turlockmike/cowork-bootstrap.md

Bootstrap: Set Up Your AI Workspace

You are setting up a new AI workspace for a user at Extend. This file is your instruction set. Follow it step by step, interacting with the user as you go. Be conversational, not robotic.

Validate after every phase. Do not move to the next phase until the current one passes. If something fails, help the user fix it before continuing.

Success Criteria

When bootstrap is complete, the user will have:

Identity confirmed — Name, role, team, and manager verified via Glean
Connectors working — Gmail, Google Calendar, Google Drive, Slack, and Glean verified with a live test (plus Atlassian and any role-specific connectors the user opted into)
Workspace structure created — docs/, notes/, meeting-prep/, tasks/, data/, .claude/skills/
Memory file written — .claude/CLAUDE.md with profile, company context, and preferences
Four skills installed — Morning plan, meeting prep, people lookup (rolodex), and one custom skill based on user's repetitive task
Live demo passed — At least one skill ran successfully against real data and the user confirmed the output format works
User knows what to do next — Daily rhythm explained, next steps given

Phase 1: Get to Know the User

Before building anything, learn about the user. Start by looking them up, then ask what you can't find.

Step 1: Look Up the User in Glean

Ask the user for their name. Then use Glean's people search to look up their profile:

Full name and title
Department and team
Manager / who they report to
Email address

Present what you found: "I looked you up — you're [Name], [Title] on [Team], reporting to [Manager]. Is that right?"

If Glean isn't connected or the lookup fails, ask the user directly for their name, role, team, and manager.

Step 2: Ask What Glean Can't Tell You

Once you have the org info, ask these questions naturally — not as a numbered quiz, have a conversation:

What does a typical day look like for you? (meetings, deep work, reactive tasks, etc.)
What are the 2-3 tasks you do most often that feel repetitive? (e.g., "I prep for my weekly 1:1 every Monday," "I check pipeline numbers every morning," "I write the same kind of status update every Friday")
How do you prefer to communicate? (brief and direct? detailed? formal? casual?)

Save all answers — you'll use them throughout setup.

Validate Phase 1

User's name, role, team, and manager are confirmed
You have their typical day, repetitive tasks, and communication preference

Phase 2: Connect and Verify Integrations

Before creating files or skills, make sure the user's connectors are working. Skills are useless without data access.

Required Connectors

Walk the user through connecting each one. After each connector, test it with a real query to verify it works.

Gmail

Tell the user: "Let's connect your email. Go to Customize → Connectors → Gmail and sign in with your Extend Google account."

Wait for them to confirm it's done.

Verify: Search for a recent email. Example: "Let me check — I'll search your inbox for something recent."

Pass: Found recent emails. Tell the user: "Gmail is connected and working."
Fail: No results or error. Help troubleshoot — wrong account? Permission denied? Try reconnecting.

Google Calendar

Tell the user: "Now let's connect your calendar. Same flow — Customize → Connectors → Google Calendar."

Verify: Check today's calendar. "Let me pull up your schedule for today."

Pass: Found calendar events (or confirmed no events today). "Calendar is connected."
Fail: Troubleshoot.

Google Drive

Tell the user: "Next is Google Drive — same Connectors menu."

Verify: Search for a recent document. "Let me see if I can find a recent doc in your Drive."

Pass: Found documents. "Drive is connected."
Fail: Troubleshoot.

Slack

Tell the user: "Now Slack — this gives me access to search messages and channels."

Verify: Search for a recent message in a channel the user would be in. "Let me check — I'll search for a recent message in one of your channels."

Pass: Found messages. "Slack is connected."
Fail: Troubleshoot.

Atlassian (Jira + Confluence) — Optional

Ask the user: "Do you use Jira or Confluence in your work?"

If yes: Walk them through connecting. "Let's get Atlassian connected — Customize → Connectors → Atlassian."

Verify: Search for a Jira ticket or Confluence page. "Let me look for a recent Jira ticket assigned to you."

Pass: Found tickets or pages. "Jira and Confluence are connected."
Fail: Troubleshoot.

If no: Skip. "No problem — we'll skip Atlassian. You can always add it later if you need it."

Glean

Tell the user: "Last core one — Glean. This gives me access to search across all your company's tools."

Verify: You already used Glean in Phase 1 to look up the user. If that worked, it's connected.

Pass: Glean worked in Phase 1. "Glean was already working — we used it to look you up."
Fail: If Phase 1 Glean lookup failed, walk them through connecting it now.

Optional Connectors

Based on the user's role, suggest additional connectors:

Engineering: "You're on the engineering team — you'll probably want GitHub connected too. Want to set that up now?"
Finance: "For finance, NetSuite and Ramp are available. Want to connect those?"
Design: "Figma is available as a connector. Want to add it?"

Don't push these — offer them. If the user wants to skip, that's fine.

Validate Phase 2

Summarize the connector status to the user:

"Here's where we stand on your connections:

Gmail: ✅

Google Calendar: ✅

Google Drive: ✅

Slack: ✅

Glean: ✅

Atlassian: ✅ / skipped

[Any other optional connectors: ✅ or skipped]

Your integrations are working. Let's build your workspace."

Do not proceed to Phase 3 until the 5 core connectors pass (Gmail, Calendar, Drive, Slack, Glean). Atlassian and other connectors are optional. If any core connector failed, help the user fix it first.

