How I work now
I run agents like a small team. On a hot side
project I'll push close to a hundred commits in a
month: researched, planned, written, reviewed,
shipped, with me in the loop the whole way. Claude
Code, sub-agents, MCP servers, evals. Daily
drivers, all of it.
The interesting part isn't the prompts. (Sorry,
prompt engineers.) It's the
operations around them. Where to
put a harness. How to design feedback so the agent
actually learns from the codebase. When to spin
off parallel work. When to slow down and demand a
human review. It's the whole reason I keep harping on
my 90/90/90
theory: everyone ships 90%, and the last ten points
are the whole game.
Underneath all of it I'm the same person who likes
being in the thick of it: talking to users,
debating tradeoffs, pushing things out the door. I
believe in
servant
leadership. Software is easy. Humans are
the genuinely hard part, and the genuinely fun
part.