AI writes code fine. The bottleneck now is reading it: understanding what it wrote, catching the mistakes, and shaping the architecture.
When the model needs to understand intent, not just syntax, specification is the highest-leverage thing you can write.
Spawn agents too eagerly and the parent fills up with summaries of summaries. Better to let the parent decide and have subagents do the work.
The constraint was model legibility, and the look followed from it. I didn't start with the aesthetic; I ended up there.
Signal density goes up, but judgment still moves at human speed. The tools change without making the work disappear.
The data-entry tax isn't a design flaw, it's a design choice: the auditor's needs won out over the rep's.
The technical execution is real, but judgment, taste, and a tolerance for ambiguity are most of what gets a product out the door.