Phase 3: Create the Workspace Structure

Create the following folder structure in the current working directory:

docs/              # Reference documents, SOPs, policies
notes/             # Meeting notes, daily plans, research
meeting-prep/      # Pre-meeting briefs
tasks/             # Action items, project tracking
data/              # Reports, exports, synced data
.claude/
  skills/          # Custom skills (you'll create these next)

Create each folder. Then explain to the user:

"I've set up your workspace. Here's what each folder is for:

docs/ — Put any reference documents here. Team SOPs, policies, guides. I'll read these for context when you ask me questions.

notes/ — Meeting notes, daily plans, research notes. Anything time-based goes here.

meeting-prep/ — I'll save pre-meeting briefs here when you ask me to prep for a meeting.

tasks/ — Your action items and project tracking. We'll manage this together.

data/ — Reports, exports, or any data you want me to work with.

.claude/skills/ — This is where your custom automations live. I'm about to create a few starter ones for you."

Validate Phase 3

All 6 directories exist
User understands what each one is for

Phase 4: Set Up Memory

Create a file at .claude/CLAUDE.md with the user's profile and preferences. This file gets loaded into every conversation automatically.

# Workspace Instructions

## About Me
- Name: [from Glean lookup]
- Role: [from Glean lookup]
- Team / Department: [from Glean lookup]
- Reports to: [from Glean lookup]
- Email: [from Glean lookup]
- Communication style: [brief/detailed, formal/casual — based on their answer]

## Company: Extend
- Extend is a Personalized Shopper Operations platform
- Four products: Shopper Intelligence, Returns & Exchanges, Delivery, Product Protection
- All powered by Shopper Intelligence (AI-driven shopper segmentation from 30+ data points)
- We sell three outcomes: improve margins, acquire customers, boost loyalty
- "Merchant" = our customer (the e-commerce store)
- "Consumer" / "Shopper" = the end buyer
- "Contract" = a protection plan purchased by a consumer
- "Claim" = when a consumer files for protection coverage
- "Pipeline" = Salesforce sales pipeline, not data pipelines

## My Preferences
- [Add any preferences they mentioned during the conversation]
- [e.g., "I prefer bullet points over paragraphs"]

## Workspace Structure
- docs/ — Reference documents and SOPs
- notes/ — Meeting notes and daily plans
- meeting-prep/ — Pre-meeting briefs
- tasks/ — Action items and project tracking
- data/ — Reports and exports
- .claude/skills/ — Custom automations

Explain to the user:

"I've created a memory file that tells me who you are and how you like to work. I'll read this at the start of every session. As we work together, I'll update it with things I learn — like which senders are VIPs in your inbox, or how you like your status updates formatted. You can also edit it directly anytime."

Validate Phase 4

.claude/CLAUDE.md exists with correct user info
Read the file back and confirm it looks right

Phase 5: Create Starter Skills

Create three skills based on the user's role and answers. Each skill is a SKILL.md file inside .claude/skills/<name>/.

Skill 1: Morning Plan

Create .claude/skills/morning/SKILL.md:

Customize the process based on the user's role and department (from the Glean lookup). For example, if they're on the sales team, include "check pipeline." If they're an engineer, include "check PRs." Use their role to infer relevant tools — don't ask.

---
name: morning
description: >-
  Start my day with a plan. Use when I say "morning", "start my day",
  "what should I focus on today", or "plan my day".
---

# Morning Plan

## Process

1. Check my calendar for today's meetings and time blocks
2. Check my email for anything that needs a response today
3. [CUSTOMIZE: Add tool-specific checks based on user's role]
4. Review my open tasks in tasks/
5. Write a daily plan with:
   - Today's schedule (meetings + open blocks)
   - Top 3 priorities (based on urgency and what's on the calendar)
   - Any heads-up items (deadlines, blocked tasks, things that need prep)
6. Save the plan to notes/YYYY-MM-DD.md

## Rules

- Keep the plan under 30 lines — this is a reference card, not a report
- Top 3 priorities should be specific and actionable, not vague
- Flag meetings that need prep (more than 2 attendees or with leadership)

Customize this for the user's role. Replace the [CUSTOMIZE] line based on their department from Glean:

Sales / Revenue Ops: "Check for any pipeline changes or deal updates in Salesforce"
Engineering: "Check for any PRs awaiting my review on GitHub"
Data Engineering / Analytics: "Check Slack (#data-eng-public) for any pipeline failures or data alerts"
Solutions / Customer Success: "Check for any escalated merchant issues or support tickets"
Product: "Check Jira for any tickets assigned to me that were updated overnight"
Finance: "Check for any pending approvals in NetSuite or Ramp"
Any role: "Check Jira for any tickets assigned to me that need attention"

Skill 2: Meeting Prep

Create .claude/skills/meeting-prep/SKILL.md:

---
name: meeting-prep
description: >-
  Prepare for an upcoming meeting. Use when I say "prep me for my meeting",
  "get ready for my 1:1", "meeting prep", or "prep for [meeting name]".
---

# Meeting Prep

## Process

1. Check my calendar for the meeting I specified (or the next upcoming one)
2. Look up the attendees — who are they, what team, what's their role?
3. Search Slack for recent threads involving these people or this topic
4. Check for any shared documents, prior meeting notes, or open action items
5. Write a prep brief with three sections:
   - **Raise** — decisions I need to make or asks I need to bring
   - **Status** — one-line FYIs that would change the conversation
   - **Watch** — emerging issues worth mentioning
6. Save the brief to meeting-prep/YYYY-MM-DD-meeting-name.md

## Rules

- Keep it to one page — a 60-second read, not a report
- Focus on what's new or changed — skip things I already know
- If there's nothing notable, say so in one line

Skill 3: People Lookup (Rolodex)

Create .claude/skills/people-lookup/SKILL.md:

This skill is useful for everyone — it answers "who is [name]?", "who do I talk to about X?", and "what team is [name] on?"

