The bottleneck now is me — my focus, and how many things I can hold in my head at once — not the machine, not its context window, its token budget, or how much the model can chew through overnight.
So the question I'm left with isn't "how do I squeeze more out of the model" but "how do I lay out my own attention." Because there are two roads. I can build a very autonomous setup, tell it "just build," and walk away. Or I can run two — for me, at most three — projects at once, leaving the tool some autonomy but keeping a hand on the plans and on what's actually happening. I tried the full letting-go: Claude drives itself, except I lose the thread, I stop understanding what's going on inside, and the token usage balloons, because the context window just keeps swelling. What works for me is the second mode, the one where I stay inside — not the one where I hand the model everything. I step into different roles: designer, note-taker, a bit of a sparring partner for the architect. And then the bottleneck is me again, and my mental capacity.
And here I hit the thing that helped most. I now know this about myself: I can't run three projects at the same stage. Three lots of architecture, or three lots of design at once, and my head hangs — everything looks the same and I get lost in it. So I started splitting them across stages. One project is in the thinking phase — I dictate notes into Malinche, and it turns them into a summary. The second is in design, the third in code, the fourth in testing. I switch between four different head-modes without pain — not between four identical ones. I surprised myself with how much it changed.
The rest fell into place around that. I work better in the CLI than in windows full of clicking — fewer things fight for my attention, I'm just buried in the terminal. The default one on the Mac is barely readable, even configured, so I moved to Ghostty and set it up for maximum legibility on a big screen. I split a 27-inch monitor into two or three windows, two similar projects side by side, and in each one two tabs: one for talking to Claude, the other a terminal for the commands I don't want to hand it. Nothing fancy, but all of it shaves off the small frictions this work has plenty of anyway.
The other half of looking after my head is breaks. It's gotten to the point where I have to design a break as carefully as the work. I only rest away from the screen — not at the computer with yet another tab open or the phone in my hand, because then my brain keeps doing the same thing: reading signs, connecting threads, gluing. I really cut off only once I disappear for an hour or two. Usually on the bike.
I'm recording this note on the bike, by the way. I ride and dictate a thought that surfaced on its own — because it's only here, once I stop chasing them, that the thoughts start showing up at all.