All day long, I talk to agents 🤖
This is the job now
Hello makers,
In the last issue, I promised to deliver a deep dive into how I work with AI day-to-day, tools and all. Here we go.
TLDR
I’ve built a second brain for my company with Claude and GitHub
I use phased planning for any new features or larger work
I rest my brain to avoid scatterbrain
If you told me in November of last year that I should be using AI to do my work I would have laughed in your face. I’d seen the code those early models produced. I wouldn’t risk my reputation on it.
Two months later, new models arrive. I bite the bullet and start tinkering. Claude one-shots a task in minutes that would have taken me days. And the code looks good. Better than mine even. Coding agents actually arrived.
Today I’m barely looking at code. I talk to agents, and this is how I do it.
My company’s second brain
This whole thing works because I’ve structured my entire company into a second brain backed by Claude, and stored on GitHub. It’s a hub-and-spoke model. Here’s a small snapshot of what that looks like.
- sixoverground-workspace
- briefshot-hub
- briefshot-ios
- briefshot-next
- linkmyphotos-hub
- linkmyphotos-android
- linkmyphotos-ios
- linkmyphotos-rails
- shinytimer-hub
- shinytimer-macos
- shinytimer-web
- sixoverground-hub
- sixoverground-next
- libraries
- claude-phases
- claude-project-managerEach directory in that list is its own GitHub repository. Parent directories have information about their children. These repositories can be added to Claude Code. I never use Chat. Always Claude Code. And it works from my phone with the Claude app.
At the very top level, I can brainstorm anything. I’ll start a new Claude Code session with sixoverground-workspace and just start chatting. Instead of relying on session memory, all my previous thoughts and memories are saved to the repo and Claude has access to everything. All my preferences, tastes, skillset, goals. It has all the context it needs to help me.
If I want to release a new Briefshot feature, I’ll start a new Claude Code session, and include the briefshot-hub, briefshot-ios, and briefshot-next repositories. Claude will know all about Briefshot from the markdown files in both the hub and the code repos. It can plan the feature across multiple surfaces, and it knows my tastes and roadmap that I’ve added to the hub.
Running out of context is no longer an issue because it’s all persisted to GitHub.
Phased Planning
I’ve been building with AI for over seven months now. At first, I was very tentative to allow the model to do anything. I’d prompt Claude to do something, review the plan, inspect every permission request, allow Claude to execute, then closely examine the diff before committing anything. This is a slow process, even with AI churning through a task.
Over time, I’ve become more comfortable with the model doing more on its own. But models have a tendency to get off track, to bloat code, to over-engineer, or to straight-up break things. There needs to be some oversight, but I can’t be inspecting every single line of code for larger tasks.
The happy medium I’ve landed on is phased planning. When I’m planning a feature with Claude, I have the model break up the feature into discrete phases. Each phase is non-destructive and self-contained. It needs to be able to be merged without breaking production. This allows me to review smaller PRs as we move through the larger feature, making it easier to steer the model when it goes off track.
I’ve since automated this whole approach. I’ve created two libraries that help run these phased plans.
Claude Project Manager
sixoverground/claude-project-manager
This is an npm library that runs on a timer. It has commands to help Claude build the phased PR. It works by triggering Claude Routines in the cloud. Each routine is designed to churn through a specific phase, and to create a PR for that phase. It waits for any GitHub actions to complete, and addresses any code review feedback from Copilot or other agents. Once the phase PR is merged, the timer triggers the next routine to work on the next phase in the plan. This continues until all phases are complete.
Claude Phases
The biggest limitation of claude-project-manager is that I can’t steer Claude Routines when I’m away from my desk. If the model hits a wall, or a CI test can’t pass, I have to direct Claude from my desktop computer. Claude Phases brings the orchestrator into the model, instead of keeping it outside. It’s a Claude plugin with two skills: one builds the phased plan and one runs the phased plan. Instead of using a Claude Routine prompt to orchestrate, the directive is saved to a markdown file in the repo. Everything runs in a Claude Code cloud session. I can watch it work from my phone, and steer it when I’m on the road as needed.
YOLO
Both of these libraries offer YOLO mode. When it’s on, the library will automatically merge phased PRs.
Note that YOLO mode never targets main. Every phase PR merges into a feature branch. The only PR that touches main is the final one, and I review that one myself. No AI code auto-merges into production, ever.
Here’s the full loop for a single phase:
Claude finishes the phase and opens a PR.
An automated code review agent inspects it and leaves comments.
The library reads the comments, decides what to address, and pushes changes to the PR.
The review agent takes another pass. This loop repeats until there’s nothing left to flag.
In parallel, the repo runs its gates: web apps get a preview deployment, mobile apps get a full build, and tests run everywhere.
If anything fails (a test, a build, a deploy), the model addresses the failure and the gates run again.
When every gate is green, the PR merges automatically, and the model moves on to the next phase.
In YOLO mode, the final PR is instructed to end with a comprehensive UAT checklist covering everything in the feature, so I can manually test the full set of changes before I merge it myself.
This has changed my workflow dramatically. Once a phased plan is defined and set up, Claude can churn through an entire feature, across multiple repos, with no breaks. When everything goes smoothly, I come back to a single, clean, thoroughly tested PR with a clear checklist of things to manually run through.
And now I can do it all from my phone.
Scatterbrain
The days of going into a zen state to do deep work for coding are over. Because you can build so much, so fast, it means you can work on multiple tasks at once. This effectively creates rapid context-switching. The resulting feeling reminds me of when I had to switch from code mode to design mode. It’s not like flipping a switch. You need to mentally break from one line of thought, and regroup.
It’s a real problem, and I’m still figuring out how to cope.
In 2020 I wrote a whole issue about protecting flow state, and mentioned 'skillboxing' to avoid context-switching. Now context-switching is the job.
Here are some things I do to try to avoid full-blown scatterbrain:
Limit tasks to 3
This number feels manageable. Any more, and I start making mistakes. Last issue I said I track up to 5 mental models. I’ve since learned my real limit.
Meditate
I meditate for 15 minutes before starting any work, and I meditate for 15 minutes when I’m done with work.
In the morning, this clears my mind so I don’t have other concerns polluting the tasks at hand. In the evening, I’m able to reset and actually rest my brain for a bit. It only works when I’m religious about it.
Take Breaks
Sometimes, I just need to walk outside and touch grass. I’ll go for a walk after lunch. I go to the gym. Anything that gives my brain a rest for a bit. And I don’t feel bad about it because my agents are still churning away on work while I’m away from my desk.
What’s Next
How are you using AI day-to-day when you’re making stuff? How do you handle scatterbrain? Reply and let me know.
Next issue: a closer look at how my company’s second brain works. Make sure to subscribe to get it in your inbox.
Keep making, and thanks for reading! 🙌
Reply and tell me what you’re making. I’m looking for anyone interested in talking about their own side-projects and maker journey, so speak up if you’d like to appear in Serial Maker. I’d also love to know what you thought of this issue, and what you want to hear about in the future. Check out the past editions if you missed them, and don’t forget to continue the conversation on Discord!
Until next time,
Craig 👋