If you can prompt it to completion, you can build it
· #building #ai
I'm not a programmer. I'm not technical in the traditional sense, and I don't plan on learning to code.
And over the past year I've shipped Mailsi, a Mac app for email, calendar and to-dos. Listen, which reads articles out loud on Mac and iPhone. RelayGraph, an attribution tool for links that get shared around. Pipeline Labs, signal-based outbound. Plus a pile of tools I built just for me, like Tandem, LimitBar, FocusDock and Still. They're all on my changelog.
The way I see it, if you can prompt the AI to completion, you can build whatever you want.
Why I'm not learning to code
Honestly, I don't see much of a reason to anymore.
When something breaks, I don't open the code and try to work out what went wrong. I get another agent, or a different model, to go in and correct the mistakes the last one made. My job is to know what I want, describe it properly, and keep pushing until it's actually done. That last part matters most. Most people stop at "close enough".
Same idea as my launch email agent. I didn't write that by hand either. I described the job, the rules and the edge cases, and kept going until it worked end to end.
The honest caveat
These products run with no issues right now. But I haven't marketed them yet, so they don't have thousands of people hammering on them and finding the weird bugs. Some are only for me. Others are built to hold thousands of users, and I'll find out how well they do when real traffic shows up.
AI doesn't write perfect software. But for someone like me, the gap between having an idea and having a working product has basically disappeared. There are limits today. But the models keep getting smarter, and every time they do, what I can build gets bigger.
Humans make slop too
People love to talk about AI slop. Fair enough, AI does make mistakes. But humans create slop as well.
If you're managing a team of 10 agents, sure, there might be some slop. If you're managing a team of 10 engineers, there's also going to be some slop. Nobody ships perfect code on the first go.
So the job is the same either way. Someone reviews the work, catches what's wrong and gets it fixed. With agents I just hand the mistakes to another agent and keep moving, and it's done in minutes instead of waiting on the next sprint.
Programming is dying, and fast
I think programming is dying, and it's dying relatively quickly. The people building with AI every day are starting to say it out loud. Two things I read this week stood out.
Proton's engineering manifesto. Joseph Cox (@josephfcox) of 404 Media posted an internal AI manifesto from Proton's CTO. It tells engineers to stop thinking of their job as writing code and to focus on helping the LLM write better code faster. It ends with "'Programming' is dead. Engineering is not, nor is expertise." Worth saying, 404 Media's reporting also covers pushback inside Proton, including engineers' security concerns. It's not a done deal everywhere.
Spawn's fleet of agents. jacob (@jsnnsa), founder of Spawn, wrote about running a ten-person company alongside a fleet of permanent AI agents that each own a part of the business, from the engine to finance to bug triage. He says they shipped 77,000 substantial changes in 30 days and spent $3.1 million on tokens in a month. The humans mostly handle taste, saying no, and pushing the agents to aim higher.
You don't have to agree with every line of either one. What's interesting is the direction. One is a big privacy company telling its engineers the job has changed. The other is a tiny startup that has already reorganised the whole company around agents. That's a real shift in how teams think about building.
What a VP of engineering told me
A good friend of mine is a VP of engineering at a pretty big tech company. I asked him straight up. Would you rather hire someone who knows how to code, or someone like me who knows how to work with AI and direct it to build whatever's needed?
He answered fast. He'd rather hire someone like me, who knows how to use AI to build things and solve problems. And he knows I'm not technical in the traditional sense.
That stuck with me. The skill he cares about is knowing what to build, breaking it into steps the AI can handle, catching when it goes wrong, and not stopping until it works.
What I'd tell anyone sitting on an idea
Stop waiting until you learn to code. Start describing what you want in plain English and get it built.
Be specific. Give it the rules and the edge cases. When it breaks, hand it to another agent and keep going. And actually finish the thing. Prompting it to completion is the whole skill.
I'm going to keep building in public here. If you're doing the same, I'd love to hear what you're making.