Not as a tool. As a thinking partner. Here's what a day at my desk really looks like — late nights, half-formed ideas, terminals on fire, and an AI that keeps pace.
I don't open Claude with a task. I open it with the shape of a problem. "Here's where I am, here's where I'm trying to go, here's what I've already ruled out." That single sentence saves an hour.
The good prompt isn't clever. It's honest about what I don't know yet.
The first version is never the shipped version. Claude writes the boring 80% — boilerplate, scaffolding, tests — and I spend my attention on the 20% that's actually load-bearing.
It's the same instinct as cooking with a sharp knife. The tool is fast; my judgment is what makes it food.
It's 1 AM. The bug isn't where I think it is. I describe what I'm seeing — not what I'm assuming — and Claude proposes three causes I'd already dismissed. One of them is right.
The win isn't speed. It's the second perspective that I'd otherwise have to wait until morning to get.
Where the work actually happens.
I tell Claude the problem, the constraints, and what I've ruled out. Then I ask. Always in that order.
If the first answer feels glib, I say so. The second answer is almost always sharper than the first.
Run the code. Read the diff. Check the source. Claude is fast — verification is free.
I save what I've learned about the project, my preferences, my mistakes. Future me thanks past me.
Done is a feature. I ship the rough version. The next prompt is always honest about what fell short.
"The best part of working with Claude isn't that it's fast. It's that it doesn't get tired when I do."— A field note, after midnight