---
name: people-lookup
description: >-
  Look up employee info. Use when I say "who is", "find person",
  "look up employee", "people search", "who do I talk to about",
  "who owns", or "who's on [team]".
---

# People Lookup

Look up an employee's info using Glean people search or Slack user search.

## Process

1. Search Glean with `app: "people"` for the person's name or topic
2. If searching by topic ("who do I talk to about X"), search Glean for the topic first to find relevant people, then look them up
3. If Glean doesn't have it, fall back to Slack user search
4. Return a structured result:
   - Name
   - Title / Role
   - Team / Department
   - Manager
   - Email
   - Slack User ID (if available)

## Rules

- Always try Glean first — it has the most complete org data
- Use Slack search as fallback, not primary
- If the person isn't found in either, say so clearly
- Don't guess or infer role/team from channel membership — verify from Glean's people directory

Skill 4: Custom Skill Based on User's Repetitive Task

Based on the user's answer to "What tasks feel repetitive?", create a third skill tailored to their specific workflow. Examples:

If they said "I write a status update every Friday": Create .claude/skills/weekly-status/SKILL.md that aggregates their week's activity from Slack, email, and tasks into a formatted update.

If they said "I check pipeline numbers every morning": Create .claude/skills/pipeline-check/SKILL.md that pulls Salesforce data and summarizes changes.

If they said "I prep the same onboarding docs for every new merchant": Create .claude/skills/merchant-onboarding/SKILL.md that generates customized onboarding materials from templates.

If they said "I triage my inbox every morning": Create .claude/skills/inbox-triage/SKILL.md that reads recent emails, categorizes by priority, and surfaces what needs attention.

Build whatever matches their answer. If they gave multiple repetitive tasks, pick the one they'd benefit from most and tell them you'll build the others later.

Validate Phase 5

Four skill files exist in .claude/skills/ (morning, meeting-prep, people-lookup, and one custom)
Read each file back and confirm it looks correct
Morning skill is customized for the user's role (not generic)
People lookup skill works — test it by looking up the user's manager

Phase 6: Live Demo and Validation

Now show the user the system works end-to-end. Pick the most relevant skill and run it live against real data.

Pick the Right Demo

If they have meetings today: Run /meeting-prep for their next meeting.
If it's morning: Run /morning to generate today's plan.
If neither: Run the custom skill you built (Skill 3) on real data.

Run It

Execute the skill. Show the output to the user. Then ask:

"Here's what I generated. Does this format work for you? Anything you'd change — too long, too short, missing something, wrong tone?"

If they give corrections: Apply them immediately. Update the skill file and/or memory file based on their feedback. Then re-run and confirm the updated output is better.

Validate Phase 6

At least one skill ran successfully against real data (not placeholder content)
User confirmed the output format works (or corrections were applied and re-validated)

Phase 7: Wrap Up and Next Steps

Summarize What Was Built

"Here's everything we set up today:

✅ Your profile and org info (from Glean)

✅ Connectors: Gmail, Calendar, Drive, Slack, Atlassian, Glean [+ any optional]

✅ Workspace: 6 folders organized by domain

✅ Memory: I know your role, team, preferences, and Extend context

✅ Skills: Morning plan, meeting prep, people lookup, and [custom skill name]

✅ Verified: [Skill name] ran on real data and you approved the format"

Suggest Next Steps

Based on everything you've learned about the user, suggest 2-3 specific next steps:

A daily habit to start: "Try running /morning at the start of your day for the next week. It takes 2 minutes and you'll have a clear plan before your first meeting."
A document to add: "If you have any team SOPs or process docs, drop them in docs/. I'll use them for context when you ask me questions about how things work."
The next skill to build: Based on their other repetitive tasks, suggest what to automate next. "Next week, let's build a skill for [their second repetitive task]. Once you've used the morning plan for a few days, you'll have a feel for how skills work."

Final Message

"Your workspace is set up and verified. You have four skills, your connectors are live, and I know who you are and how you work. Every correction you give me makes this better — the system learns over time. Let's get to work."

Reference

Getting Started Guide — Full guide with the 5 levels of Cowork
Extend Setup — Extend-specific connectors, roles, and skill ideas
Philosophy — Why this approach works
Official Cowork Docs — Anthropic's setup guide
Learn Cowork IN Cowork — Free 12-lesson interactive course

Cowork Setup for Extend

Extend-specific configuration for Claude Cowork. Follow the Getting Started Guide first for the general setup, then use this doc to configure Cowork for how Extend works.

