, let’s talk about what Claude can access and what to think about.

Your Security Boundaries

Cloud isolation protects your computer from code Cowork runs, but it does not make every task safe. Claude can still act through the folders, connectors, browser, and applications you authorize.

Folder permissions limit which local files Claude can reach. Share what Claude needs, nothing more.

Connectors route data through Anthropic’s cloud. When Claude pulls your Slack messages or HubSpot deals, that information passes through Anthropic’s infrastructure. If you’re on an Enterprise plan, you get additional controls: role-based access, spend limits, usage analytics, and OpenTelemetry integration for your SIEM pipeline.

Approval modes determine how much Cowork can do without stopping. Use manual approval for sensitive data or consequential actions. Claude always asks before permanently deleting files, but you should still keep backups.

Computer use operates outside the sandbox. When it controls your mouse and keyboard, it can interact with the applications you approved as your logged-in user. Watch it closely.

Best Practices

• Create a dedicated workspace folder rather than sharing your entire Documents directory
• Don’t put files with credentials, API keys, or secrets in shared folders
• Review the execution plan before approving tasks that touch connectors
• If you’re handling client data, check with your legal team about Anthropic’s data processing terms
• On Enterprise plans, use RBAC to control which team members can access which connectors
• Start with a sandbox folder for experimentation before connecting real working directories

Tips, Tricks, and Best Practices

After months of daily Cowork usage, here’s what I’ve learned.

Write Better Task Descriptions

The single biggest factor in output quality is how you describe the task. Be specific about:

• Input: Where the data is and what format it’s in
• Output: What format you want, where to save it, what to name it
• Constraints: Word counts, formatting rules, what to include or exclude
• Examples: Even with memory, a concrete reference is more reliable than “like the report I asked for last week.” Point to a file: “Use /Examples/good-report.md as a reference for structure and tone.”

Bad prompt:

Create a report about these files.

Good prompt:

Read the three PDFs in /Research/competitor-analysis/. Create a markdown report with these sections: Executive Summary, Key Findings (one per competitor), Comparison Table, and Recommendations. Save as /Reports/2026-08-competitor-analysis.md. Keep it under 2,000 words.

Use Context Files Aggressively

I add to my context files constantly. Every time I correct Claude on the same thing twice, I add it to the relevant context file. It’s like training a new hire. The more context you front-load, the fewer corrections you need later.

Start Small, Build Up

Don’t try to automate your entire workflow on day one. Start with one simple task (like the Downloads organizer). Get comfortable with how Cowork plans and executes. Then add context files. Then try a connector. Then set up your first schedule. Build incrementally.

My own progression:

1. Week 1: File organization and document creation
2. Week 2: Added context files, started getting drafts in my voice
3. Week 3: Connected HubSpot and Slack, started pulling live data
4. Week 4: Set up first scheduled tasks (morning briefing, weekly metrics)
5. Week 6: Started continuing cloud tasks from my phone
6. Week 8: Built custom plugins for recurring deliverables
7. Ongoing: Refinements, new workflows, expanding what I delegate

Each piece builds on the last. The compound effect is real. By month three, I had hours back every week.

When to Use Cowork vs. Chat vs. Claude Code

Use Case	Best Tool
Quick question or brainstorm	Chat
Single document, no file access needed	Chat
Multi-file tasks, batch processing	Cowork
Tasks that need your filesystem	Cowork
Tasks that connect to external tools	Cowork
Recurring/scheduled tasks	Cowork
Software development	Claude Code
Data science and notebooks	Claude Code

Managing Token Usage

Cowork consumes more tokens than regular chat because it’s planning, using tools, and often running sub-agents. On the Pro plan, heavy usage can burn through your cap faster than expected.

A few ways to manage this:

• Be intentional about what you schedule. Each scheduled run uses tokens whether you look at the results or not.
• Large batch tasks (processing dozens of files in parallel) consume tokens quickly.
• Extended thinking (Settings > Cowork) improves quality on complex reasoning but uses significantly more tokens. Keep it off by default, enable it for specific tasks.
• If you consistently hit your cap, the Max plan ($100/mo) gives you substantially more headroom.

Now Go Build Your Assistant

Cowork turned Claude from a chatbot I talked to into an assistant I delegate to. That shift is bigger than it sounds. Instead of spending time gathering information, formatting documents, and keeping up with operational tasks, I spend it making decisions and doing creative work.

The personal AI assistant is not a future vision. It is a product you can set up today with a paid Claude plan, a few good instructions, and carefully chosen connectors.

Start with one task. Organize your Downloads folder. Set up your context files. Connect a tool. Schedule your morning briefing. Build from there.

If you’re a developer looking for the coding equivalent, the Claude Code guide covers the same agentic power with developer tools. And if you want to understand the agent architecture underneath both products, my guide to AI agents goes deep on how this all works.

For product details that change faster than this guide, use Anthropic’s current documentation for getting started with Cowork, cloud sessions across devices, scheduled tasks, and Cowork safety.