AI Workforce Clinic · Members-only session recap
Everybody's asking me the same question this month: what's the difference between a prompt, a skill and an agent, and which one am I supposed to be building? Friday I answered it in sixty seconds, and the room spent the other forty-nine minutes showing me why the answer was the easy part.
"You guys just left me behind."
Daniel said it at minute thirty-five, laughing, from the intersection where he says he was sitting while the rest of us drove off. 34:57 He was right, and it was the most useful sentence of the session.
Here's what set it up. Mike had asked the room how everybody actually talks to their agents. VS Code? Terminal? Cowork? The desktop app? 32:33 Kristin walked through her whole setup: ChatGPT Pro open in Chrome all day, Claude's Chrome extension beside it, Claude desktop, and only Codex or Claude Code when she's talking to her agents, because Codex told her to stop chatting with them anywhere else. Then Daniel said the thing half the people reading this are thinking. He hasn't opened Claude Code or Codex at all.
Mike gave the explanation I wish I'd had a year ago. Claude Chat and Cowork can see your agents. They can read the rules. Then they do their own thing and pretend to be the agent. Point Claude Code at the agent folder as its root and it follows the rules. The catch is the word "code." It scares people off the one rung where the agents actually live. 35:22
That's when I did the sixty-second version. 36:25
A prompt is a single problem. You have it once, you ask, you get an answer. 36:38
A skill is the same problem showing up five times a week, or five times a day. You write the method down once and invoke it every time. 36:50
An agent is the third level. It needs to do the job five times a day and make decisions along the way that you didn't pre-program. It has guidelines and constraints, a goal it works toward on its own, and a loop, so it gets better every day from the outlier cases it hits. 37:06
The tools line up the same way. Chat is for prompting. Cowork is for skills. Claude Code and Codex are for agents. 38:10
Now look at who was in the room. Daniel's projects live in Claude's cloud, and he told us there's a chasm in his head between a project and an agent. 45:37 Kristin has four agents she built by hand and knows cold, plus a set she downloaded from somebody's workshop that she describes with one question: who are you? Mike has nine agents and talks to exactly one of them, and on Wednesday his developer started building the formal version of that for a customer who wants a team now. I've spent the week one rung above that, mapping how a company shares agents without every employee building their own. And Natalija is working the rung underneath all of it: getting the humans to climb at all.
One ladder. Five people. Every one of them standing on a legitimate rung. Does that make sense? The rung you're on this week is the right rung. The only question is whether you've hit the same problem five times. When you have, you're ready for the next one.
One more thing from the show-and-tell, because it decides where most people stall. The skills I use most are the skills I built. I downloaded a pile of them from GitHub because people said they were great, and I don't have their use cases. Kristin said the same about her agents: the ones she built, she knows how to use. 42:40 Your fifth repetition of your own problem is worth more than anybody's library.
The play · Climb one rung
Screenshot this one. It's the whole sixty seconds in a grid. 38:34
| Prompt | Skill | Agent | |
|---|---|---|---|
| The job | One problem, once. | The same problem, five times a week. | A goal, with decisions along the way you didn't pre-program. |
| What you hand it | The problem. | Your method, written down once. | Guidelines, constraints, a goal, and the connections it needs. |
| What it decides | Nothing. You do. | Nothing new. It repeats your method. | The calls inside its guardrails, on its own. |
| Where it lives | Chat | Cowork | Claude Code or Codex |
| How it gets better | You re-prompt. | You edit the skill. | It learns from the outlier cases, every day. |
| Time to climb when | You catch yourself typing it again. | It needs to decide something, or reach a tool. | Somebody else on your team needs it too. Now you need governance. |
Daniel The Roofing Guys · Tulsa
PropsFor saying it out loud. "You guys just left me behind." That sentence bought the room the clearest ten minutes of the session, and he was taking screenshots of Kristin's folders the whole time she talked. 34:57
ShippedHis website's phone number finally rings somewhere. He found out last week it had never been wired to ring at all, fixed it, and now they're busy. 10:54 He also ran one Facebook ad for thirty days once, to say he'd done it, and he's planning the next campaign the way I built mine on Friday, on the $20 ChatGPT plan he already has.