Available Connectors

These connectors are available to all Extend employees today. Install all that are relevant to your work:

Connector	What it gives you
Gmail	Email triage, drafting, searching threads
Google Calendar	Meeting prep, scheduling, daily planning
Google Drive	Reading/writing shared docs, slides, sheets
Slack	Searching conversations, posting updates, reading channels
Atlassian (Jira + Confluence)	Tickets, sprint boards, wiki, SOPs, runbooks
GitHub	PRs, issues, code, CI status
Glean	Internal knowledge search, people lookup, cross-tool search
Coralogix	Log search, metrics, traces, alert investigation
Figma	Design files, components, prototypes
Lucid	Diagrams, flowcharts, architecture drawings
NetSuite	Financial data, accounting, ERP
Ramp	Expense management, corporate card transactions

Requesting Additional Connectors

The list above covers what's available today. If you need a data source that isn't listed (e.g., Salesforce, Snowflake, Tableau, or any other tool):

Type of connector	Who to contact
Sales & Analytics tools	Mitch
Engineering & Infrastructure tools	IT / DevOps

Until a connector is set up, you can still work with data by exporting it as CSV or JSON to your data/ folder and asking Cowork to analyze it.

Extend Org Context

Add this to your company.md file so Cowork understands Extend:

# Company: Extend

## What We Do
Extend is a Personalized Shopper Operations platform. Our tagline: "Smarter commerce. Happier customers. One unified platform." We turn shopper behavior into a strategic advantage by combining personalization with operational efficiency across the customer journey.

## Four Products

### Shopper Intelligence
AI-powered analytics that segments customers based on 30+ data points (purchase history, return patterns, fraud risk, account age, payment behavior). Replaces one-size-fits-all policies with individualized perks and policies informed by actual shopper actions. "The future of commerce starts with Shopper Intelligence."

### Returns & Exchanges
Smart return policies built to drive loyalty. Rewards valuable shoppers (instant exchanges, VIP return experiences) while detecting and preventing abuse (automated flagging, manual review for high-risk). Turns returns from a cost center into a revenue driver with dynamic, behavior-driven policies.

### Delivery
Shipping protection, order tracking, and automated claim resolution. Includes branded tracking emails/SMS, WISMO reduction, and coverage for lost/stolen/damaged deliveries — offered as a VIP value-add or purchasable add-on.

### Product Protection
Warranty coverage for product failures and accidental damage (drops, breaks, spills). Extend handles claims directly — repairs or replaces, 24/7 online support. Custom-branded checkout integration. Drives incremental revenue while building customer loyalty.

All four products are powered by Shopper Intelligence and integrate via API or pre-built connectors for Shopify and BigCommerce.

## Three Business Outcomes We Sell
1. **Improve Margins** — Reduce risk, boost profits
2. **Acquire New Customers** — Win customers with embedded protection
3. **Boost Customer Loyalty** — Personalized perks for best shoppers

## Key Terminology
- "Merchant" = our customer (the e-commerce store)
- "Consumer" / "Shopper" = the end buyer
- "Contract" = a protection plan purchased by a consumer
- "Claim" = when a consumer files for protection coverage
- "Pipeline" = Salesforce sales pipeline (not data pipelines)
- "The platform" = Extend's Shopper Operations platform

## Key Tools
- Code: GitHub (org: helloextend)
- Project Management: Jira
- Documentation: Confluence
- Communication: Slack
- Calendar/Email: Google Workspace
- Internal Search: Glean
- Identity: Okta
- Observability: Coralogix
- Finance/ERP: NetSuite
- Expenses: Ramp
- Design: Figma
- Diagrams: Lucid

Note: You don't need to manually add org hierarchy or your reporting chain. The bootstrap process (see below) uses Glean's people search to look that up automatically.

Skills by Role

Starter skill ideas for each Extend team. Build them as you need them — don't try to create all at once.

Sales & Revenue Ops

Skill	Trigger	What it does
`/pipeline-brief`	"pipeline update", "deal summary"	Pull Salesforce pipeline changes, flag risks, summarize stalled deals
`/prospect-prep`	"prep me for the call with [company]"	Research a prospect: company background, contacts, recent news, Salesforce history
`/qbr-prep`	"prepare QBR", "quarterly review"	Aggregate pipeline data, win/loss trends, key metrics for QBR materials

Data Engineering & Analytics

Skill	Trigger	What it does
`/query-review`	"review this query", "check this SQL"	Review a Snowflake query for performance, conventions, and correctness
`/pipeline-alert`	"what broke", "pipeline failure"	Search Slack (#data-eng-public, alert channels) for recent pipeline failures, summarize
`/metric-lookup`	"how is [metric] defined", "where is [metric] dashboarded"	Search Confluence + Tableau for metric definition, owner, and dashboard location
`/data-doc`	"document this model", "write docs for [table]"	Generate documentation for a dbt model or Snowflake table

Solutions & Customer Success

Skill	Trigger	What it does
`/merchant-status`	"status on [merchant]", "what's going on with [merchant]"	Pull contract status, recent support tickets, integration health, open Jira issues
`/integration-debug`	"troubleshoot [merchant] integration"	Search logs, docs, and Slack for known issues with a merchant's integration
`/onboarding-prep`	"onboarding for [merchant]"	Generate customized onboarding materials based on merchant's platform and products

Product

Skill	Trigger	What it does
`/write-prd`	"write a PRD for [feature]"	Synthesize customer feedback, support tickets, analytics data into a PRD
`/competitive-analysis`	"compare us to [competitor]"	Research competitor features, pricing, positioning, recent announcements
`/user-research-summary`	"summarize research on [topic]"	Pull Confluence research docs, Slack discussions, and synthesize into themes

