If you want to build real software with Claude and you don't write the code, you're going to hit five walls. I hit every one of them.
I'm a marketer. I don't write the code. With Claude Code I've built a planner app that runs on my phone, a land navigation tool for soldiers, a Mac cleaner, a check printer, and a mileage tracker that runs on my own server.
This post is the map. For each wall I'll tell you what it looks like, what got me past it, and where to go next. Then I'll show you what I built and answer the three reasons most people never start. I'm not going to show you an income screenshot. I'm going to show you software that runs and what broke on the way.
I hear the same five things from almost everyone who tries this.
Which AI do I use? Is the free one enough, or do I have to pay? You watch someone build an app for 20 minutes and it looks great. Then a comment underneath says, "that's a demo, not a business." And they're right. Nobody showed you how to tell which idea is worth building or which tool is worth paying for.
What got me past it: asking the AI to argue with me before it built anything. I wanted a mileage tracker. The AI told me to just buy one. I overruled it, so it sent six skeptics after my plan, and I ended up running a free open source tracker on my own server instead of writing my own. The whole story is in Mileage Tracking Apps for Businesses: Build One or Buy One?.
If your question is the money side (free, $20 or $100), read Is Claude Code Free? first.
You try it yourself and the AI says it's done. It isn't, and you can't tell because you don't read code. That's worse than a normal bug, because the AI sounds sure. Or it does work, and it looks like every other AI-made website on the internet. One commenter put it better than I can: "How could you know?"
What got me past it: checking the work against something the AI didn't write. When I had Claude build check printing software, it filled in my bank's routing number from memory, and it was wrong. I caught it before anything printed by checking the number against a public lookup. That's the check printer post, and it's a plain way to check the work that doesn't need you to read code.
You've heard the stories. One report said an AI agent deleted 48,000 files in just over 100 seconds and then apologized. Typing "please make it secure" is not a security plan. The real question underneath is simple: who gave the AI that much access in the first place?
What got me past it: I built a Mac cleaner, which is about the worst place you could let an AI loose, and I gave it four guardrails so deleting isn't possible. Then I planted a delete line on purpose to try to break it. The check failed, and the app can't ship with that line in it. That's in Claude Code Deleted 48,000 Files.
You hit your usage limit in the middle of a job, or you're staring at the Max plan wondering if it's worth it. Everybody has an opinion. Almost nobody shows the real math.
What got me past it: watching the actual numbers. My AI chief of staff sent out more than 100 jobs in two days, and my bill didn't change. The math is in Is Claude Max Worth It?.
You spend an evening teaching Claude your project. You open a new session the next morning and it starts cold, like you've never met. So you sit there babysitting it, and the thing that was supposed to free you up becomes a second job.
What got me past it: the short version is this. Start a new chat before you hit about 250,000 tokens, and make Claude write down where it stopped so the next session picks up there. The long version is how I stopped babysitting altogether: I give one agent the list at night and wake up to decisions, not chores. That's the overnight agent post.
You should ask for proof, so here's only what actually exists.
I'm not telling you any of this made me rich. I'm telling you it runs, and I built it without writing the code.
Me neither. I'm a marketer. Claude writes the code.
Some developers will tell you vibe coding isn't for people with no programming experience, and they have a point. Somebody has to check the work. But checking the work is not writing code. It's asking the AI, "how did you come up with that?" It's giving it exact steps instead of "don't mess up." And it's having a second AI review the first one.
If you plan to sell what you build, pay an engineer to look at it before real users do. You can get a lot further than you think before you need one, and you'll know exactly when you do.
It can. I run mine on the $100 a month Max plan, and the Max post shows exactly what that buys. The expensive part is long sessions that drag old context along with them. So wall five is a money problem too.
Yes. Mine did. The difference between a demo and software you can use is what happens next: a check that runs on every build, and a rule that says "work only from the data I give you." You won't build something that never breaks. You'll build something that tells you when it did.
This is the order I'd go in:
Don't binge all of it. Pick the wall that stings most right now and start there.
And if you only do one thing tonight, ask the AI to search your code for passwords and keys and show you what it found. Demos rarely show that part.
If you're about to put something you built with AI in front of real users, run it through a checklist first. I made one. It's one page, it's free, and it's built from the five ways my own apps broke: the pre-launch checklist. Put in your email and it unlocks right there.
New builds go up every week, and I show what broke, not just what worked. Subscribe on YouTube so the next one on your wall finds you, and tell me in the comments on the video what broke first for you. I read every one.
I send what I'm building, what worked, and what broke. Free, no fluff.
Free. Unsubscribe anytime.I send what I'm building, what worked, and what broke. Free, no fluff.
Free. Unsubscribe anytime.I'm a marketer who builds his own tools. More at adamgarceau.com.