How to Build an AI Chief of Staff in Claude (Full Setup)

Full video transcript

We’re going to be using Claude Cowork, training our AI chief of staff and three specialized AI agents designed to support operations and revenue growth. Welcome into today’s Build Your AI Chief of Staff workshop. We’re hiring our AI chief of staff to support us on revenue and operations. This training today is going to lay the foundation, starting with building our Atlas folder, creating the guidelines and instructions that we will be training our AI chief of staff on. Just the same way you would want any human coming into your business to know how you operate — standard operating procedures, what you like or dislike, how you speak, who your key partners are — all the things that you want your AI chief of staff to do is going to live in this Atlas folder. And then we’re going to be utilizing Claude Cowork as the body that is helping us with actually implementing the AI on top of this folder. Without context, everything else with AI gets very generalized. So much of the work here in this process is foundational — it’s setting up the folder and the architecture to hold our context so that our chief of staff can do the work that we want it to do most effectively.

So what is the Atlas folder? Picture Atlas, the Titan that is holding up the world on its shoulders. Shoulders rounded, legs braced, the whole world on its back. That imagery fits because it often feels like we carry the whole weight of the world in our business — clients, prospects, admin, decisions, inbox, calendar — inside our own heads. If that’s all laying on us as founders, leaders, executives, decision makers, that’s a lot of weight, and sometimes we can’t take it off our shoulders because we don’t have anywhere to put it. Then stuff falls apart or breaks, and too much weight for too long leads to burnout.

So the Atlas folder, in the simplest terms, is quite literally a folder that lives on our computer. It’s where all of our context, thought processes, decisions, frameworks, clients — everything we need to know about our business — gets organized over time and used as a frame of reference for building out more. You can call this whatever you want. Bruce Fleck, a longtime community member, has done additional Obsidian-related content and an intelligent workspace creation process along similar lines. Don’t get too hung up on the name Atlas — it’s just how I picture it: organizing into separate folders so the system has clear structure, then handing that off so Atlas can hold up our world while we lead — becoming an agentic CEO.

Being an agentic CEO doesn’t mean just prompting AI and re-explaining your business context over and over in a one-way conversation. It means stepping into a more dynamic role: delegating, orchestrating, and architecting the support systems that not long ago would have required hiring a virtual assistant or administrative support. I’m not trying to grow a massive team of humans or take on payroll to deliver a service I already know how to deliver — either by delivering it myself, finding a strategic partner, or building my own systems. You may want to hire and build a large team for your business, and that’s fine too, but for me it’s about being more of a solopreneur-style operator, and a big reason I don’t need to hire additional VAs is that I’ve figured out how to set up these automations with Cowork.

Here’s a day-in-the-life example of an agentic CEO. It’s a Tuesday. A client session ends at 11:00 a.m. You have another prospecting call at 11:30. You open Claude and say, “Prep me for this 11:30.” By 11:05, your chief of staff has already pulled in the prospect’s last email, prior discussion notes, their Instagram, and three questions to ask. You’re not spending 30 minutes prepping — you’re showing up fully present, not scrambling, not context-switching. You hop on the call, finish it, and your chief of staff logs the decision, updates the files, drafts follow-ups in your voice, and creates tasks in your project management system that you can commit to or delegate. By noon you close your laptop and move to your next client, and nothing fell through the cracks. This is a real example of running a Granola voice recorder during calls, integrated with Claude via MCP, then having a workflow process the transcript and prepare you for the next call. If you’re spending 30 minutes to an hour on admin work around sales and discovery calls, that time either needs to go to an assistant or to your AI chief of staff.

This training isn’t a typical software tutorial — it’s building the infrastructure and filling your Atlas folder today so you can keep building on it over the next few days without it becoming thin or creating more work. Most of what people call an “AI employee” is literally a skill.md file in Claude, and today’s plugin gives you more than a hundred to start from. We’re focused on one chief of staff plus three specialized teammates. Your chief of staff runs in Claude Desktop and Claude Cowork; some of you with the Max plan can also use Claude Code, though Cowork works fine on the Pro plan. The goal is a system that creates a kind of perfect memory for where all of your chief of staff’s knowledge about your business lives.

One real example: while working in one Claude thread and another thread was making updates in a different folder, the first thread flagged, unprompted, that something happening in the other thread might create a conflict — because the Atlas folder setup gave it real-time insight and monitoring across everything being worked on. That’s the reassurance that the system is working and staying on top of things the way it’s supposed to. This memory gets better and better the more you train and refine it over time, and it’s always available whenever you want to delegate a task, with infinite patience — no HR complaints about being told your idea needs work.

