AI Workforce Clinic · Members-only session recap

Own the folders, rent the model

I've spent a month telling this room to build agents and never once said where they should live. Friday, Mike said it in one sentence from a car with a broken microphone, and it reframed everything we've been doing since July.

SessionFriday, September 25, 2026
Runtime49 minutes
In the roomMike, Kristin, Natalija, Heather
RecordingFull transcript
ShiftRun AI workers
The shift

"If you decide you don't like Claude anymore, that team is in your environment."

Mike said it at minute forty-six, near the end, after the real conversation had already happened. 46:51 Everybody nodded and we said goodbye. I've been thinking about it since.

Here's what set it up. Mike asked the room a plumbing question: what does your setup actually look like? Folder structure, VS Code, the desktop app, where does the work happen? 37:50 He builds agent teams for clients, he's moving from individual setups to company-wide ones, and he wanted to know whether the people a few steps ahead of him had found something better.

Kristin shared her screen. ChatGPT running in her client projects, Claude on her machine through Code, and a local environment she calls her KKAI brain. 40:19 Inside it: a folder for connected apps, her knowledge files, her skills, her clients, the context about her business and her voice, even headshots for her agents so everything sits in one place. She dropped Cowork entirely because she thinks it's out of date now. 40:13

Mike looked at it and said that's the same structure he's been implementing for people. Two businesses that share no tools, no clients and no industry landed on the same shape. That's not a coincidence, that's the shape the work wants.

Then he named why it matters, and this is the part I'd underline. It isn't a filing convention. It's an operating team. 45:39 The folders aren't where you keep the output, they're where the team works. You give an instruction and something happens in there. His example from the day before: he was sitting with a client who didn't have Claude set up properly, so he turned to his orchestrator and said go find the file we built for the other client, rewrite it for this one, tell me when it's ready. 44:18 Three sentences. It came back while they kept working and he forwarded it from the room.

Now put those two things together. The team is a set of folders on your machine. The folders hold the context, the rules, the history, the skills, the output. The model is the thing you plug into them.

Which means the model is the part you can throw away.

I proved that this month without meaning to. When I installed Codex it scanned my machine, found how I'd set everything up for Claude, and asked in one click whether OpenAI should get the same access. I said sure. 42:51 I started out careful about which work went to which. I've gotten sloppy since, because nothing has broken and neither one has failed to know something. It's Coke or Pepsi. 43:37 Their pricing proves it: a new model drops, the price comes down, and the other one matches inside the hour. 47:06

Does that change what you should be building this week? It changes where you build it. Every hour you spend teaching a chat window about your business is an hour you'll spend again on the next vendor. Every hour you spend writing that same thing into a folder is an hour you own. Mike's version is cleaner than mine: how you use AI is a choice, and this is the choice. 19:42

The play · move your team out of the chat window

1. Make one folder and call it what it is
Your AI team's home. Not a project, not a chat. A folder on your machine or your drive that you could hand to a new employee.
2. Write the business down inside it
Who you serve, how you talk, what you sell, what you never say. Kristin's version covers her business and her voice in plain markdown files. This is the part that makes every model you ever plug in smarter on day one.
3. Give each repeated job its own folder
One per agent. Its rules, its examples, its output. A skill you invoke five times a week earns a folder, and that folder is the difference between an agent and a good prompt.
4. Point your workbench at it, not at a chat
Claude Code or Codex, opened at that folder as its root. Chat windows read your rules and then do their own thing. A workbench pointed at the folder follows them.
5. Test the portability on purpose
Plug the other vendor into the same folders and run one real job. If it works, you've stopped renting your own business back from a model.
The repair

When it does something you never asked for, don't just undo it

Natalija brought the best problem of the day. She'd written a two-page piece going out to seventeen hundred member companies, it was beautiful, and then she collected feedback from everybody and Frankensteined it to death. 22:24 She had to start over from a clean version. Fine, that happens.

Here's the part that got her. She was on version twenty, she pasted in an email that happened to contain her marketing colleague's comments, and Claude went ahead and incorporated them. She never said incorporate. 23:11 She fixed it by pointing it back at the version she wanted. 23:31 Then she asked the question that made the session worth publishing: how do I teach it not to do that again?

I stopped her there. Did you ask it to diagnose why it happened and how it won't happen next time? 23:49 She hadn't. Almost nobody does. We revert, we sigh, we move on, and we hand the same mistake a fresh opportunity next week.

Reverting fixes the document. Diagnosing fixes the worker. That's the whole difference between a tool you use and a team that gets better, and it costs you one extra message.

The prompt · paste this the next time it goes off-script

Say what it did, not how you feel about it
"You incorporated edits from a document I pasted for reference. I never asked you to apply them."
Ask for the diagnosis
"Diagnose why you did that. What in my instructions or in the context made you think applying them was in scope?"
Ask for the rule
"Write me one line I can put in your instructions so this never happens again. Keep it short enough that you'll actually follow it."
Then put the line where it lives
Into the folder, not into the chat. A rule that lives in a conversation dies with the conversation. That's the shift and the repair meeting in the same place.
Around the room

Four people, four different weeks

Mike D&M Solutions · Software

PropsFor naming the mindset shift the rest of us had been circling since July, and for asking the question that got everybody's setup on the table instead of guessing at it.

