Why It Matters
Every few weeks a new feature ships and agency owners feel pressure to rebuild their AI setup around whichever tool looks best that month. Dustin Thompson’s answer is to stop tying your work to any one tool. Keep your skills in a single home, ideally a GitHub repo, and point Claude, Codex, or whatever comes next at it. Committing a working skill gives you a version to roll back to, and team plans let people install the right skills with a click, with an admin view showing which ones get used. The bigger point is that tools matter less than strategy. Start by automating reporting, give the model real context so you don’t get internet-average output, and test new models on tasks where you already know what good looks like before switching anything over.
Who This Is For
Built for Founders members who are agency owners, media buyers, or operators already using Claude or Codex and wondering whether to switch, add a second tool, or share skills across a team. Especially useful if you feel overwhelmed by constant AI releases, aren’t sure where to start automating, or want a setup that doesn’t need rebuilding every time a model or feature drops.
We pulled this blog from an AI DTC WTF podcast with Thomas Moen and Dustin Thompson; Thomas was sitting at the beach in Croatia and Dustin was at his place in New York. Before I get into it, a quick disclosure: Dustin is a Foxwell Founders member, and he's the person who has been rebuilding a lot of our processes at Foxwell Founders and Foxwell Digital, including the Foxwell bot. So if this reads like I'm a fan, it should.
Thomas kicked things off with a question I've been hearing constantly. For about six months, every agency owner and media buyer I know went deep on Claude: skills, scheduled tasks, Cowork, all of it. Then in the last few weeks OpenAI shipped a pile of updates to Codex, ChatGPT for work and the Mac app, and people started moving over. Nobody wants to rebuild their whole setup every time one of these companies ships a feature. So how do you set things up so you can use whichever tool is best this month?
Dustin Uses Both, All Day
Dustin bounces between Claude and Codex all day without any trouble. Part of that is practical, since splitting work across both lets him spread out his subscription usage. The other part is that the features really do differ week to week.
His example: Claude just added the ability to record you doing work on your computer and turn that into a skill. Codex had something like it a month or two earlier. In his words, the two are neck and neck right now. If you've only ever worked in Claude, he thinks Codex is worth trying.
What makes switching painless for him is that every skill he has is shared between Claude and Codex. If he opens Grok Build, he still has all of them. For coding work he keeps a CLAUDE.md file, but it's a symlink to his AGENTS.md file, so the two are literally the same file. He edits it once and both tools know how he wants to work and what skills he has.
What a Symlink Is, in Plain English
Thomas asked Dustin to explain this one for those of us who aren't living in the terminal, and I'm glad he did. You keep one canonical version of a file, usually on your computer or a network drive like OneDrive. A symlink is a pointer that tells Claude or Codex where that file lives. You change the file in one place and every tool sees the change. You never end up with two copy-pasted versions drifting apart.
If that still sounds too technical, Codex has done a lot of the work for you. The first time you log in, the onboarding asks whether you want to import your memories from another AI provider, meaning Claude. When you open a new project, it tells you it found skills and memories and asks if you want to bring them over. You click yes and you're done.
The bigger habit Dustin pushed is keeping your skills in one central place, so you always know where to point whatever tool you're using. Skills are just text files and can live anywhere.
Put your skills in GitHub
Thomas keeps his skills on GitHub. (I've heard of people using Dropbox.) Dustin's preference is GitHub, and he knows that can sound intimidating. His advice for non-technical folks is to ask Claude or Codex to set it up with you. Tell it you want your skills on GitHub and let it walk you through it, including which skills should be private and which can be public, and how to pull them into Claude, Codex or whatever harness you use. These agents are very good at GitHub.
The main reason he gives for GitHub over Dropbox or a local drive is version history. Once a skill is working, you commit it and push it, and that becomes the working version. Then you're free to mess around with future versions, because you know the good one is always there to go back to. Reverting doesn't require any Git knowledge either. You tell Claude, "I think I broke this skill, can we go back to the previous version?" and it's done in a couple of minutes.
Stay a step ahead. The Foxwell Digital Download brings you what's working in paid social and performance marketing right now.
Sharing Skills Across a Team
I asked about teams because I'm getting pitched constantly by software companies that want to come on the podcast and talk about tools that bring all of this together. I've heard mixed reviews on some of them, HQ included. People tell me it's confusing and has a big learning curve.
Dustin was upfront that he's a team of one, so he hasn't tested those tools deeply. His take is that you can get pretty far by getting everyone comfortable pulling skills from GitHub. Claude and Codex team plans already let you sync a private GitHub account to the team account and make the right skills available. Your people open Cowork, see the skills available to them and click install.
The part I liked is the admin view. You can see who's using which skills, which tells you whether a skill is useful. If nobody's touching it, either the skill is broken or the team was never taught the process that makes it worth using. Thomas has more than 30 people using his internal skills library on GitHub this way, and he says it works well.
Scheduled Tasks Can Live in Either Tool
Thomas has been moving away from Claude Cowork. He had something like 96 scheduled tasks running there and moved most of them to OpenClaw. So when should you run scheduled work in Codex versus Claude?
