Dev Notes
Dev Notes: March 13, 2026
I've been running four+ Claude sessions at once lately.
Thread one is chewing on the DevEx roadmap. Thread two is researching an architecture change. Thread three is something I threw in right before bed and haven't checked back on yet. Thread four is a side project idea I wanted to flesh out before I forgot it.
This is how I work now. It's fast. But it's also messy.
The problem isn't Claude. The problem is that these threads are islands. Project A can't see Project B. If thread two figures out something thread one should know about, I'm the one carrying that context manually. The AI is smart, but it doesn't know what the other version of itself just figured out.
TJ and I got into this on the podcast this week. There's no clean answer yet. My current plan is to wire up Open Claw on a DigitalOcean droplet, connect it through CloudFlare tunnels, and route messages through Discord. The Obsidian vault would sync headlessly as the connective tissue. If this works out, I'll probably move Open Claw to the NAS or a new dedicated machine.
It's ambitious for a plan that's still just a plan. But the problem is real: when AI becomes core to how you think, managing your AI sessions becomes its own cognitive overhead. The tool that speeds up your thinking creates a new coordination burden you didn't have before. Nobody really talks about that part yet.
This Week on Slightly Caffeinated
New Adventures, Iris Long Running Jobs, and Multi-Threaded Conversations
TJ found his next adventure: starting Monday at Luma Brighter Learning, working on Laravel and AI initiatives. I moved into a full-time DevX platform role. We cover both transitions, TJ breaks down why he rebuilt Iris's long-running agent tasks as daemon processes instead of queued jobs (graceful shutdowns and no more cascading timeouts), and we dig into the multi-threaded conversation problem from above.
New episode drops tomorrow (Saturday) at 8:30am Eastern.
What I'm Learning
This week I officially moved from a 20% side allocation on DevEx and AI tooling to a full-time role. I've been doing the work informally for months. Now it's the whole job.
The first week of a new charter is surprisingly tricky. The instinct is to plan: write the roadmap, map the 60-day sprint, get stakeholders aligned. But planning too long means shipping nothing, and nothing kills a new team's credibility faster than an empty first few weeks.
My approach: find the most annoying daily friction points for engineers and start shipping fixes fast. The longer-horizon infrastructure work matters, but it needs quick wins to build trust first. Speed now, then depth.
Dev Tool of the Week
I've been using this to revamp a lot of my Claude skills, both building new ones and tightening up older ones that had drifted. It ships with four modes: Create, Eval, Improve, and Benchmark. Under the hood, four composable agents handle the work: an Executor that runs your skill against test prompts, a Grader that checks outputs against expectations, a Comparator that does blind A/B comparisons between versions, and an Analyzer that suggests targeted fixes.
The part that actually changed how I work: the eval tool. You define test prompts, it splits them 60/40 train/test, runs each query multiple times for reliability, and iterates on your skill description up to five times automatically. The result is a description optimized to fire when you want it to without triggering on things you don't. If your skills feel flaky or you've never stress-tested whether they trigger correctly, start here.
That's it for this week. If you've hit the multi-threaded AI session problem and found a solution, I want to hear it. Hit reply.
-Chris