2/11/26: Claude Cowork Setup: How One Founder Replaced $14K of Agency Work in 30 Minutes
Just as a reminder, we have CHANGED FORMATS, which we are really excited about, where we have practitioners and GTM leaders/professionals showing off workflows that they are using right now to help you know what is working.
Just in case you missed it, we had last week going over how to use Claude Code to scrape podcasts for competitive research.
Week before we dove in with Mindstudio.ai for their business development and how they use AI agents.
Make sure that you participate in the quick 5-10 min AI Led survey for the GTM AI 2026 report, will be sharing all the research in a couple weeks when we have enough respondents.
Now lets dig in, who would have thought how much intel podcasts that your competition either hosts or been on has in store for you? Lets dig in.
You can go to Youtube, Apple, Spotify as well as a whole other host of locations to hear the podcast or see the video interview.
Stop Building AI Workflows. Start Running an AI Company.
Steve Cunningham killed his dream job on purpose. He was the best in the world at reading business books and turning them into actionable summaries. Then ChatGPT showed up and did it in 5 minutes. Instead of clinging to the old model, he burned it down, rebuilt, and now runs an operation where 30 minutes of his time replaces $14,000 and 70 man-hours of work. The gap between people experimenting with AI and people operating with AI just became a canyon.
I brought Steve on the GTM AI podcast because he was one of my earliest AI mentors in my AI journey. Three years ago, when I was scraping together every piece of AI education I could find, his program was one of the ones that rewired how I thought about this technology. Not as a tech guy. As a marketer. That perspective shaped everything I built afterward. What he showed me on this episode is the most practical, most immediately executable agent setup I have seen anyone demonstrate publicly.
Here is what actually matters from the conversation, broken into two parts. First, the core insights and the quotes that frame them. Second, a full breakdown playbook so you can build this yourself today.
PART 1: THE CONVERSATION REVIEW
1) Your Business Just Got Eaten. You Just Don’t Know It Yet.
Steve’s company, Read It For Me, was a 10-year business built on his superpower: reading business books and summarizing them better than anyone. He was probably the best in the world at it. Then AI arrived and did the same thing in minutes.
His words: “I realized very quickly that I could get a decent book summary in about 5 minutes, where in the past, it would have taken me 8 hours to read the book, and make notes and be super detail-oriented about it. And this AI thing, almost immediately was as good or better than I was.”
And then the kicker: “It didn’t take me long to figure out that if I could figure out how to do that, everybody could figure out how to do that, and once everybody figured out how to do it, the business would be very easy to replace.”
This is the part most GTM leaders skip over. They hear the AI disruption story and think it applies to other industries, other roles, other companies. Steve lived it. His 10-year business got swallowed. Not overnight, but steadily. The lesson is not that AI kills businesses. The lesson is that if AI can replicate your core value proposition, the clock is already ticking, and you do not get a notification when the countdown starts.
He even said something that punched me in the gut: “If I could take a time machine back to November 2022 and destroy all the AI, I would still do it, because that job, which allowed me to read books for a living and share what I was learning with people, as a lifelong learner, that was the best I could imagine myself doing for the rest of my life.”
The man loved his job. AI killed it anyway. That is not a cautionary tale. That is a blueprint for urgency.
2) The Org Chart Thinking Lens: Treat AI Like a Company, Not a Tool
This is the insight that separates Steve from 99% of the AI creators I talk to. He stopped thinking about AI as a tool. He started thinking about it as a company.
He referenced OpenAI’s progression framework from a few years back: chatbots, then reasoning models, then agents, then innovators, then organizations. At the time, the idea of AI running an organization sounded absurd. Steve’s realization: “What if I just set this up like I’m running a company with agents? And that’s basically how I’ve been thinking about it for... it feels like months, but it’s been about a week, and it’s transformed the way I am able to operate every single day.”
One week. Not months of elaborate planning. One week of reorganizing his folder structure and mental model around the metaphor of running a company, and his productivity exploded.
His setup inside Claude Cowork (which sits on top of Claude Code for non-technical users):
An “Operate” folder as the root
A COO file that onboards the AI to every new task
Functional folders for each area of the business (marketing, product, etc.)