Engineering

Skill	Trigger	What it does
`/pr-summary`	"summarize this PR", "what changed in [repo]"	Read a PR's diff and produce a human-readable summary with risk assessment
`/incident-timeline`	"what happened with [incident]"	Compile timeline from Slack alerts, Jira tickets, and deploy logs
`/architecture-review`	"review this design"	Evaluate a design doc against Extend's architecture patterns and conventions

Finance

Skill	Trigger	What it does
`/revenue-reconciliation`	"reconcile revenue for [period]"	Compare Salesforce bookings vs. Snowflake revenue vs. billing system
`/budget-variance`	"budget variance for [team/period]"	Analyze spend vs. budget, flag overages, summarize for leadership
`/contract-review`	"review this contract"	Flag non-standard terms, compare against Extend's standard agreement

Marketing

Skill	Trigger	What it does
`/content-draft`	"draft a blog post about [topic]"	Research topic, pull internal data/examples, draft in Extend's voice
`/campaign-report`	"how did [campaign] perform"	Pull campaign metrics, compare to benchmarks, recommend adjustments

HR & People

Skill	Trigger	What it does
`/candidate-screen`	"screen this resume for [role]"	Compare resume against job requirements, flag strengths and gaps
`/offer-prep`	"prepare offer for [candidate]"	Generate offer letter from template with role-specific compensation data

Recommended Plugins

Pre-built plugins from the Cowork marketplace useful for Extend:

Plugin	Best for	What it does
Legal	Anyone reviewing contracts	Contract review, NDA analysis, compliance checks
Sales	Revenue ops, account managers	Pipeline management, prospect research, proposal drafting
Finance	Finance team	Budget analysis, invoice review, financial reporting
Productivity	Everyone	Email management, meeting prep, task prioritization
Project Management	PMs, engineering leads	Sprint planning, status updates, roadmap management

Install from: Customize → Browse Plugins

Getting Help

Slack: Post in #devx-private for Cowork setup questions
Connector requests: Sales/Analytics tools → Mitch | Engineering tools → IT/DevOps
Guide: Getting Started Guide — General Cowork setup (the 5 levels)
Video: The 5 Levels of Claude Cowork — Watch the full walkthrough
Interactive course: Learn Cowork IN Cowork — 12 hands-on lessons
Bootstrap: cowork-bootstrap.md — Point Cowork at this file and say "bootstrap from this" for guided setup

Getting Started with Claude Cowork: The 5 Levels

A practical guide to setting up Cowork and unlocking its full potential — from first prompt to fully automated workflows.

Based on The 5 Levels of Claude Cowork. This guide is the text version — watch the video for screen recordings, or read this for a step-by-step reference.

What is Cowork?

Cowork is the AI that works alongside you — not just answers questions. It's built into the Claude Desktop app and brings the same power that developers get from Claude Code to everyone else.

What makes it different from a chatbot:

It can read, edit, and create files on your computer
It can connect to your apps — Gmail, Google Drive, Slack, Salesforce, Notion, and more
It can break complex tasks into steps and complete them start to finish
It can run tasks on a schedule while you're not there
It learns about you over time and gets better with every interaction

Most people open Cowork, type one prompt, and wonder why it feels like every other chatbot. That's because there are five levels to master — and most people are stuck on level one.

Level 1: Import — Bring Your Context

If you've been using ChatGPT, Gemini, or another AI tool, you don't have to start from scratch. Claude has an import feature that transfers your memory from another AI in under 60 seconds.

How to do it

Go to Claude's import page (linked in the video description)
Click Get Started — you'll get a prompt to copy
Paste that prompt into ChatGPT (or whatever AI you've been using)
ChatGPT will export everything it knows about you as a structured document
Copy that output, bring it back to Claude, and paste it in
Claude now has all the context your other AI had

Why this matters: Without context, Claude gives generic responses. With your imported memory, it already knows your role, your business, your preferences, and your communication style from day one.

Level 2: Foundation — Set Up Your Brain

This is like onboarding a new hire. You write a few files once, and Cowork reads them every time it starts a task.

Step 1: Download the Claude Desktop App

Download from claude.com/download for Mac or Windows. Open it and switch to Cowork mode (top of the screen — you'll see Chat, Cowork, and Code).

Step 2: Create Your Workspace Folder

Click Choose Folder and create a new folder (e.g., Claude Workspace). Click Allow so Cowork can read and write files in it.

Step 3: Create Your Foundation Files

These are the files Cowork reads to understand you and your work:

Instructions file (.claude/CLAUDE.md): Your profile — name, role, team, communication preferences, standing rules. Think of this as the "employee handbook" Cowork reads at the start of every session.

Goals file (goals.md): Your priorities on a quarterly, monthly, and weekly basis. Without goals, Cowork gives generic responses. With goals, everything it does is oriented toward what you're actually trying to accomplish.

Company context (company.md): Your business, your brand, your tech stack, the platforms you use. This gives Cowork the vocabulary and context to speak your language.

Tip: You don't have to write these from scratch. Tell Cowork: "Help me create a goals.md file. Ask me questions about my quarterly goals, monthly targets, and weekly priorities." It will interview you and generate the file.

How It Works

Think of the instructions file as a kitchen bible. After every session, the chef updates the book. So every time a new session starts, Cowork already knows exactly what to do — without you explaining it again. This is a self-learning system that gets better every time you use it.

Level 3: Workflows — Plugins and Skills

This is where Cowork goes from assistant to employee. Workflows are triggered by a single command and execute multiple steps automatically.

What Are Plugins?