Claude is the platform being used here as the interface for working with these folders — not ChatGPT, not Gemini. The real asset being built isn’t the Claude subscription itself; it’s the Atlas folder, which is portable to whatever platform comes next. You don’t want everything sitting inside a Claude account only, in case usage limits or account issues ever get in the way — you want your business context and training to live in folders you own and can bring anywhere.

The folder is where the chief of staff reads its guidelines, gets its daily briefing, and checks in on different projects. Skills are the standard operating procedures your AI employee executes toward a defined outcome. Connectors are the MCP tools set up at the system level so your chief of staff has access to them (and you can build custom connectors too). Scheduled tasks run recurring work — daily, multiple times a day, weekly, monthly — on a cadence. Compared to having a bunch of scattered Claude/ChatGPT threads where context evaporates or gets lost, Cowork’s automations, skills, and persistent memory system create a window of access that uplevels memory and lets you delegate to and orchestrate multiple AI employees in parallel on multiple tasks at once, rather than one prompt at a time.

Examples of things you can delegate: client work — prep sessions, post-session recaps, progress notes, follow-ups, prospecting research, discovery call prep, proposal drafts, follow-up sequences; content — batching posts, repurposing content, weekly content calendars; retention — check-ins, at-risk flags, re-engagement drafts (skipping retention creates a leaky bucket); operations — Monday morning brief, weekly review, inbox triage, meeting-to-action; and research — competitive research scans, market intelligence, buyer psychology briefs. One example from the community: Adella, founder of Eterniva (featured on Shark Tank), used a weekend build to offload morning briefing, project management, and inbox triage, freeing up roughly 5 to 10 hours a week. There’s also personal/travel delegation — booking appointments, itineraries, packing, time zone prep.

The starting point for this workshop is a morning brief, connecting Gmail and calendar as an MCP connector, and mapping time sinks and revenue-generation opportunities. Requirements: the Claude Desktop app (from claude.ai/downloads), the Atlas folder setup files and plugin from the classroom, and setting up the Atlas folder itself. Local storage is recommended for speed and ease, though a Google Drive folder (with the desktop extension installed) or Dropbox also work for people whose business already lives there. Name the folder whatever you’d like — often the name of your chief of staff (Riff, in the presenter’s case), and you can always rename it later or ask the AI what it wants to be called.

Onboarding questions to think through as you start: Who are you at the bare minimum? Who do you serve? How do you sound? What are the first things you want to hand off? What are you already using daily? Treat this as training a new hire — if you hired a person and gave them no documentation, no processes, and no standards, you wouldn’t blame them for failing. Everyone bringing existing skill and experience to this can succeed by adapting these tools to what’s already working.

Project folders live within the main Atlas/AOS folder — think of the top-level folder as a window Claude can see into, with sub-folders as smaller windows that can be given focused, specific access. Obsidian is a layer on top of that same folder that visualizes it — it’s just an interface for a set of markdown files in folders; a free app you install at obsidian.md, with an optional $4-5/month sync add-on.

Walking through the install live: two Claude Desktop chat panes and an Obsidian window open side by side. A new folder was created (Maya Atlas folder, for a fictional fractional-CMO business named Maya used as a demo), the AICOS plugin zip was added to it, along with a “Maya Chen workshop creator brief” markdown file as seed business context. In Obsidian, “Open folder as vault” was used to point at that same folder, and the one markdown file already there appeared inside Obsidian’s graph view.

In Claude Cowork, under Customize, there are Skills, Connectors, and Personal Plugins. Skills are individual skill.md workflow files that can be shared, downloaded, edited, or uninstalled. Connectors are the tools within a skill — Gmail, Google Calendar, Granola, Microsoft Teams, Slack, Fireflies, and others — companies are increasingly shipping their own MCP connectors so more work can happen from one dashboard.

To install: in Customize, go to Personal Plugins, click Create/Upload, and select the zipped V3 Atlas folder plugin everyone downloaded from the classroom — the same file for everyone, customized to each business through the folder’s own context. The chief of staff acts as orchestrator, able to create, modify, and chain together any of the 100+ skills in the plugin — the point isn’t the number of skills but the outcome they’re targeted at.

Next: in Cowork, select “work in a project,” choose a different folder if needed, and navigate to the Atlas folder (in the demo, the Maya Atlas folder) rather than any pre-existing folder from a previous session — this step matters a lot on a first-time install, since picking the wrong folder means the setup runs against the wrong context. If running this for two separate businesses, you’d maintain two separate Atlas folders and run the plugin setup on each one independently.