Inside each function: advisors (for decision-making), agents (for execution), context files (brand voice, strategy docs), design systems, templates, and a projects folder
Every time he starts a new task, he tells the AI: “Go read the COO folder and get yourself onboarded.” The AI reads its instructions, understands the business context, and then dives into the specific functional area. If he is doing marketing work, the AI then reads the marketing context files, design system, and templates before touching anything.
This is not fancy prompting. This is organizational design applied to AI. And it works because it mirrors how actual companies operate: hire someone, onboard them, give them the department playbook, and let them execute.
3) HTML Is the Language of the Humans Plus Agents Economy
This one caught me off guard. Steve made a compelling case that the file format wars are already over, and HTML won.
His argument breaks down like this:
Pre-AI, knowledge workers lived in Microsoft Office. Word docs, PowerPoints, Excel sheets. Those file types are terrible for LLMs. There is a massive amount of messy code underneath the presentation layer, and AI cannot parse it efficiently.
Software engineers gravitated toward Markdown files because that is how they work. But Markdown is bad for humans. Hard to read, no images, no flow diagrams, no visual hierarchy.
Steve’s conclusion: “We need a language that is both good for large language models and is good for human beings. And that language happens to be HTML. It works everywhere. The agents can read it incredibly well.”
He backed this up with a real-time observation. An Anthropic team member had just posted about using HTML to increase the bandwidth of communication between humans and models. Claude Code was building HTML files so humans could understand code architecture through diagrams and visual layouts. The model providers are already validating this direction.
The practical unlock: Steve built a drag-and-drop publishing tool. Create an HTML file with AI, download it, drag it into the tool, and it is live on the internet instantly. No WordPress. No dev team. No CMS friction. Just idea to published page in minutes.
He told a story about watching someone on Twitter build a beautiful presentation in HTML, and then the punchline was that they exported it to PowerPoint. His reaction: “You just did all this work to do it in the exact right way, and then you did a lot more work to make it worse, for humans and for agents.”
For GTM teams, this is massive. Custom proposals, landing pages, internal documentation, training materials, competitive battlecards. All of it can be HTML. All of it can be created by AI, read by AI, edited by AI, and consumed by humans with full visual fidelity.
4) The Invoice Method: Quantify Your AI Output or You Will Undervalue It
Steve does something at the end of every AI work session that I have not seen anyone else do. He prints a receipt.
He asks the AI to calculate what the work would have cost if done by external contractors or agencies. Line items. Hourly rates. Timeline estimates. Total project cost.
One example he showed: designing the entire interface for his new Humans Plus Agents platform. 30 minutes of his time. The receipt: 70 man-hours, 2-3 week project timeline, $14,000 in estimated external cost. Actual cost: 30 minutes and about a dollar in compute.
Why this matters: “One of the things that this enables you to do is to start to understand the scope of how much work you can actually get done. And if you think in terms of things that would have taken weeks, I can get done in 30 minutes, you start to think about how the work gets done, which rewires how you plan, and how you execute.”
He treats every day like a quarter. Every task like a project. That mental reframe only works when you have the receipts to prove the output is real. Without quantification, you default back to old planning cycles. With it, you compress your strategic planning horizon from quarterly to daily.
5) The Continuous Improvement Loop: Your System Gets Better While You Sleep
This was the most operationally impressive part of the demo. Steve built a self-improving system.
At the end of every work session, the AI does three things:
First, it generates the invoice (the receipt we just talked about).
Second, it creates a wrap-up that documents what got done, what system improvements were made during the session, and what the AI recommends improving even if Steve did not catch it during the work.
Third, and this is the unlock, Steve sends the AI off overnight to implement those improvements. He wakes up every morning with a better system than he had when he went to sleep.
His words: “At the end of every day, I’ve got a list of here are all the things we can do to actually make our company, my company of agents, better for tomorrow morning. And then I have, at the end of the day, the AI go off and work on that. So even overnight, the system is getting better so that every day, it’s like a next version of our company.”
He compared it to running quarterly cycles: “It almost feels like it’s a new quarter, because we’ve done a whole cycle of projects, we’ve accomplished a whole bunch of work, we need to regroup, make the whole thing better for the next time.”
