One of the coolest things about AI today is that there are so many different ways you can link these apps together to help you work your best. I know everybody’s got their own flavor (and my flavor is Pumpkin Spice), but I thought I’d share the setup I use to be as productive as I can.
I spend most of my time in OpenAI’s Codex and Claude Code, even when I’m not coding. I use them for research, planning, and writing too. With access to my files and the apps I’ve connected, they can work directly on those tasks. The names make them sound more specialized than the way I actually use them.
What took me a while was figuring out how to use all of that without giving myself another job moving information between the AI apps. I can spend time working through a plan in Codex, then want to talk it through with Claude later. Having to stop and package up everything we already figured out gets old pretty quickly.
I’ve written about work continuing between sessions. Here are a few of the habits behind that.
I talk instead of typing
One of the least expensive improvements to my workflow was remembering that my Mac already has Dictation. I didn’t need to buy another transcription app just to speak my request into Codex or Claude.
I use dictation when I’m at the Mac and know roughly what I want to ask. It turns my speech into a message I can read and fix before sending. On a Mac, it’s under System Settings, Keyboard, Dictation. Apple’s guide has the settings and shortcut options.
When I want to talk through a strategy, I use conversational voice mode in Claude or ChatGPT. That’s a different activity for me. Sometimes I know the request and just don’t feel like typing it all out. Other times I need the back-and-forth to figure out what I’m asking in the first place.
I give my AI apps somewhere to find the work
I use Clarabit, which I built as my own personal system for keeping track of work, to hold the tasks, plans, decisions, and documents for my projects. Over the last several months, I’ve been working on making it available to others. I’ve got a waitlist if you want to try it out.
I connect my AI apps to Clarabit so they can read and update the work directly. That connection uses MCP, a way for AI apps to work with other software. The setup guides for Claude, ChatGPT, and Codex cover connecting apps. If I switch from Codex to Claude or ChatGPT, I can ask it to pick up the same project.
I keep a short set of instructions in the project folder on my computer: AGENTS.md for Codex or CLAUDE.md for Claude Code. It tells the AI to read the project details from Clarabit and save plans, decisions, and progress back there. That way I don’t have different versions scattered across AI apps and computers.
When I open a new chat, I ask what’s waiting on me or blocked because it needs my input. I’m generally the bottleneck in this flow, and I ask the AI to tell me when I’m holding things up. I choose what to work on, then it gets the details.
The AI might work for fifteen minutes or chug for an hour. I start a second task, then a third, and by the time I get back to the first one, I need to remember what I wanted to review. I’ve typed some version of “what should I be reviewing here?” more times than I can count.
You could try a small version of this with one project brief in a place both AI apps can access. Ask one AI app to save your decisions and what’s still open, then see whether another can pick up the work.
I keep the raw conversation
Granola has become useful to me for more than meeting notes. I read its summarized notes, but when I give the conversation to another AI app, I want all the raw stuff we said. I use Granola’s MCP connection so the AI can pull up the conversation directly. Granola’s setup guide walks through connecting it to your AI app.
If I remember explaining something well a month ago, I can pull up that discussion. Otherwise, I’m relying on a memory I’ve often got wrong or mixed up with another conversation (ugh, getting older). The summary helps me find it, but the original wording is often what I need.
This matters for writing, and it matters when I’m trying to remember why we made a decision. The hesitation or side discussion can be useful context. I’ll ask for the relevant transcript instead of giving another AI app a summary of a summary.
It also lets my AI apps pull out repeating patterns in the topics I’m talking about across conversations. That helps drive what I write about and what I build. Sometimes I’ve been coming back to the same idea in several conversations before I’ve sat down to do anything with it.
I leave feedback on the actual document
If the AI creates a plan or article draft, it becomes a document in Clarabit. In Codex, I can open that document right next to the chat. Having the document and the conversation together makes it much easier to collaborate with the AI. It feels a lot more like working with a teammate.
I read the document and leave comments on the exact text that feels wrong. Sometimes I’m asking for a bigger change. Sometimes it’s, “I would never write this,” or, “This isn’t the way I said it in the Granola recording.”
Then I tell the AI I’ve added feedback. It reads the comments, revises the document, and creates a new version. I read it again. We go back and forth until the document says what I mean in language that sounds like me.
The comments stay attached to the work, which saves me from copying each one into another chat. It also means I can switch AI apps without trying to reconstruct which paragraph I was talking about or which exported document had the latest edits.
For software I’m building, I still open it up and test it directly, no matter how many tests pass or how well the AI thinks it did. I need to make sure the flow works and feels intuitive. I’d estimate that about one out of four times it doesn’t, and I have the AI start over with a little better direction.
I have an old MacBook doing the waiting
I have an old MacBook at home that gives longer-running work a place to run. It’s plugged in, connected to the internet, and configured to stay awake while on power. I use Tailscale to put it and my everyday laptop on the same private network. It makes the two computers feel like they’re sitting next to each other, even when I’m somewhere completely different.
Claude Code Remote Control and Codex Remote let you keep chatting with the AI working on your home computer from another device. Those links cover getting each one set up. Tailscale connects my computers; the apps’ remote modes let me talk to the AI doing the work.
From my laptop, it feels like I’m working right there, but the AI is doing the work on the Mac at home the whole time. When I close the laptop and walk away, that work can keep going because I haven’t closed the computer doing it. The home Mac still needs to stay awake and online, and the AI may need to wait for my input.
You don’t need the whole arrangement. Pick one task, try dictating the request or leaving feedback directly on its document, and see whether it saves you some back-and-forth.
In the next article, I’ll share how I use Claude through CarPlay to talk through work during a commute or a longer drive. I can work through ideas and decisions aloud, then pick up the hands-on work when I’m back at a computer.
Everybody’s got their own way of putting these apps together, and I’d like to hear what’s working for you. What’s one small habit or connection between apps that’s made AI more useful in your work? Share it in the comments so other people can try it too.
I write about using AI to build and run a company. Get new posts by email.