ShippedProgress on the company-level agent team. 00:42 He and his developer took the developer's working setup and started replicating it onto other machines to share the agents. It isn't a full solution yet, they're testing approaches to handle the sync. Plus several individual agent setups for his own customers this week.

The play he gave awaySitting with a client who hadn't set Claude up properly, he told his orchestrator to find the guided setup doc he'd written for another client, rewrite it for this one, and report back. 44:18 It was ready before the meeting ended. Put your team to work in the moment, in three sentences.

The tipWhen a chat keeps giving you the same wrong answer, change the model. 27:47 Opus to Sonnet, better to worse, it doesn't matter, different is the point. And the sharp corollary: an orchestrator checking another agent's work with the same model underneath isn't a second opinion. 28:30 It's the same brain agreeing with itself.

On the resistance questionThe hard part is judging who's open to it. Once somebody is willing to make a choice about how they work, that kicks the door open. 19:57 Finding their tolerance for change is the door. I said so out loud because he's right.

Kristin Marketing and web agency

PropsShe put her whole system on the screen when Mike asked, which is the most generous thing anybody did all session.

The insight"Managing humans is more time consuming than managing agents." 10:40 She's been in the weeds for weeks and is working out how to lean harder on the agent side and trust her people to own their end. She doesn't have time to micromanage either one.

The move worth stealingBefore she joined the call she asked Claude what the two of them had done together in the past week. 11:14 It listed everything and she realized she'd made far more progress than it felt like. When you're heads down, your own memory is the unreliable narrator. Ask the thing that kept the records.

Where she isA user interface and training module with six permission types across multiple task flows, a fast-turn site for a city government office, and retainer work with a print deadline six weeks out covering trade show displays, signage and Spanish translation. She's using AI to pressure-test her thinking and to show her things visually when she can't picture them, then doing the design herself. Her read: her brain is still better at the design, AI is the business partner she kicks ideas around with.

What she's doing about the deadlineI asked how much of the hair-on-fire she'd exposed to her agents. Not the print schedule, not yet. 15:20 That's her afternoon with her new project manager: feed in the budget, the print resources and the production plan, and have it build the Gantt chart that says when each piece has to reach the printer. The project is bigger than her head, which is exactly when you hand it over.

Natalija Business development

PropsFor bringing the two failures nobody else wanted to admit to, which is how the room earns its keep.

ShippedHer CFO got on board. He didn't join the class, he built his first report on his own, and she applauded the heck out of it to keep the reinforcement coming. 21:24 Her read on why: not her nudging. The CEO adopted it and his college-age son is pushing from the inside. The young people in the building are doing the convincing for us.

The frictionChatGPT image work. Ask for one tweak, say don't change anything else, and it garbles the letters and changes five other things. 24:37 I gave her the mechanism, because it helps to know it isn't you: the model re-renders every pixel from scratch and makes fresh decisions on all of it. There's no stopgap. Kristin's the same: she needs one image split into three and cannot get it, so she crops it herself. 25:34

Her best use of it this weekNothing to do with work. A public-records pull for a local campaign she cares about, where the full list was sitting in published budget documents the whole time and AI found it faster than twenty emails back and forth. And a check-engine light: she looked up the symptoms before going to the shop so she walked in with a clue. 34:09 It got one thing wrong, and she still walked in ahead.

The workflowWrite the image prompt with Claude, render it with ChatGPT. She gets what she wants that way, and she credited it back to a tip from this room.

Heather Local classes · enrollment-driven

Where she isStretched thin and carrying a lot outside the business. She said she's doing too much and just needs to get through it. I'm glad she showed up anyway, and I told her so.

What the room gave herSlow down and put the administrivia on AI before you put the strategy on it. 05:23 Meta's Muse rides along on the Instagram account, it's free, it works from your phone, and it'll handle forms and errands while you're driving. When you're at capacity, the first job for AI isn't growth. It's giving you back an hour.

The objection she namedHer circle isn't resistant to change, they're resistant to what the data centers do to the planet. 35:19 That's a real position and it deserves a straight answer, not a sales pitch. I went down the rabbit hole on it one night. The energy and water cost is real, the back-room dealing around these sites is worse than most people know, and the trajectory is still clear. Both things are true at once, and telling somebody their concern is nonsense has never once moved them.

What I brought

Two things from my week

A new logo across 861 pages, in ninety minutes

I've wanted to change the PitchKitchen logo for a long time. The old one was two speech bubbles, and our competitor StoryBrand uses speech bubbles too. 02:15 I'm hammering them hard on answer engine optimization these days, so looking like them is the last thing I want.