The compounding effect of this is staggering. If your system gets 5% better every day, in 30 days you are operating at a fundamentally different level than when you started. Most teams improve their AI workflows once a quarter if they are lucky. Steve’s improve every night.
6) The Three Things That Matter in the Humans Plus Agents Economy
When I asked Steve what is actually new versus just doing old things faster, he dropped a framework that is worth the entire episode.
Three things that will matter:
Distribution. You can build anything overnight. A full Instagram clone, a custom CRM, whatever you want. But if you have zero people to talk to, it does not matter. “The product is not... the distribution is one of the things that is worth spending a lot of time on right now.” In a world where production costs approach zero, the scarce resource is attention and audience.
Deployment and feedback loops. Get something deployed at a company. Get your hooks in. Then listen. “It’s like you’re in perpetual beta. Or perpetual lean startups.” The old model was build the perfect product, then sell it. The new model is deploy something useful, get feedback, ship the improvement overnight, and repeat. Steve said he can have a feature request built and deployed by the next morning. The companies that can run this cycle fastest win.
Data. “The only thing that’s going to be about ever in this humans plus agents economy is going to be the data that you collect while people are doing that work.” Steve is building tools that track how work gets done: manual, with SOPs, or with agents. That data creates a pipeline for identifying what work should be automated next, and it becomes proprietary intelligence that competitors cannot replicate.
7) The Bezos Question Revisited: What Will NOT Change?
I asked Steve to revisit the Jeff Bezos question he used to reference in his training: instead of asking what will change in the next 5 years, ask what will NOT change.
His answer was deceptively simple: “People and companies want their problems to go away. The understanding of what the problems are and how you can utilize whatever technology is available to you to solve them is how you build a business.”
And then he said something that reframed the entire conversation: “Everybody gets so caught up in Gamma, and how do we make better PowerPoints. Nobody cares about making better PowerPoints. What they care about is communicating ideas better. And so, if you can find a way to do those things for companies, then you’re going to be fine.”
Tool obsession is the trap. Problem obsession is the strategy. The tools change every 3 weeks. The problems are eternal: make more money, save more time, reduce more risk, communicate more clearly. Steve reminded me that the people who win are not the ones who master every new AI tool. They are the ones who never lose sight of the problem they are solving.
8) The Uncomfortable Truth About Your Team
Steve said the quiet part out loud: “I could lay off half of our company tomorrow, and it would have no impact. And I’m not going to, it’s not the reason why we’re doing this, but it’s also like, they don’t have work to do. The agents are doing the work now, so what do we do?”
That is not a threat. That is a design challenge. If agents are absorbing execution work at this pace, the question is not whether your team structure changes. It is how fast you can redeploy people into work that agents cannot do yet. Steve’s answer: turn the team into a research lab. Use them to discover new products and services, test new value propositions, and run the feedback loops that agents cannot run alone.
The companies that figure out the human-agent team structure first will have a 12-24 month head start on everyone else.
PART 2: THE BREAKDOWN PLAYBOOK
Everything below is designed to be executable today. Not next quarter. Today. You do not need to be technical. Steve learned to use the terminal over the holidays. If he can build this, you can build this.
Step 1: Set Up Your Operating Folder Structure
Time required: 45-60 minutes
Tool: Claude Cowork (or Claude Code if you are comfortable with a terminal)
Create a root folder called “Operate” (or whatever makes sense for your business). Inside it, build the following structure:
COO file (HTML or MD): This is the master onboarding document for every AI session. It should contain your company overview, your role, what you are working on right now, and the folder structure so the AI knows where to find everything. Think of this as the new hire packet you would give a chief of staff on day one.
Functional folders: One for each major area of your business. For GTM leaders, this usually means Marketing, Sales, RevOps, Product, and Customer Success. You do not need all of them on day one. Start with the one where you spend the most time.
Inside each functional folder, create these subfolders:
Advisors: decision-making frameworks, competitive intelligence, strategic guidelines
Agents: specific instructions for recurring tasks (email drafts, reporting, content creation)
Context: brand voice docs, target persona descriptions, positioning statements
Design System: visual standards, color codes, typography rules, layout preferences
Templates: formats for common outputs (proposals, reports, presentations, emails)
Projects: where all active work happens, one subfolder per project
Why this works: You are not just organizing files. You are creating the organizational memory that makes every AI session start at 80% understanding instead of zero. The AI reads the COO file, understands the business, navigates to the right function, reads the context, and executes with full awareness of your brand, strategy, and standards.