Where he isProjects in Claude's cloud, on the Cowork side, and they aren't agents. His words: maybe there's a chasm in my mindset between a project and an agent. 45:37 He's risk-preferred and says so. He tells Claude to send the email, and he gives it the browser. If something breaks he'll fix it.
Last week he saidDolores maturity report. Nobody raised it Friday, me included. Carried.
The homeworkOpen Claude Code. Have it create the project and the folders from what's already in the cloud, and one routine that runs every day. 48:47 His own summary: "I need to let Claude create the folder systems." We're not doing this by hand. Let Claude do it.
AlsoHe bought StoryCraft for Disruptors. It arrived Friday. The Magnetic Messaging Framework is in it, if you ever want to fix your story. 11:38
Kristin Marketing and web agency
PropsFor a show-and-tell she offered before anybody asked, aimed straight at Daniel. And for opening with "I don't have as much to share this week" and then giving the most useful ten minutes of the hour.
ShippedShe hired people for the first time. AI built the templates and SOPs she trained them with, and one of them now runs review, project management and task triage over the others. Her line: give the people the keys. 20:40 Her VA suggested the next SOP herself: how the team engages with AI.
The play Kristin gave away · Her folder system
A downloadable "AI brain" from a workshop, rebuilt around her own business, and she said so. 38:45
Last week she saidParse the eleven. Friday's version: the four she built, she knows. The downloaded ones get one question, who are you. She's leaning toward cutting them and rebuilding when a real use case shows up. That's the parse, in progress. 43:04
Wrestling withHeads-down on client work while the new team learns. She said this room is what makes her step back and look at the big picture. 22:26 My answer, for the record: no pressure to use AI here. The trust you're building with people and the trust the AIs are earning with you both take time.
NextA cleaning-service client wants a hiring campaign on Meta, and she's testing the Astra move on it. 05:58 She's dropping a screenshot of the folder system in Slack for Daniel. And she pitched me a video series: this is the common problem, this is how I solve it, starting with the chief of staff, then the content writer trained on your people.
Mike D&M Solutions · Software
PropsFor the clearest explanation of the hour, from the driver's seat with the camera off. Chat and Cowork see your agents and read the rules, then do their own thing and pretend. Claude Code, pointed at the agent folder, follows them. And "code" scares people.
Where he isNine agents. He talks to one, the coordinator, and reminds it how to use certain groups because of drift. 29:02 Right now the sharing runs on OneDrive replication. It works, and there's a more formal way with better performance, so on Wednesday his developer got the task to build it: an agent team that functions company-wide. A customer heard the plan, said it sounds great, and doesn't want to wait. Next week Mike builds his team.
GovernanceHis answer to my rabbit hole. A baseline set of rules at the company level. Employees build inside the boundaries. The company decides how far any agent can go. 30:56
What I told himThe risk in spinning up agents for a client is that the client won't like them, or won't know how to work with them, and you end up babysitting every month. Kristin's take is the one I'd bet on: build them together. That's where the value is. 43:25
Last week he saidBook the 1:1 and bring the MCP notes. Tuesday's open session was the two of us, and it turned into the multiplayer conversation that produced this week's blog post. I'm counting it. 22:46
Natalija Business development
PropsShe named the theme of the day before I did, and I said so in the room.
ShippedHer engineer is retiring and the company needs a plan for what AI covers. She asked Claude, ChatGPT and Grok to each recommend a plan for her CEO, and they agreed with each other and with what I'd found. 04:15 Marketing copy for a national grocery association's women's group, outcomes first, with Claude as the coach. Image prompts written by Claude, dimensions included, rendered by ChatGPT. What used to take her marketing person days now takes a minute.