Plugins are skill packs for every job — productivity, marketing, sales, finance, HR, legal, design, operations. Each plugin bundles commands and workflows that you can invoke with a slash command.

Using Pre-Built Plugins

Click Customize (top left)
Click Browse Plugins
Browse by category — legal, finance, marketing, etc.
Click Install on any plugin you want

Each plugin adds commands you can invoke. For example, the Legal plugin gives you:

/brief — Generate a contextual briefing for legal work
/review-contract — Upload a contract and get a full review with flagged clauses

Example: Upload a contract PDF, type /review-contract, and in 2 minutes you have a full review document with green (approved) and yellow (needs attention) flags on every clause.

Building Your Own Plugins

You don't need to be a developer. Two ways to do it:

Method 1: Natural language. Do a task in Cowork, then say: "Turn this into a plugin I can run with one command." Cowork creates the plugin files for you.

Method 2: Write a skill file. Create a Markdown file describing the workflow. A skill is just a set of instructions in plain language:

---
name: weekly-status
description: Generate my weekly status update. Use when I say "weekly status" or "status update".
---

# Weekly Status Update

## Process
1. Check my completed tasks from this week
2. Check Slack for highlights from my channels
3. Check email for any threads that need mention
4. Write a status update with: Done, In Progress, Blocked, Next Week
5. Save to notes/weekly-status-YYYY-MM-DD.md

Sharing Plugins

Once you build a plugin, you can share the folder with teammates. They drop it into their system and have the same workflow instantly. Admins can also provision plugins across the whole organization.

Level 4: Ecosystem — Connect Your Apps

An employee isn't useful if they can't access your tools. Connectors give Cowork access to the apps you use every day.

Built-In Connectors

Click Customize (top left)
Click Connectors
Click the + button to browse available apps
Click any app → sign in → authorize

Available connectors include: Gmail, Google Calendar, Google Drive, Slack, Notion, Figma, Canva, Salesforce, and 30+ more.

What This Unlocks

With connectors, a single prompt can touch multiple apps:

"Check my email, look for calendar conflicts today, draft a reply to the urgent one, and save a summary to Google Drive."

That's four apps accessed in one prompt — email, calendar, email again for the reply, and Drive for the save.

Expanding Beyond Built-In Connectors

If your app isn't in the built-in list, use Zapier MCP to connect to 8,000+ additional apps:

Go to zapier.com/mcp
Create a new MCP server, select Claude Cowork as the client
Add the tools you need (Airtable, Zendesk, HubSpot, etc.)
Back in Cowork, go to Connectors → search for Zapier → add it

Now those apps are available in Cowork alongside the built-in connectors.

Level 5: Automation — Run Tasks Without You

This is where Cowork becomes a true employee. Scheduled tasks run automatically — daily, weekly, hourly, whatever cadence you choose.

How to Create a Scheduled Task

Click Schedule in Cowork
Click New Task
Give it a name, a prompt describing what to do, and choose a frequency
Select which folder it should work in
Choose the model (Opus for best quality, Sonnet for speed)

Examples of Scheduled Tasks

Task	Schedule	What it does
Morning brief	Daily, 7:00 AM	Summarize emails, calendar, and priorities
Competitor research	Daily, 8:00 AM	Scrape competitors, generate a trends dashboard
Weekly report	Monday, 9:00 AM	Pull data from multiple sources, format, save to Drive
Inbox cleanup	Daily, 6:00 PM	Archive low-priority emails, flag urgent ones

Important Notes

Your computer must be awake and Claude Desktop must be open for scheduled tasks to run
If your laptop is closed, tasks wait until it's open again
After each run, Cowork can rewrite its own instructions to get better

The Natural Way to Build Automations

You don't have to go to the Schedule screen first. The best automations come from working naturally:

Ask Cowork to do something
It does it well
You say: "That was great — turn this into a scheduled task that runs every Monday at 9 AM"

That's it. One sentence turns a one-off task into a recurring automation.

The Compounding Effect

Here's what most people miss: the system gets better every time you use it.

After every session, Cowork updates its understanding of you
Every correction you give gets stored in memory
Every skill you build makes the next week more efficient
Every connector you add gives Cowork more context to work with

After a week, you'll notice Cowork stops asking basic questions. After a month, it anticipates what you need. After three months, you have a personal operating system that knows your work, your preferences, and your rhythm.

The difference between "using Claude" and building a workspace is the difference between using a calculator and building a spreadsheet. Both compute. The spreadsheet compounds.

Quick Reference

What you want	What to do
Download Cowork	claude.com/download — switch to Cowork mode
Import context from ChatGPT	claude.ai/import — 60-second transfer
Set up foundation	Create `CLAUDE.md`, `goals.md`, `company.md` in your workspace
Install a plugin	Customize → Browse Plugins → Install
Build a custom skill	Create a Markdown file with instructions in `.claude/skills/<name>/`
Connect an app	Customize → Connectors → + → select app → authorize
Connect 8000+ apps	zapier.com/mcp → create server → add to Connectors
Schedule a task	Schedule → New Task → set name, prompt, frequency
Turn a task into automation	Do it once, then say "turn this into a scheduled task"

Learn More

Watch

The 5 Levels of Claude Cowork — The video this guide is based on. Watch for screen recordings of every step.
Claude Cowork: From Beginner to Expert in 18 Minutes — Compact walkthrough from setup to power-user techniques.