Common mistake to avoid: Do not try to make this perfect on day one. Steve’s first week was, in his words, “a disaster.” His second week was better. By week three, it was transformative. The structure evolves. Start messy, improve daily.
Step 2: Build Your Onboarding Protocol
Time required: 20-30 minutes
Tool: Any text editor or the AI itself
Write a COO onboarding file. This is the single most important document in your system. Every AI session starts by reading this file, so it needs to contain:
Company context: What do you do, who do you serve, what stage are you at, what are the key metrics
Current priorities: What are the top 3-5 things you are focused on this quarter (or this week, since Steve treats every day like a quarter)
Folder map: Where everything lives so the AI can self-navigate
Working style preferences: How you like outputs formatted, what tone you prefer, what mistakes to avoid
Quality standards: What “good” looks like for your business
Here is a starter prompt to generate your first version:
“I need you to help me create a COO onboarding document for my AI operating system. Here is my company context: [describe your business, role, and current priorities]. Create an HTML file that an AI agent would read at the start of every work session to understand my business, my standards, and where to find the relevant context files. Organize it so the AI can quickly orient itself and begin productive work immediately.”
Pro tip from Steve: Use Whisper or another dictation tool to talk through this instead of typing. Steve said, “Whenever I’m typing now, I feel like I’m living in the 1950s. When you’re typing, you want to get the words right. But you don’t do that when you’re talking.” Dictation lets you do free-form conscious thinking without the friction of perfecting every word.
Step 3: Switch to HTML for All AI-Created Documents
Time required: 0 minutes (it is a decision, not a task)
Tool: Any LLM
Stop asking AI to create Word documents, PowerPoints, and PDFs as your default output. Start asking for HTML.
Why:
LLMs are trained on the entire internet, which is built on HTML. They generate better HTML than any other file format.
HTML is readable by both humans (open it in any browser) and AI agents (clean, parseable code).
HTML supports full visual design: images, charts, flow diagrams, interactive elements, responsive layouts.
HTML is instantly publishable. No conversion, no CMS, no dev team required.
Where to start:
Proposals: Instead of PDF proposals, create HTML proposals. They can include analytics tracking (Google Analytics on a proposal page tells you exactly how long prospects spend reading each section).
Internal docs: Strategy documents, meeting summaries, project briefs. All HTML.
Training materials: Course outlines, onboarding guides, SOPs. All HTML.
Design mockups: Steve designed his entire platform interface as HTML files generated by Claude Cowork.
The publishing unlock: If you need to share HTML files externally, you need a way to host them. Steve built a simple drag-and-drop tool for this. You can start with GitHub Pages (free), Netlify (free tier), or even just emailing the HTML file (it opens in any browser). The friction of publishing is the last bottleneck, and it is already solved.
Step 4: Implement the Invoice Method
Time required: 2 minutes at the end of each work session
Tool: Your AI agent
At the end of every significant AI work session, ask the AI to generate an invoice. Here is the prompt:
“Generate an invoice for the work we just completed. Include: line items for each deliverable, estimated hours if done by a human contractor or agency, estimated cost based on market rates, total project timeline if done traditionally, and the actual time and compute cost for this session. Format it as a receipt I can save.”
Why this changes your behavior:
When you see that a 30-minute AI session replaced $14,000 and 3 weeks of agency work, your brain starts planning differently. You stop thinking in quarterly cycles. You start thinking in daily sprints. You stop gatekeeping projects behind budget approvals. You start shipping.
This also gives you ammunition for internal conversations. If you are trying to get budget for AI tools, showing a stack of receipts with real dollar comparisons is more persuasive than any vendor pitch deck.
Track this over time. Create a simple spreadsheet or running document that logs: date, project, your time invested, AI compute cost, estimated traditional cost, estimated traditional timeline. After 30 days, the ROI story tells itself.
Step 5: Build the Continuous Improvement Loop
Time required: 5-10 minutes at the end of each day
Tool: Your AI agent
This is what separates Steve’s system from everyone else’s. At the end of each project or work session, ask the AI to generate a wrap-up with three sections:
Section 1: What we accomplished. A clear list of deliverables and outcomes from the session.
