$ cat choices/claude-code.md
the call
I reach for agentic coding as the volume layer. Pointed at the right problem, driven by someone who can tell when it's lying. I don't hand it the map and walk away.
Producing code stopped being the bottleneck. An agent will write you dozens of plausible, well-formatted commits a day without coffee or ego. That’s real leverage, but only if the scarce thing is supplied separately: judgment. Multiply great judgment by a fleet of agents and you get something genuinely new. Multiply no judgment by the same fleet and you get a beautifully formatted disaster, faster. Anything times zero is still zero.
I don’t let an agent own a system no human understands. The moment generation outruns comprehension, you’ve traded a slow team for a fast one that’s accumulating debt nobody can read. If there’s no one who can review the output and smell when it’s wrong, the agent isn’t an asset. It’s a liability with a green test suite.
This site. I built im·a·cto with Claude. I brought the frame, the experience, and every call on what mattered and what to cut; the model did most of the drafting. That’s the pattern done right: a human holding the map, an agent holding the volume. I say it plainly in the footer because pretending the tool wasn’t in the room would be exactly the dishonesty this whole site argues against. — see: about · writing / Volume Is Free Now
AI expands who can build. It doesn’t replace the person holding the map. The engineer’s job moved from producing to directing: fewer people, each steering a fleet, each accountable not for how much they typed but for what they caught and what they refused to ship. Verification became the core competency. Knowing what’s true is the whole job now.
the gaps — what it costs even when it’s right
It goes fast at the wrong things too. An agent can’t tell a “move fast” problem from a load-bearing one. It floors both. The danger isn’t bad code; it’s confident, plausible, wrong code that passes its own tests.
You don’t die by outage. You die by entropy. Drop agents in without extracting the map first and nothing visibly breaks. The demos work, the line stays up, and every quarter the code drifts further from anyone’s understanding, until a “simple” change takes three weeks.
The test was written by the same thing that wrote the bug. Green doesn’t mean correct when the generator also generated the proof. Verification has to come from somewhere the model isn’t: a human, an adversarial check, a real user path.