Then type “run” with a space before the slash so the slash command registers, and run /run AIOS setup. The model in use during the demo was Opus 4.7, which uses more tokens; Sonnet 4.6 is suggested for Pro-plan users who want to conserve usage early on. Watching the right-hand panel, Claude begins reading the CLAUDE.md file and the skills markdown files, pulling in the seeded business context. (Note: global instructions from the presenter’s own account may have still been influencing this particular run since they hadn’t been reset for the demo.)

Claude then runs a series of onboarding questions: What are the top three things you’d hand off to an employee on Monday? (In the demo: following up on client proposals, keeping track of the different teams managing agency projects, and helping with client deliverables from a market research service.) What’s the one thing that, if systematized this week, would create the most leverage? (In the demo: buying back time lost to email inbox and client/team communications.) Where should the memory backend live — local folder, Google Drive, or Notion? (Local was chosen, with Obsidian layered on top.) Then naming the three specialized agents — time buyback, retention, and revenue generation — plus an optional request for writing samples (emails, blog posts) to learn your voice, which can be skipped in favor of defaults.

As setup proceeds, Claude begins actually creating the individualized folder structure — described as “Tier one connectors,” start-up folders for one teammate and three agents inside the Atlas folder — building phase by phase. New folders and README files appear in real time on the right-hand panel, since Cowork can create folders and files just like a person clicking around a computer. Some cross-contamination from the presenter’s own pre-existing global instructions shows up in this particular demo run (e.g., references to Gen AI University), which is expected when not starting from a completely blank account.

Once connectors like Gmail and Calendar are already set up, Claude can reference them directly during setup — in the demo it referenced the actual live workshop happening on the calendar. First-time users without connectors configured won’t get this until they add them, which is recommended as part of the foundational setup.

Claude then creates named, scheduled agents. In the demo: Mercury (time buyback) — a Monday buyer pulse and a Friday weekly review, plus a daily morning brief at 9:00 a.m. Central time on weekdays; June (retention/LTV) — a Tuesday 10:00 a.m. client check-in pass, which can be extended with skills like tracking renewal dates for clients who subscribed a set number of months ago; and ACE (revenue generation) — a Monday 11:00 a.m. Central time buyer pulse. Useful commands surfaced during setup include requesting a morning brief on demand, a quick status check, and a “save session” command that runs a memory-saving skill at the end of a session.

Not everything needs to be built and tested in one sitting — it’s fine to focus on one agent first. The suggested framing: if your AI chief of staff showed up Monday morning to help, what would that support look like? What’s taking two, three, five, or more hours of your time that you’d love to offload? One community member, Hannah, went from spending about 10 hours per client deliverable (branding and marketing work) down to about an hour after operationalizing the process — freeing capacity either to take on more clients or redirect focus elsewhere.

On additional connectors: for Outlook, Microsoft 365 was suggested as covering SharePoint, OneDrive, Outlook, and Teams. Other available connectors mentioned include Glean, HubSpot, Slack, Miro, and Booking.com, among many others — the guiding question remains what would be useful if help showed up and supported you on a given Monday morning.

On global instructions: under Settings > Cowork > Global Instructions is where preferences, conventions, and context that Claude should always know are stored — separate from the Atlas folder’s business-specific content, and applied across every session and folder. The presenter’s own global instructions include a memory-backend note (Notion write-back as mandatory in his setup, subject to updates as needed) and a personal “tier zero override”: never let work override time with his wife Rachel, family, friends, or music, because the business is designed to serve life, not the other way around. This can be run and edited through a /setup-global-instructions command, and can direct Claude toward routine actions (competitive analysis, pulse checks, whatever matters to your workflow) at the start of every session.

In response to a question about whether global instructions apply across all of Claude, not just one folder — yes, and if you run multiple businesses or projects, global instructions can direct Claude to be aware of that and point sessions at the right folder(s) accordingly, or even reference multiple Atlas folders in the same session when a task spans more than one business.

Another feature mentioned is “dispatch,” available in both Cowork and Claude Code, which — with computer use/browser actions enabled — can sync a phone with a running Claude session so a task can be handed off remotely, letting Claude continue working while you’re away from the machine.

Homework and action items: the foundation is built — the Atlas folder, the plugin, the chief of staff, three specialized agents scaffolded, and Obsidian layered on top for visualization. The assignment is to identify what the AI chief of staff should do first that will be most effective — building out further from the simple morning-routine setup already in place — focused on either buying back time or generating more revenue without hiring additional people, using the tools already covered (skill delegation, connectors like Fireflies, Granola, and Fathom for transcription) so nothing falls through the cracks over the next few days.