The frictionClaude goes overboard. Ask for one thing, get extra tools, a phone version, Samsung versus iPhone. "Slow down. I didn't ask." 15:01 ChatGPT's version: change one thing and it changes three.
The challengeHer CEO, her CFO and the owner's son have real problems Claude could solve, and she can't get them to spend one hour on a Tuesday. 13:37 Thirty years of convincing risk-averse retailers to try something new taught her the move: take people by the hand. It's going to be okay. Follow me. To a CFO, in his language: one hour, and you'll see ten times back. She won't do the work for them, and she's right. It's their job.
What the room gave herRun a hands-on workshop for the three of them. I'll sit in the passenger seat. 19:22 Her sharper version: solve one of their problems in a session, bring it back, and say now you show up. And they pay her for it.
Last week she saidPost Ruben Hassad's usage chart in Slack. It hasn't landed yet. Carried.
This group has been word of mouth for months, and word of mouth only goes so far when I don't broadcast the AI work anywhere but TikTok. This week I decided to spend money to let people know the Clinic exists. 01:44 I don't know which buttons to click in Meta Ads Manager, so I told ChatGPT (Astra) my goals and asked it to take the browser. It wrote the ad copy and the headline, used one picture of my nine agents and a video I uploaded, set the audience and the budget, and stopped at the publish button for me. Three people registered for Tuesday by Friday.
I ran the same job in Claude first. Slower, clumsy, and it kept fighting me for the tab. ChatGPT worked in its own tab while I worked in mine. 08:27 For browser work, right now, it's five to ten times faster. Next month it might flip. It's a leapfrog game, and I'm fine plugging either one in.
Mike and I spent Tuesday on one question: how does a team share AI employees without everyone building their own? 22:46 I wrote it up on the site: How can our team share AI employees without everyone building their own version? The short form: every agent has a human owner, like an org chart. Governance says which humans can change which agent, because one person's improvement to the invoicing agent can break it for everyone else. An SOP for how humans and agents work together is a requirement everybody signs. Onboarding runs both ways: new people learn what each agent does, and the agents learn which human to bring which question to. I had Claude research the best practice companies are deploying. Its answer: it's the Wild West. There is no approved model. Everybody's fending for themselves.
I promised a special section this week, because the ladder describes the tools, and the tools were the easy part. Every hard thing in the room on Friday was a human changing how they work.
I'm almost resigned to it. This is who we are as a species: very, very slow to change how we do things. 18:06 Natalija has thirty years of proof from retail, and her method is the one that works. Take people by the hand. Tell them it's going to be okay. Walk with them.
Kristin's biggest wins this week came from her human team. AI wrote the SOPs and the templates, and the people took the keys. 21:08 The next SOP on her list came from her VA, unprompted: how the humans on the team will engage with the AI. I told her the thing I believe. The trust you build with people and the trust the AIs earn with you both take time, and you get your time back through both.
That's the whole multiplayer problem in one agency. Put a human in charge of every agent. Decide who can change what. Onboard the people to the agents and the agents to the people. 24:06 Every conference room in every company is having this conversation right now, and it's going to take years, because the capability is already here and the people are still people.
This is change at the pace of how people and humans change.
Greg, minute 26 26:19
Mike's customer wants a team next week, and here's my caution, which I said to Mike's face: an agent built for someone who wasn't in the room when it was built goes unused, or gets babysat. Kristin's answer is the one I'd bet on. Build them together. The hands-on part is the value. 43:25
The ads are running, so Tuesday will have new faces asking what to install and where to click. Be gentle. Getting somebody from that question to a five-agent workforce takes a few weeks, and it only happens if we meet them on the rung they're standing on. Everyone has to start where they are. I feel a calling to help people do this. 31:51
Three moves from Friday's room, every one of them aimed at humans:
Nobody raised these Friday, me included, so they stay on the record. Everyone's agent audit: find what it emails you that it could have decided alone. Heather's Google Business Profile cadence and Rev's ISBNs, neither of them in the room. Rev said in Slack it was a rough week at the day job. Mike's 1:1 is counted above.