Dustin's answer surprised me a little. He doesn't think there's a category of scheduled tasks where one is clearly better. Capability-wise, he puts them about even. He just runs a task wherever the related project already lives. If you want to move one, open a new chat in the other tool and say, "I have this scheduled task over in Codex, can you move it here?"
Where Agencies Should Start
I asked Dustin what he sees agencies and marketers trying to accomplish now that he's been inside a lot of different businesses. Ad creative is what everyone talks about online, and he called it one of the harder things to automate because of how many moving pieces there are.
His number one recommendation is reporting. Automate pulling the data, verifying it and building the client deliverables. After that, look at any business process you're still doing by hand. It isn't exciting, but it's the kind of work that suits scheduled tasks well, and it takes a lot off an owner's plate.
We just ran our Foxwell Founders member survey, and overwhelm came up again and again as a challenge. Some of the builds I've seen help most with that are simple: a daily summary of what you need to do today and what's still outstanding. I asked Dustin whether people should just start by talking to the models about their situation, and he said absolutely.
Use voice mode (ChatGPT had just released desktop voice mode the night before we recorded, and Dustin was tired because he didn't get access until 10:30 that night), use Wispr Flow, or type. Do a brain dump: here's my situation, here's what I do, here are the tasks I'm stuck doing. Tell it the goal, whether that's being more productive, more profitable or getting more creative out the door for clients. Then go back and forth for 30 minutes or an hour. By the end it'll lay out what you could build and the order to build it in.
Thomas added something he sees with his own team. People ask the AI how to do something, it explains, and then they go do it themselves. They're surprised to learn it can use the browser or organize their files. His advice is to remember who's the boss and ask it to do the work. Dustin sees the same thing and said sometimes the models get lazy and just explain the steps. He replies "do it for me," and usually it says it can and goes and does it.
Have a Strategy, and Don't Chase Every New Model
My view is that the people winning with this have a clear idea of what's going to make the business more money, and they use AI to get there. The people who struggle are on a lot of side quests. Dustin agreed. If you go in without a plan for how you'll use it, you'll probably end up with more cost and the same productivity, or less, because people get distracted and burn tokens on projects nobody ends up using. He does side quests all the time, but he treats them as learning, separate from the business.
The same goes for new models. A new release will sound cool and will probably cost more, but that doesn't mean it's better for the work you're doing today. Dustin's example was the urge to say, "Fable 5 is out, let's run everything on it because it's smarter." Check first whether it helps you grow revenue or margin.
I asked how he keeps track of what each model is good at, since this stuff changes seemingly by the hour. He does look at benchmark sites, with a grain of salt, because some labs get good at training models to do well on benchmarks and you don't really know what's going on inside those tests. His real test is running a new model on a task where he already knows what good output looks like. He also reads the model's reasoning as it works to get a feel for its personality, for lack of a better word. That takes time, but it shows him how it thinks about a problem before he judges the output. For coding models, he runs the same set of evals on the same project every time something new comes out and compares the scores.
His advice for agency owners is that you don't have to rush. Let a new model filter through the community first. If something is working, sticking with it is fine. He also shared a phrase he picked up on X that I can't stop thinking about: watch out for token dealers. When Anthropic is pushing a new model, they're a token dealer. Keep that in mind.
Context Separates Good Output from Average
On a recent episode, Jess from Fireteam told us they have an AI copywriter that's better than any human. Then he explained that before it writes anything, they hold three hours of brainstorming meetings with the AI listening in. It picks up the nuance in how the team talks about the product.
Dustin put it simply. If you don't give the model the right context, you'll get internet-average knowledge back, because that's what it has to work with. Unless you're fine being average, give it more.
This is also why the Foxwell bot and the Foxwell MCP have been so useful for our members. A high percentage of people in our survey said the bot is a main part of how they use Foxwell Founders. It has years of the best media buyers in the world talking through problems behind closed doors, so the answers come with real context.
What Dustin Is Building
Besides tinkering with the Foxwell bot, Dustin has been building more advanced skills for his own daily work in media buying and development, and automating them as much as he can. One I want to copy: once a week, a scheduled task on both Codex and Claude goes through his computers and repos looking for content ideas. It surfaces things he worked on and would have forgotten, so he can decide what's worth writing or posting about.
He's also working on a model orchestrator skill for Claude Code. He spent a weekend testing the same idea in Cowork with mixed results. Subagents in Cowork do work, but in most cases running them was less token-efficient than running one model.
Practical Steps to Get Started
Here's what I'd take from this conversation:
Pick one home for your skills, ideally a GitHub repo, and point every tool at it. Ask Claude or Codex to help you set it up.
Commit a skill once it works so you always have a version to go back to.
Automate reporting before you try to automate creative.
Spend 30 minutes talking your situation through with a model, then ask it to do the work, not just explain it.
Test new models on a task where you already know the right answer before you switch anything over.
If you need help building AI systems for your agency or brand, Dustin is the guy. And if you want to be in the room where conversations like this happen every day, apply to Foxwell Founders at foxwellfounders.com.
Until next time, take care and keep it real.
Dustin Thompson heads up Konnected Interactive.
Check out Konnected Interactive at konnectedinteractive.com
You can also connect with Dustin on LinkedIn
If you want to hear the whole conversation, watch the full episode here:
Note: This blog was written by a human, but AI helped draft the outline and synthesize ideas.
