There used to be a gap between the thing in my head and the thing on a screen. Design it, hand it over, wait, get something back that was nearly it. The gap is mostly gone now. I can go from an idea to something live in an afternoon, and the interesting consequence is not the speed. It is that the expensive part of my work moved.
Four moves, in a loop. Plan, build, verify, ship. Then again.
The rule underneath all of it: I don’t write the code, I direct it. The AI types faster than I can and never gets tired. What it cannot do is decide what is worth building, or when the result is good. That part stays mine. It always was.
Plan
Every session starts blank. The AI does not know my project, my stack, or why I picked the blue button yesterday. So the first move is to give it two things: a memory, and a plan.
The memory is one file it reads at the start of every session. The project in a paragraph: what it does, what it is built on, the decisions already made. Every tool has one of these now under a different name (CLAUDE.md, GEMINI.md, AGENTS.md). Same idea, a brain that survives the conversation.
The plan is the work broken into phases, and each phase into steps small enough to check. Not “build the app.” More like: phase one, get a form on the screen. Phase two, save what someone types. Every phase ends in something I can see with my own eyes.
Plans survive, conversations don’t. The plan goes in a file, so tomorrow, or a brand-new session, picks up where I left off instead of re-explaining everything.
Build
This is the part I already knew, in a new medium. I sketch first, roughly, so the idea is clear in my own head before I ask for anything. Then I ask for options.
Never one. Three. Trying things is close to free now, so I explore wide and then choose, the same way I would in Figma. Keep one, delete the other two, and tell the AI why I kept it.
That last bit matters more than it looks. If the decision is not recorded, next week I re-argue it from scratch, or the AI does. Today’s decisions should stay decided.
Verify
This is the step that gets skipped, and it is the one standing between a demo and a product.
After every small step I take a quick look. Does it run. Click it once. Move on. After every phase I do a real review: click everything, read what changed, check that it still holds together. Step checks catch typos. Phase checks catch the logic mistakes that pile up quietly.
Then the part that is only mine. Is the empty state handled. Does the error message sound like a person or a machine. Is the spacing right. The AI will happily call something done. Whether it is actually good is my call, and it was always the job. It just used to be buried under the typing.
Ship
I save the work and push it. About a minute later it is on a real web address. No servers to configure, no scripts to run.
Then the part that is easy to skip: send it to one person. A friend, on WhatsApp, two minutes, and watch what they tap first. A live link nobody opens is the same as no link at all.
Doesn’t the AI just do this itself now?
Fair question. There are agents that take the whole loop: give them a goal and they plan, build, verify and ship on their own.
Autonomy owns the how, not the what. An agent can type the code and wire up the deploy. It still cannot decide what is worth building, or whether the result is any good for a real person. Hand it a goal and accept what comes back, and I have handed off again, just to a different black box.
So agents do not retire the loop. They move it up a level: less typing, more deciding.
Then again
Back to the top. Plan, build, verify, ship, again. The first time round it takes a day. The tenth time it is a Tuesday afternoon.
Notice what is not in that loop: writing code. What actually transfers is not a language or a framework, both of which will be different in a year. It is the rhythm. Knowing when to plan, when to generate, when to stop and judge, when to ship.
Plan, build, verify, ship. Then again.