After spending a month building 250+ skills inside Claude Cowork, James hit a wall. Not because the tool doesn’t work — it works brilliantly. The problem was setup. In this tutorial, James walks through the exact mistakes he made and the mental model that turned chaos into a working AI operating system.
If you’re just getting started with Claude Cowork, or you’ve been using it but feel like things aren’t quite clicking, these are the mistakes you want to avoid — and the framework that makes everything make sense.
Mistake #1: No Folder Structure
🔴 The Problem
James started creating outputs immediately — skills, drafts, agents — without organizing where anything lived. The result: lost outputs, no clear home for anything, and agents that couldn’t find the files they needed to do their jobs.
Think of it like starting a new job and dumping every document you ever create onto the same desk with no folders, no labels, and no filing system. When something goes wrong at 2am, you can’t find anything.
James ended up with separate folders for agents, marketing, taxes, YouTube projects, and mind maps — all disconnected, all disorganized. When he needed to find something, it was buried somewhere in one of a thousand chats.
✅ The Fix
Set up your folder structure before you create your first skill or run your first agent. The Cowork “Work in Folder” setting is right there when you create a new task — use it intentionally, not as an afterthought.
A simple structure that works:
- inputs/ — raw material coming in (transcripts, recordings, briefs)
- outputs/ — final deliverables going out
- active-work/ — tasks in progress
- shared/ — business context all agents can read
Five minutes of setup now saves hours of confusion later. Agents can only use what they can find.
Mistake #2: Not Using Projects for Persistent Context
🔴 The Problem
Every time James started a new Cowork session, he had to re-paste his ICP (ideal customer profile), brand guidelines, and system instructions. Over and over. Same context, every session, by hand.
It’s like re-introducing yourself to your assistant every single morning before you can give them any work. They never remember who you are or what you do.
✅ The Fix
Use Claude’s Projects feature. When you create a new Cowork task, there’s a “plus” icon and an “Include in Project” option — use it.
Projects work like NotebookLM — you add documents to a project once, and they’re automatically available in every session that runs under that project. You can add memory, uploaded documents, and previous chat history all in one place.
Add your ICP, brand voice, system instructions, and standing context files to a Project. From that point forward, every session starts already knowing who you are, who you serve, and how you work. One setup. Zero re-explanation.
Mistake #3: Scheduling Giant Multi-Skill Agents
🔴 The Problem
James built large automated agents that tried to run 5, 6, or 7 skills in a single scheduled task. These agents kept getting stuck — stalling mid-run, waiting for permissions that weren’t there, or running out of context before finishing.
When he checked his scheduled tasks, many showed incomplete runs. Some had read a file that got stuck. Others were waiting for a permission prompt nobody was there to approve.
A big agent running lots of tasks at once is like trying to carry 10 grocery bags in one trip. Something drops every time.
✅ The Fix
Break large agents into smaller, focused scheduled tasks that each do one job — and schedule them to run at different times. The outputs from one run feed into the next.
Ask yourself: Can I split this into Task A at 8am and Task B at 9am, with the daily tasks consolidating at the end of the week?
Smaller runs are more reliable, easier to debug, and easier to restart if something goes wrong. Individual skills and single-purpose tasks are much more stable than mega-agents.
Mistake #4: Using the Web Version Instead of the Desktop App
🔴 The Problem
Claude’s web version and desktop app look almost identical — but they are not the same. Cowork mode, MCP connectors (WordPress, ContentStudio, local file system), and several critical tools only work in the desktop app.
James was trying to connect MCP connectors that were only available on the desktop, or discovering that the connector existed on desktop but had no web equivalent. Using the web version for Cowork is like trying to use a power tool with no cord plugged in — it looks right, but half the features don’t work.
✅ The Fix
Download and use the Claude desktop app. This is where Cowork mode lives. This is where your MCP connectors run. This is where local file system access works.
If you need connectors like WordPress (via AI Engine), ContentStudio, the local file system, FluentCommunity, or YouTube transcript tools — you need the desktop app. Even if some connectors exist on the web, keep everything on the desktop for consistency.
Use the web version for quick questions and one-off tasks. Use the desktop app for anything automation-related.
The Real Fix: Build an AI Operating System
The four mistakes above all point to the same root cause: treating Claude Cowork as a collection of separate tools instead of as a system. Once you see it as an operating system, the right setup becomes obvious.
James introduces a three-part framework he calls the AI Operating System:
🏢 Front Office — Consistent Context
Everything your AI needs to know before it touches any task: your ICP, mission statement, brand voice, standards, systems, and business rules. This lives in your Project (and a CLAUDE.md file) and gets loaded automatically into every session.
Think of the front office as your AI’s job description and company handbook — always visible, always applied. Every AI action inherits this context automatically.
🗄️ Back Office — Structured Data
The storage layer. Organized markdown files in a single root folder, optionally synced with Obsidian (a free tool that stores everything as markdown — the same format Claude creates). This is where outputs land, where agents find source material, and where your business knowledge compounds over time.
James is building his back office by setting a single root folder that maps to his Obsidian vault. Everything Claude creates lands there. Everything Claude needs is findable there.
📋 Business Rules
The guardrails that govern how AI acts: what it can do, what it should never do, what it should always check before publishing. These live in your shared context files and are inherited by every agent — ensuring consistency across every run, every output, every time.
Front office + back office + business rules = an AI operating system that runs consistently, not randomly. The exciting part is this: once this is set up, your agents have everything they need to work on your behalf without you having to babysit them.
Your Next Steps
- Create your folder structure first. Before your next session, set up inputs/, outputs/, active-work/, and shared/ folders in your Claude Cowork workspace.
- Set up a Project in Claude. Add your ICP, brand guidelines, and key context files. Use the “Include in Project” option when creating Cowork tasks. Never paste context manually again.
- Audit your scheduled agents. If any task is running more than 2–3 skills at once, split it into smaller focused runs at different times of day. Let daily outputs consolidate weekly.
- Switch to the desktop app. If you’re using the web version for Cowork tasks or MCP connectors, download the desktop app today.
- Draft your AI Operating System. Write a simple front office document (who you serve, your brand voice, your standards) and save it to your shared/ folder. This becomes the blueprint every agent follows.
This is Part 1 of an ongoing series on building an AI Operating System for your education business. James will be covering the full setup — front office context files, back office folder architecture, and business rule orchestration — in upcoming videos.
Join the free community at trainingsites.io to get all the resources, templates, and updates as they’re released.