Section 2: System improvements made during the session. If the AI went off-track during the work and you corrected it, you should have also asked it to update the relevant context documents so the mistake does not repeat. This section logs those fixes. Steve’s exact approach: “If I’m in the middle of doing something, I find something that I don’t like, I’ll say, go and change that, but also go and change the context documents, because clearly you didn’t understand the instruction. Go find out why you didn’t understand the instruction and fix that.”
Section 3: Recommended system improvements. Ask the AI to identify improvements you did not think of. “Give me any system improvements that we didn’t think about during this session.” This generates a punch list of refinements to your context files, templates, agent instructions, or design system.
The overnight upgrade: At the end of your work day, take the recommended improvements from all your sessions and ask the AI to implement them. Update the context files, refine the templates, improve the agent instructions. When you wake up, your system is better than when you went to sleep.
The compounding math: If your system improves 5% per day (a conservative estimate based on Steve’s results), in 14 days you are operating at roughly double your starting efficiency. In 30 days, you are at a completely different level. This is not theoretical. Steve said his productivity from week one to week three was not even comparable.
Step 6: Reframe Your Team Structure
Time required: This is a strategic exercise, not a one-day task
Framework: Distribution, Deployment, Data
Steve’s three priorities for the humans-plus-agents economy give you a framework for thinking about where human effort should go:
Distribution roles: Who on your team is building audience, creating content, developing partnerships, and expanding reach? In a world where production costs approach zero, the people who control distribution control the business.
Deployment and feedback roles: Who is getting your products and services in front of customers, collecting feedback, and translating that into the next iteration? The feedback loop cycle time is now measured in hours, not quarters. You need people who can run that loop at speed.
Data roles: Who is tracking how work gets done, identifying what can be automated next, and building the proprietary data layer that becomes your moat? Steve is tracking manual work, SOP-driven work, and agent-driven work to create a pipeline of automation opportunities.
The uncomfortable audit: Look at your current team. How many people are doing execution work that agents can now do? Steve said he could lay off half his company with no impact on output. He chose not to, because the strategic play is redeploying those people into distribution, deployment, and data roles. But the audit is necessary. If you do not proactively redesign roles, you will reactively scramble when the output gap becomes undeniable.
Step 7: Apply the Bezos Filter to Every Decision
Time required: 0 minutes (it is a mental model)
Before you invest time in mastering a new AI tool, ask: “Am I solving a problem that will not change, or am I chasing a tool that will be obsolete in 3 weeks?”
Steve’s version: “Nobody cares about making better PowerPoints. What they care about is communicating ideas better.”
The filter:
Will customers still need this problem solved in 2 years? If yes, invest.
Is this tool the only way to solve this problem? If no, stay problem-focused and tool-agnostic.
Am I learning a skill (problem-solving, system design, feedback loops) or learning a feature (this button in this app)? Skills compound. Features expire.
The Bottom Line
Steve Cunningham went from having his dream job destroyed by AI to building an operation where 30 minutes replaces 3 weeks. The difference was not technical skill. It was a mental model shift: stop treating AI as a tool and start treating it as a company you are building and running.
The playbook is not complicated:
Organize your AI environment like an org chart
Onboard your AI like you would onboard a chief of staff
Use HTML as your universal file format
Quantify every AI session with an invoice
Improve your system every single night
Redeploy your human team toward distribution, deployment, and data
Stay obsessed with problems, not tools
The gap between people experimenting with AI and people operating with AI is no longer about access to technology. It is about operating system design. Steve built his in 3 weeks. You can start building yours today.
The people who move first on this get 6-12 months of competitive advantage before it becomes table stakes. And based on how fast things are moving, that window is closing faster than anyone expects.
Steve Cunningham is the founder of Humans Plus Agents (formerly Simple Academy) and has spent 3+ years helping companies adopt AI. He is also the creator of Read It For Me. His Black Belt AI Workflow Engineering training is available through a special link for GTM AI Academy listeners (check the show notes).



Thanks for doing the breakdown of Steve's Cowork Setup. It looks to be a winner.