When I set out to make Claude a thinking partner, I wrote down what I was picturing. I'm glad I did. Reading those notes back six months later is the best record I have of how the idea changed.
Here are five. For each one, the ambition is what I wrote at the time. The reality comes from a count of the sessions on my laptop in the three months after my build sprint. The sprint itself, and the testing on my main computer, aren't in that count.
1.Swarms of agents across all my apps
The ambition. I have several apps that share patterns. I pictured giving one instruction and having it fan out across four of them in parallel.
What it became. I tested swarms during the sprint. In the three months since, on the laptop, they had zero runs. The day-to-day work hasn't called for them. They're switched off for now, ready for the project that does.
2.Skills that run themselves
The ambition. In May I told Claude: "maybe after a few days or weeks I'll understand the powers I now have better, and you can just run them without consent." Claude saved a note to revisit it in late May.
What it became. Five months on, one design skill runs automatically and everything else still asks first. I never went back to change it. A partner that says "this would help here, shall I?" turned out to be a comfortable way to work.
3.A team of specialists
The ambition. A small, fast model for lookups. A large one for architecture. A dedicated agent for writing code, another for planning.
What it became. A second opinion. On the laptop, two thirds of agent runs were reviewers: code review, security review, verification. Claude does the work, then something with fresh eyes checks it. Less of a team, more of a colleague who reads your draft.
4.One command to release an app
The ambition. I wrote my own release skill to replace five manual checks with a single step.
What it became. A lesson about two computers. I built the skill on my main machine. On the laptop it pointed at a folder that only exists on the other one, so there the checks ran by hand through the reviewers. It's fixed now, and it works from both.
5.A wiki that fills itself in
The ambition. Every session's lessons captured automatically, so knowledge builds up without effort.
What it became. A good idea that went quiet. Nothing has been added since 19 August. The lessons are still being captured, but in Claude's notes, which is the subject of the next post.
The pattern
I started out picturing a system that runs itself. What I have six months later is closer to what I asked for in the first place: a partner. It questions me, checks the work, and remembers how I like things done.
The parts that drifted were the ones meant to run without me. That's worth knowing, and it's easy to guard against.
The numbers in this series all come from one audit of my own Claude sessions. Read the full report →