For a while my days with AI coding tools went one of two ways. Either I was arguing with the model — re-explaining, re-prompting, correcting the same mistake for the third time — or I had swung the other way and was pasting whatever it gave me with a quiet trust me bro and hoping the tests, if there were tests, would catch it. Both left me tired. One felt like babysitting; the other felt like gambling.
What actually changed for me
The shift wasn't a better prompt. It was a change of role. I stopped trying to be the author of every line and started being the steward of a loop — someone who sets the boundaries, asks for a plan before any code, and then runs a fast, local check that tells me the truth in under a second.
That's the whole idea, and it's smaller than it sounds. Give the agent a clear, written spec of what "done" means. Let it propose. Verify deterministically — a real test, a real run, on my own machine, not vibes. When it fails, feed the actual error back instead of a vague "that's wrong." The loop tightens. I stay calm because I'm no longer guessing, and the model stays useful because it's correcting against something real.
Why I keep it on the command line
Most of the friction I used to feel came from the gaps between tools — copying out of one window, pasting into another, losing context on the way. On the command line those gaps mostly close. The feedback is immediate, the agent can touch the files directly, and I can see exactly what ran. Fewer middlemen between intent and result. That immediacy is also what keeps the work humane: a sub-second loop doesn't leave you holding your breath waiting for a slow, silent process to maybe break.
Where this goes
I'm not interested in teaching anyone to memorize syntax. I'm interested in the handful of habits that make this loop trustworthy: write the spec, audit the plan, never trust without a test, capture the error and let it teach the next attempt. Once those are second nature, the size of the thing you're building stops being the scary part.
If this resonates and you want the longer, hands-on version — the setup, the verification protocol, building a couple of real things end to end — that's what I'm putting together next, and I'll be doing some of it live so you can watch the loop in motion. Follow along, and bring a small project you actually want to ship.