AI Workforce Clinic · Members-only session recap
You hit your usage limit in the middle of a job and sit there waiting for the meter to reset. Mike stopped waiting. Two weeks after this room decided the model is the part you can throw away, he showed up having built it.
"I can now run two separate Claude accounts and Codex out of the same thing."
Two weeks ago this room landed on an idea: your AI team lives in your folders, which makes the model the part you can throw away. That recap is here. The homework was to plug the other vendor into your folders and run one real task. Friday, Mike came back with the whole thing built. 02:14
Here's his setup. A personal Claude account on the 20x plan, which he kept because the team plan tops out at 5x. A team account for his company. And one Codex account. 02:40 He worked with Claude on and off for two weeks to build a launcher that starts any one of the three on its own, or all three at once. 03:02 Every seat runs his chief of staff and the same agents. He gave each one a different name so he always knows which account he's sitting in. 03:49
Why go to the trouble? He'd maxed out an account a couple of times in the middle of a workflow. Now when one fills up, he starts the next workflow in the other account. 04:17 Each job stays in the seat that started it. Nothing waits on a meter, and he gets the full value of every subscription he pays for.
Codex was the hard part. The two vendors look for different file names and read them a little differently, so Mike had Claude build conversions. When Codex starts up, it reads the same files the way it needs to, and he never maintains two copies of anything. 03:56 He's honest that a couple of things still don't work exactly the way they do in Claude. 03:38 Keeping the team running is the bar, and it clears it.
Then came the part I'd underline. On September 25, Mike warned us that an orchestrator checking another agent with the same model underneath isn't a second opinion. Friday he brought the fix. He built a reviewer on Claude, with Fable 5 underneath, and a reviewer on Codex, and each system can call the other's. 04:45 Claude's work gets audited by OpenAI. OpenAI's work gets audited by Claude. That's a real second opinion, and it runs on subscriptions he was already paying for.
I called it nirvana, and I meant it. 05:09 I'd been doing a smaller version all week. I added Muse to my stack, pointed it at the same folders Claude and ChatGPT use, and picked up from one to the other mid-task. 01:43 The phrase I landed on was free from any one AI. 12:14 With this much change going on, marrying one vendor is the riskiest move on the board.
Natalija showed the other side of the same coin. She's running four windows across Claude and ChatGPT and forgetting which one is which. 06:53 Her biggest headache this week was figuring out where the folders are and where the conversations are. 07:02 That's exactly the problem the folder solves. One home for the team, and every model you open reads from it. Does that make sense?
Mike closed his turn with the sentence I'd want every new member to hear. When he joined, he was still trying to get his head around the team concept. Now his day goes like this: get up, figure out what's going on, figure out how to give it to Claude. 05:26 On Wednesday night Claude ran his 7 PM job while he was at a Red Clay Strays concert, and he checked from the crowd that it ran. 05:44 That's what this buys you. Your evening back.
The play · make your AI team vendor-proof
Mike D&M Solutions · Software
PropsFor doing the homework from two weeks ago and then going further. He plugged the other vendor in, then made the vendors check each other's work.
ShippedThe three-seat launcher above: two Claude accounts and Codex, any one alone or all three at once, one team underneath. 03:02
Loop closedOn September 25 he said he'd tell the room which sync approach won for the company-wide team. He did, in Slack on October 1: company-wide agents shared through SharePoint, controlled and managed at the company level, with every step validated. The one-step installer is still in progress.
The habit worth stealingWalking and talking. He puts Claude in chat mode, goes for a walk, and comes back an hour later with the thing figured out and kicked off. 10:26
Wrestling withCodex still doesn't do a couple of things the way Claude does. 03:38 And by his own account he does very little marketing, which is why he wants to hear how Natalija's next experiment goes. 18:24
Natalija Coffee kiosks · Business development
PropsFor the most honest performance review any AI has gotten in this room.
ShippedA one-minute sizzle reel, cut by Claude Code from a local TV segment about her company that aired fourteen years ago, to pitch the station on a comeback story. 08:01 She dictated it halfway through a dog walk, about twenty minutes start to finish. 09:14 She handed it the original video file and asked for a then-and-now. 09:27 The part about her company ran only five seconds, so it had little to work with, and she didn't love the "now" visuals because she hadn't coached it much. 09:40 She still called it fairly good, and it beat her expectation that ChatGPT would be the creative one. 10:05
Her performance reviewFor writing and posts, ChatGPT delivered the least value, even turned up to its highest model. She tells it the same thing ten times. Claude is proving more efficient. 07:11 For graphics and images she still reaches for ChatGPT. 07:41 My one-line summary: Claude Code is great at video editing, and ChatGPT still wins on images. She said "100." 10:11
The pushback she's gettingShe posts AI-made images of herself on LinkedIn, and the complaints come from traditional marketers. 15:37 Her read: they feel it's cheating, and some of it comes from people whose work it threatens. 16:27 Her own numbers say otherwise. Posts and images she turns around in five minutes get likes and comments, and nobody has called them slop. 16:36 Her rule is the right one: invest the time so it doesn't look ridiculous. 16:53 And her evidence the tide is turning: retirees on the trail using AI. 17:08
NextHer part-time social media help just moved on. She already has the posting schedule, and she wants Claude creating the campaigns by theme, holiday and platform. 17:46 Mike wants the report.
From September 25 Loops, said plainly
KeptMike's sync answer, above. Mine too: "plug the other vendor in" was the room's homework, and Muse now runs on the same folders as Claude and ChatGPT.
Not raisedKristin's print-deadline Gantt chart and Natalija's diagnosis prompt. Kristin sent her regrets for Friday. Both carry, no scolding.
From SlackRev isn't in the room these days, but he posted on October 7 something he's been building on his own, code he owns and can change on a whim to support whatever comes next. That's this week's whole theme in one sentence.
Muse now sits in my stack next to Claude Code and ChatGPT. It connects to the same folder structure and knows everything that's going on, so I can pick up from Claude and go right to Muse. 11:46 I gave it a job. Every morning it reads my email, tells me who I need to answer, and drafts the replies from everything it knows about the projects. 12:00 That's the vendor-proof idea in my own week: three companies' models working one set of files.
I ran an experiment for a LinkedIn post. I recorded a raw take of myself talking to Claude about what I wanted, almost like a prompt, and let it do the edit. 12:55 It put text and graphics on screen, then went out and clipped related video of Sam Altman and Anthropic to tie into what I was saying. 13:13 It's rough, and good enough for a lot of purposes, which is the same thing Natalija found the same week. 12:38
My prediction: we're about to see a lot more people recording themselves. Most people held back out of fear, because a good edit cost real money and time. 13:21 That cost just fell through the floor.
Nora is my agent for networking on LinkedIn, and she doesn't do volume. Maybe twenty people a week. 18:48 She picks them from deep research on their business and their website 18:55, then writes a short note in lowercase that quotes a strong line from their own site back to them, names the messaging pattern I see all the time, and offers my five-second read on how the company lands with a buyer who's never heard of it. 19:00
This week a founder wrote back: "I'd welcome it." 19:29 The door's open. Next he gets a quick video and a full gift with a real read on where his messaging stands in the market. These are good leads, and we wouldn't have them without putting AI to work on the sales and marketing side of the business. 19:43
Mike asked the question everybody asks: doesn't LinkedIn block agents? 19:57 It's the hardest platform for it, harder than Meta. At a human pace, about twenty a week, it's working for me. 20:50
Carried, no scolding: Kristin's Gantt chart, Natalija's diagnosis prompt, Daniel's first Claude Code folders.