Try

Learn Cowork IN Cowork — Free 12-lesson interactive course (~2 hours). Download the folder, open in Cowork, say "Read START-HERE.md and start lesson 1." Claude teaches you by doing.

Read

Get Started with Cowork — Official Anthropic guide
Cowork Product Page — Overview of features
Use Plugins in Cowork — Plugin setup and management
How to Build a Plugin from Scratch — Official plugin tutorial
Anthropic Knowledge Work Plugins — Open-source plugin repo
Cowork Guide for Power Users: 50+ Tips — Deep dive on plugins, skills, sub-agents, memory
Claude Skills 2.0 — Guide for Non-Techies — Skills focused on non-technical users

Extend-Specific

Extend Cowork Setup — Which connectors to install, role-specific skills, and Extend tools

Philosophy

The AI-Native Workspace: Philosophy — Why this approach works. For leadership and anyone who wants the deeper thinking.

The AI-Native Workspace: Philosophy

Date: 2026-03-20 Audience: Leadership adopting Cowork across the organization Companion doc: Getting Started Guide — the practical how-to for team members

The Core Insight

Most people who adopt AI tools treat them like a chatbot: you open a window, ask a question, get an answer, close the window. The context dies. Tomorrow you start over.

This approach does something fundamentally different. It treats the AI as a permanent assistant operating inside a structured environment. The AI doesn't just answer questions — it reads your files, manages your tasks, prepares you for meetings, triages your inbox, automates your work and remembers what happened yesterday. It does this because the workspace is designed to make all of that possible.

The tool is not the product. The workspace is the product. Cowork is just the engine. The workspace you build — the folder structure, the skills, the memory, the data — that's what makes it powerful.

The Principles

This workspace is built on principles from Unix, the 50-year-old operating system philosophy. These principles weren't designed for AI, but they turn out to be perfectly suited for it.

1. Everything is a File

There is no database. No Airtable. No Notion backend. Every piece of data the AI works with is a file — JSON, Markdown, CSV.

Why this matters for AI: An LLM can read a file natively. It cannot query a database, open a Notion page, or parse a proprietary format without special tooling. Files are the universal interface between humans and AI. You can read them. The AI can read them. Anyone can inspect them.

The implication: keep your working state in files and folders — not trapped inside apps. Your meeting notes? A file. Your project tracker? A file. Your daily plan? A file. The AI can read all of them.

2. Small Tools That Compose

In Unix, you compose small programs: ls | grep | sort. Each does one thing well.

In a workspace, the equivalent is skills. Each skill is a procedure that does one thing — a Markdown file with natural-language instructions that the AI follows. Not code. Not a plugin. An SOP that the AI executes.

Skills compose the same way Unix programs do: a morning skill invokes the email triage skill and the meeting prep skill. Each does its job and passes its output forward. The system is built from small, reusable pieces.

3. Separate Data Acquisition from Presentation

The workspace keeps getting data and viewing data separate. Integrations (MCP) pull data from external sources — Slack, Gmail, Salesforce, Snowflake. The client (Cowork) reads that data and presents it.

This means:

The client never breaks because an API changed
You can swap out any data source without touching your workflow
The AI, any visualizations, and your skills all speak the same language

4. Plain Text Over Proprietary Formats

Everything in the workspace is plain text: Markdown, JSON, CSV. Not Word docs. Not PowerPoint.

Why: Plain text is the only format that works everywhere — the AI reads it without conversion, you can search it with any tool, changes are trackable, and it will still be readable in 20 years. Other tools and skills can consume it.

This doesn't mean you can never use Google Docs or Slides. It means your working state — the data the AI operates on — should be in plain text. The finished product can be exported to any format.

Memory: The Foundation

Memory is what makes this a system instead of a chatbot. Without memory, every conversation starts from zero. With memory, the AI builds up an understanding of:

Who you are — your role, responsibilities, preferences, communication style
How you work — which skills you use, what corrections you've given, what approaches you prefer
What's happening — current projects, ongoing initiatives, recent decisions, organizational context
What you've learned — patterns, principles, and rules extracted from experience

Memory is stored in text files that the AI reads at the start of each session. Over time, these files grow into a comprehensive profile. The AI stops asking questions you've already answered. It starts anticipating what you need.

In Cowork, memory works through multiple layers: Global Instructions (persist across all sessions), Projects (persistent workspaces with custom instructions and reference documents), Dispatch (a continuous conversation thread carrying context forward), and Knowledge Bases (coming soon — topic-specific persistent memory containers).

This is foundational. Without memory, skills are stateless procedures. With memory, skills get personalized. Your email triage skill doesn't just sort email — it knows which senders are VIPs, which threads you've been ignoring on purpose, and what your actual priorities are this week.

Two Modes of Work

Every day with an AI workspace has two kinds of activity:

Mode 1: Doing the Work. You use the AI to prepare for meetings, triage email, draft documents, research topics, query data, and manage tasks. Every role has dozens of these tasks — the AI does them faster, more consistently, and with more context than doing them manually.

Mode 2: Building Automation. When you notice a pattern — "I do this same prep before every 1:1" — you turn it into a skill. Now the AI does it automatically every time. The Skill Builder exists for exactly this: you describe what you want automated, it creates the skill, you test it, refine it, and it's part of your system forever.

This is the key difference from using a chatbot. A chatbot resets every conversation. A workspace accumulates. Every skill you build, every preference you teach it, every correction you make — it all persists. The system gets better every day because you're investing in the workspace, not just consuming the AI.

