Guide: Build Your AI Team in Cowork

Guide: Build Your AI Team in Cowork

Most ecommerce operators use Claude the way they'd use a calculator : open it, ask one thing, close it. That works for quick questions, but it leaves the real leverage on the table.

Cowork is a different layer entirely. It lives on your desktop, reads and writes real files, connects to live data sources like your marketplace analytics, Notion, Gmail, and ad accounts, and runs multi-step work while you do something else. A single Cowork session is one worker doing one job. A team is what happens when you build repeatable specialists - each one owning a function, loaded with your standards, and triggerable on demand or on a schedule.

The shift is from operator to director. You stop being the person who pulls the performance report and start being the person who reviews the three decisions it produced.

This guide walks you through building that team from scratch: auditing your work, defining roles, writing SOPs, and staffing your first specialist — all in a single 90-minute sitting.

What You'll Need

Software & Access

• Claude Desktop app (download from claude.com/download)

• A paid Claude plan (Max tier recommended for production-level usage)

Connected Data Sources (as needed per role)

• Notion MCP — for inventory checks and price tracking

• Gmail / Google Calendar — for morning briefing roles

• Meta MCP — for ad creative analysis

Your Own Context

• Again it will depend on what employee you’re building

Step 1: Run the Delegation Audit

Before you build anything, sort your actual week into two columns. 

Judgment work stays with you — supplier negotiations, product launch decisions, brand positioning, channel strategy.

 Production work is delegable — pulling reports, harvesting keywords, drafting copy, tracking competitor prices, auditing listing content.

The test: if a solid SOP could let a sharp VA do it to your standard, it's production. Those tasks are your first hires.

Write the list. Two columns. This becomes your hiring plan.

Step 2: Draw Your Org Chart

Name the roles, not features. "My Customer Researcher" is a colleague with ownership. "A skill that analyzes reviews" is a feature with no accountability.

A typical ecommerce org might include a Search Analyst (keyword and funnel diagnostics across your sales channels), a PPC Analyst (weekly negative keyword harvests and wasted-spend flags), a Customer Researcher (review and feedback mining into personas), a Listing Strategist (product page copy optimized for conversion — whether that's an Amazon listing, a Shopify PDP, or a marketplace detail page), Competitive Intel (competitor listing teardowns and gap maps), a Copywriter (on-brand copy across channels — email, social, ads, product pages), an Ad Creative Analyst (Meta, Amazon, Google ad diagnostics and creative fatigue reads), and an Ops Gatekeeper (inventory and supplier checks before scaling spend).

You won't build all of them. The chart lets you see the whole org and choose which seats to fill first.

Step 3: Write the Job Description Before You Build

This is the step everyone skips and the reason most AI specialists produce generic output. You cannot build a specialist for a process you haven't defined.

Take the PPC Analyst as an example. The job description isn't "clean up wasted spend." It's a documented SOP with process steps, rules, and a quality bar:

Process: Pull 30 days of search term data from the relevant ad platform. Flag terms with 10+ clicks and zero orders. Flag terms above 2× target ACOS or CPA with spend over your threshold. Identify irrelevant terms by pattern — wrong use case, competitor brand names you don't want, size or variant mismatches. Recommend each as an exact or phrase negative with a one-line reason.

Rules: Never negate a term that has produced even one sale without flagging it. Never negate a term with fewer than 10 clicks. When ambiguous, list it under "Needs a human."

Quality bar: Every recommendation has a defensible reason. Nothing gets negated on a hunch.

The same principle applies whether you're writing a job description for a Shopify SEO audit, a marketplace listing rewrite, or a Meta ad performance summary. Specificity is the quality. Do this for every role before you build it. The roles that are easy to write down are your best first hires.

Step 4: Build the Shared Handbook

Every specialist reads the same company context before doing any work. Create a folder structure on your computer:

Cowork/
├── HANDBOOK/
├── ROLES/
├── OUTPUTS/
└── TEMPLATES/

The HANDBOOK holds three files. First, about-the-business.md — let Cowork interview you to build it:

You are building my about-the-business.md file for my Cowork handbook.
Interview me using AskUserQuestion — one question at a time, 15 questions,
push back on vague answers. Cover: what I sell and to whom, which channels
I sell on, my positioning, my margins and priorities, what a great
deliverable looks like in my world, and my hard non-negotiables. Then
compile it into one markdown file under 1,500 tokens — condensed prose
and bullets, no raw transcript. Save it in HANDBOOK/.

Second, brand-voice.md — how your copy sounds and the words it must never use. Third, the-handbook.md — your north-star goals and constraints this quarter.

Then set your global instructions once in Settings → Cowork → Edit Global Instructions:

Before every task, read every file in HANDBOOK/.
Never read OUTPUTS/ or TEMPLATES/ unless I point to a specific file.
Save all deliverables in OUTPUTS/ under a subfolder named for the project.
If the brief is unclear, use AskUserQuestion. Don't fill gaps with filler.
Deliver the work. Don't over-explain it.

Step 5: Hire in the Right Order

Do not build eight specialists in a weekend. Pick one role based on two criteria: how often you do it (weekly beats monthly) and how cleanly you could write the SOP.

For most operators the right first hire is the boring, repeatable one — the PPC Analyst, the weekly ad summary, or the price tracker. Build it, get it working, feel the hours come back. Then build the next seat.

Step 6: Build the Specialist

A specialist in Cowork is a skill with a SKILL.md file at its core — everything you wrote in Step 3 goes here. Give it access to live data through connectors (DataForSEO for marketplace and search data, Notion for inventory and SOPs, Meta for ad performance, Gmail for communications), then trigger the whole workflow from a single command like /harvest.

It reads your handbook, follows your SOP, uses your rules, and drops a clean deliverable in your OUTPUTS folder. That's an employee, not a prompt.

The rest of the guide — including the full 90-minute setup walkthrough, pipeline examples, and the limits you'll hit — you can find here, a free gift from me :)

• Amazon tops $3 trillion market cap after a Q2 cloud surge, lifting 2026 capex to $220 billion. Jassy says even that leaves AWS short of AI demand, and calls 2028 demand striking.

• The Trump administration is quietly building toward a de facto ban on Chinese AI models — and critics say OpenAI and Anthropic are lobbying for it every few months to eliminate their cheapest competition.

• Agency experts debunk five AI visibility myths — including why your old SEO playbook, website-only optimization, and GEO "hacks" won't cut it in a zero-click, LLM-driven search world.

• Meta just launched a standalone app for Facebook Marketplace power sellers — with AI-powered listing creation, bulk tools, and performance insights

• Companies went from rewarding employees for spending the most on AI tokens to "thrift-maxxing" with cheaper and Chinese models — and it's threatening the valuations of OpenAI and Anthropic right before their IPOs.

• ARK projects AI agents handle 25% of digital transactions by 2030, up from 2%, worth over $8 trillion. Protocols MCP, ACP and UCP unify the stack for 90 second agentic purchases.

  

🎨 Hedra - Turn an image plus audio into talking, expressive AI avatar videos.

🎧 Runway - Pro-grade AI video generation and editing for creators and studios.

🤖 Synthesia - Avatar-led explainer and training videos in 140+ languages.

⚙️ Decagon - AI support agents that resolve refunds, account and order tasks end to end.

📹 Crayo - Auto-generate faceless short-form videos with captions and voiceover.

😍It was great, all I needed

🙂Good, not great

🙄It sucked