Dev Notes
Dev Notes: June 15, 2026
Every plan to get a team using AI tools treats it as a distribution problem. Ship the tool. Run the training. Send the Slack announcement. Wait for the usage graph to climb.
It climbs partway, then stops. There's a floor under that curve, and no amount of rollout pushes through it.
I've watched this up close for a few months now. The engineers who already push small branches, write clear PR descriptions, and leave good review comments pick up AI tools fast. They were already getting their thinking out of their heads and onto the page, so handing that thinking to a model is a short step. The engineers who didn't do those things before don't suddenly start because Claude is in the editor. The tool doesn't create the habit. It assumes one.
So adding another layer of tooling widens the gap instead of closing it. Your strong adopters get stronger. Everyone under the floor stays put, and now they're further behind.
Here's the part the rollout plans miss: this is a confidence problem before it's a tooling problem. People under the floor aren't confused about the buttons. They're unsure of their own judgment, so they won't trust a model's output either. Documentation and announcements don't move that. Sitting with someone, live, through one real task does.
If your adoption numbers plateaued, stop shipping features at the problem. The next click won't move someone who doesn't trust their own read of the diff.
This Week on Slightly Caffeinated
New Grinder, Solo, and AI Burnout
TJ and I dig into Aaron Francis's Solo, then get honest about the other side of running this much AI: the multi-session burnout, the zombie-brain afternoons, and the missing sense of accomplishment when you ship 1,500 lines and feel nothing. More on Solo itself below.
Out now.
What I'm Learning
We run an AI reviewer via AWS Bedrock on every pull request within a few pilot repos, and for weeks it was the slow, flaky part of the pipeline. Ten to fifteen minutes a review, with enough timeouts that I get pinged daily about it.
I'd built it as an agent: hand the model some tools and let it explore the PR over 25 or 30 round trips. That felt like the modern way to do it. It was also the problem. Every round trip was another chance to stall, and a big PR would burn the whole budget and post nothing.
So I tried the opposite. Build the full context up front, hand it over in one shot, ask for a single reasoning pass. No tools, no loop. Review time dropped to about 90 seconds, with a five-minute worst case across a day and a half of live PRs. The prompt got sharper too, because I could watch every change land on real reviews instead of guessing.
An agent loop isn't free. For a bounded task, one good pass over the right context beats letting the model wander toward it.
Dev Tool of the Week
Aaron Francis's terminal app for running Claude Code (and other agents), with a multi-session orchestrator built in.
The unlock for me: a light Opus orchestrator spins up cheaper Sonnet worker sessions in the same project, hands each one a runbook, and routes them around shared files instead of letting them collide. I've been running my PR review skill in Solo work trees and doing my morning triage this way. The part that sells it: it drives your normal Claude Code CLI on your existing subscription, not the headless path, so you're not paying extra credits to orchestrate.
That's it for this week. Hit reply if your team's AI adoption hit a floor like this one. I want to know what moved people past it, or whether anything did.
– Chris