I started a chat at 10:33 that morning. 02:37 We went through iterations, stick with the bubble or go abstract, and landed on a butterfly, because the work I do is transformation. By 11:07 the typography and the mark were settled. Then I asked for a full corporate brand guidelines document: the mark, the colors, the type, and the don'ts, so nobody adds a gradient or a colored dot to it later. 03:15 That was done at 11:28.

Then the real ask. Replace the logo on the homepage, show me the preview because it needs different sizing, and then go through every page on the site and swap it everywhere, favicon included. It worked through thirteen separate use cases and finished all 861 pages while Kristin and I were talking. 16:44 I was down to 1% of my usage for the period, and I'd do it again.

For perspective: this site had six pages in March 2026. 17:07 I moved it to Cloudflare with Claude, went hard on answer engine optimization, and it's 861 now. Kristin's reaction is the right one. A brand doing that logo change the old way books months for it.

Which brings up the thing I don't know how to handle. We designed a site for a client back in March and handed over Figma files, because Figma files are what they asked for. It took them six months to turn those into a live WordPress site and move their blog. 18:47 Do I wake up their head of marketing, or leave them sleeping? 18:23 I asked the room and I still don't have a clean answer. Mike's is the closest thing to one: find out what their tolerance for change is first.

One more thing I've noticed about my own week. The old business process is so smooth now that the scarce thing isn't production. 15:58 If I want to grow, my time goes to the thing AI is worst at: building trust with people, and rebuilding trust I had with people years ago. That's where my attention is.

Muse as the personal assistant I stopped trying to build

I've been using Muse, Meta's assistant, and I hit the usage limit on it. 05:23 It's less sophisticated than Claude Code and far easier, which turns out to be the right trade for personal admin. In one week it researched every PitchKitchen lead before I opened the mail, built eighteen eBay listings, flagged the watermarks and blurry photos in them, filed my power outage report and confirmed when the power came back, delivered print-ready art for a button, tracked down how to cancel boat insurance, and faxed a prescription transfer to a doctor's office. 07:00 That's ninety minutes back on the personal side of one week.

The one that surprised me: a property-line question with a neighbor. Late at night, on my phone, I asked it to find the deeds for all six houses on the easement and tell me which of them carry the right to drive on it. 32:49 It built the chart. Four of the five have it named specifically, mine doesn't, and it told me exactly where to go for the record it couldn't pull. I had no idea how much public record is sitting there waiting for somebody to ask. 33:24

The button was for the Fieldy I'm wearing, a recorder on a cord that listens to my conversations with a button to shut it off. 38:12 I want people to know when they're being transcribed, so I had it make me something to wear that says so. I'm running it for a month as an experiment to see what I can learn about how I work from everything I say in a week. That's the whole hypothesis, and I'll report back.

Tool desk

What crossed the desk

Muse (Meta)
Free, attaches to your Instagram account, runs on your phone while you drive. Weak at building systems, excellent at errands, forms, returns, complaints and public records. The personal-admin tool most of this room is missing. 06:24
Codex
On install it scans your machine, finds your Claude setup, and offers the same access in one click. 42:51 This is the portability proof, and it's why the folders matter more than the vendor.
Cowork
Kristin has stopped using it. Her read is that it's out of date for agent work, and she keeps her agents on Code. 40:13
Magnific (formerly Freepik)
Kristin's escape hatch when ChatGPT images won't cooperate. Around fifteen models behind one door, Nano Banana among them. She has ChatGPT write the prompt, drops it in, and tests models until something lands. 28:57
AI and food photography
It can't do appetizing. Ground meat comes out as perfect little pellets with none of the imperfection real meat has, and raw burgers look wrong. 29:39 If you sell food, budget for a photographer.
Switching models to break a loop
Mike's trick, and it's free. Same chat, different model, better or worse doesn't matter. 27:47
Fieldy
A wearable recorder with a mute button. Greg's running a month-long experiment on what his own conversations teach him. Wear a sign if you do this. 38:12
The jagged edge of the frontier
Ethan Mollick's phrase, and the most useful frame for a week like this one. The same model in the same session does something unbelievable and something idiotic. 31:42 It isn't binary and it isn't you.
Before Tuesday

What each of us is conquering next

  • EveryoneOpen the folder your AI team works out of. If you don't have one, that's the week's job. If you do, plug the other vendor into it and run one real task, so you know for a fact you're not locked in.
  • KristinThe print schedule goes to the agents this afternoon: budget, print resources, production plan, out comes the Gantt chart with every printer deadline on it.
  • NatalijaRun the diagnosis prompt on the edits Claude applied without being asked, and put the one-line rule it writes into your folder, not into the chat.
  • MikeKeep going on the sync for the company-wide team, and tell us which approach won. Half this room needs that answer.
  • HeatherOne errand to Muse this week. Not a strategy, an errand. Take the hour back.
  • MeThis page. And a straight answer on whether to wake up a client who's still doing it the old way, which I don't have yet.