Agents: Specialty Systems That Learn

Agents are specialized versions of Claude that focus on a specific domain. Each agent has a defined expertise, memory that persists across sessions, skills it can invoke, and specific tool access.

Think of agents as specialist team members. You don't want one generalist handling everything — you want a pipeline analyst who knows your data warehouse, a customer ops expert who knows your merchant workflows, and a competitive intelligence researcher who knows your market.

Agents get better over time because they have memory. The pipeline analyst remembers which deals you asked about last week. The data quality monitor remembers which tables are known to be flaky. The merchant specialist remembers that a particular vendor has a recurring integration issue.

In Cowork, agents are packaged as plugins — installable bundles of skills, connectors, and sub-agent definitions. Enterprise admins can set up a private plugin marketplace, provision plugins per-user, and auto-install across the organization. Anthropic's open-source plugin repo provides a starting point at github.com/anthropics/knowledge-work-plugins.

The Daily Rhythm

The workspace supports a structured daily cycle:

Morning: AI gathers calendar, email, tasks, recent activity — writes a plan with top 3 priorities. You review and adjust.
Work blocks: Invoke skills as needed. Manage tasks. The AI maintains context across the session.
Before each meeting: AI produces a brief — Raise (decisions needed), Status (FYIs), Watch (emerging issues).
Evening: AI reviews accomplishments vs. plan, rolls over unfinished tasks, suggests tomorrow's priorities.

This cycle creates continuity. The AI reads yesterday's evening notes to inform today's morning plan. You stop losing context between days. After a week, you have a complete record of what was planned, what happened, and what was learned — without writing any of it manually.

How Traditional Setups Differ

The Traditional Knowledge Worker

10+ browser tabs: Gmail, Calendar, Jira, Slack, Confluence, Salesforce, Snowflake, Tableau, Google Docs, Sheets
Manual context switching: Copy-paste between tools, re-explain context every time
No automation: Every recurring task is done manually
No persistence: Yesterday's context is gone unless you took notes
No reflection: The day ends without structured review
Data trapped in apps: Can't search across tools, can't compose workflows

The AI-Native Workspace

One primary interface: Cowork — the AI as your interaction point
Structured workspace: Files and folders organized by domain
Skills for recurring tasks: Morning plan, email triage, meeting prep — all automated
Persistent memory: The AI remembers your preferences, your org chart, your communication style
Daily rituals: Structured morning planning and evening reflection
Connected data: MCP integrations bring Slack, Google, Jira, Salesforce into the AI's context

The fundamental shift: from app-hopping to workspace-operating. Instead of switching between 10 tools, you interact with one AI that has access to your workspace and your integrations.

Visualizations and Custom Tools

Cowork can build interactive tools through two mechanisms:

Artifacts — Claude generates HTML, React components, charts, and dashboards on demand. You describe what you want to see, the AI builds it. Artifacts are session-scoped (generated fresh each time), but skills make them reproducible: "show me this week's pipeline" produces a consistent visualization every time.

MCP Apps — Over 75 third-party tools (Salesforce, Amplitude, Figma, Hex, Slack, Asana, Monday.com) can render interactive dashboards and tools directly inside the Cowork conversation with live data connections. A Salesforce MCP App renders your actual pipeline with filters and drill-down. Hex runs SQL queries and displays charts.

You don't need a designer or developer. You describe what you want, the AI builds it or surfaces it from your connected tools. If it's not right, you correct it. The AI remembers your preferences for next time.

What Makes This Different From "Just Using Claude"

Just using Claude	Workspace-driven Claude
New chat every time	Persistent workspace with memory
You provide all context	AI reads your files and integrations
Manual everything	Skills automate recurring workflows
No history	Daily logs, task history, meeting preps
One-size-fits-all	Custom agents for your domain
Consumption	Compounding — gets better every day

The difference is between using a calculator and building a spreadsheet. Both compute numbers. But the spreadsheet remembers, compounds, and grows with your work.

The Principles (Summary)

The workspace is the product. The AI is just the engine. Invest in your workspace structure.
Everything is a file. Plain text is the universal interface between you and the AI.
Small tools that compose. Each skill does one thing well. Combine them for complex workflows.
Separate data from display. Integrations get data. Cowork presents it.
Memory is foundational. Without memory, it's a chatbot. With memory, it's a system that knows you.
Two modes of work. Do the work. Then build automation to do it faster. Both are productive.
Agents for specialization. Specialty systems with memory outperform one generalist.
Daily rituals create continuity. Morning plan → work → evening retro. The AI connects the days.
Iterate, don't architect. Start with one skill. Add more as patterns emerge.

Appendix: The Unix Principles

These principles, articulated by Doug McIlroy, Ken Thompson, and Rob Pike in the 1970s, guide this entire approach:

Make each program do one thing well. → Each skill has one job.
Expect the output of every program to become the input to another. → Skills write files that other skills read.
Design and build software to be tried early. → Start with minimal skills, iterate based on actual use.
Use tools in preference to unskilled help. → Build a skill rather than doing it manually next time.
Store data in flat text files. → JSON, Markdown, CSV. No databases.
Use shell scripts to increase leverage. → Skills are the new shell scripts.
Prototype before polishing. → Make it work → make it lean → production.

These principles are 50 years old and they're more relevant now than ever. AI agents are the new Unix programs. Your workspace is the new filesystem. Skills are the new shell scripts. The pipe is the file.