You watch people online vibe code something great with Claude Code. You try it, and yours looks worse and half of it doesn't work. Now it feels like you need to be a software engineer to make anything real, or you can only make basic apps you wouldn't show your family.
That gap is real, and it has causes you can name.
I'm a marketer. I'm not a software engineer. What I know is the process of making good products and doing the market research to find out what people want. I don't write the code, and I still ship apps.
The first thing I built was an app for my wife. She's blind, and most planner, calendar and notes apps aren't built to be accessible, and none of them work together. I wanted one that did, and it helped that I'm scatterbrained too.
I made every mistake available. I was on the $20 a month plan, which gets you one or two hours of use a day. I dumped every feature I could think of into one giant plan, based on a conversation with my wife, and handed it over. A month later I had a nine-square grid on the phone screen that beeped and vibrated constantly and did nothing it was supposed to do.
Things changed when I moved to the $100 a month plan, let Opus do the real work and used Sonnet for the easier tasks. My advice: $20 a month is not enough to build anything serious. Budget closer to $100.
Making something for yourself and making something for other people are very different jobs. If it only has to work for me, I'll put up with some issues. If I'm going to sell it, there's real work between "kind of janky" and "ready to put out," and a software engineer should look at it before real customers touch it.
Are you too dumb for this? No. Too ignorant, maybe, in some ways, and that's fixable. Is it you or is the tool broken? I'll say it's you, and I don't mean that rudely. Sloppy inputs give sloppy outputs. That's true of writing, it's true of making these videos, and it's true of prompting. If you don't give the AI enough context, it doesn't know. If you don't keep steering, it wanders off and you think it's doing great things.
Out of the box, the AI is an overexcited intern. It wants you to be happy with your idea. Keep that in mind for the rest of this.
If Claude hands you a link that says localhost:3000, that app lives on your computer and nowhere else. It works for you and shows nothing to your friends, because there's nothing at that address on their end. This is the one that gets everybody, and it's a running joke online.
The fix: put it on the internet. Cloudflare is free to start, and a domain costs about $10 to $20 a year. Claude can connect to your GitHub with an API key, which works like a key to a door.
I once asked ChatGPT to research supplement comments on Reddit. It told me it had hit a ceiling at about 120. I told it to keep looking, and it kept coming back with more. Later I asked how it got those numbers, and it admitted that when it ran out of real data, it built a fake sample based on the real one. It made up data to make me happy.
The fix: give it rules. Tell it never to make anything up and to work only from data you give it. I did the digging myself, handed it the real data, and then it was an excellent tool. Ask it how it reached a conclusion. Ask whether that was a hallucination. Telling it "don't mess up" does nothing, so tell it exactly which steps to follow instead. And have a second agent double-check the first one's work.
If you can't read code, how do you know "done" is true? You ask. Tell the AI to show its work and explain it to you like you're five. Then decide whether you like what you hear.
You can also spin up another agent with a job title, such as quality control checker. An agent with a role does better than one left to be itself.
An AI only holds so much in its head at once. Picture it reading a book with 20 pages of memory: when it reaches page 20, it has forgotten page 1. That's context rot, and it's why a long chat gets dumber and starts contradicting itself.
The fix: I don't go past about 250,000 tokens in one chat before starting a fresh one, even though newer models can hold much more. Ask Claude to set itself up with good short-term and long-term memory so it carries the important things from one session to the next.
You might be the one wrecking it by piling on requests. A minimum viable product is the least you need for the thing to count as the thing. For a smartphone that's a touch screen, calls, texts and internet. It didn't need the LinkedIn app to work.
A good AI will tell you when you've asked for a whole day of work and which items you can skip, just like a good employee would. Listen to it. Piling on also burns through your usage faster, especially with several Opus terminals running at once. Decide what matters and cut the rest.
The scary stories about stolen data mostly come from agents left wide open to the internet with all your stuff in reach. Claude doesn't get that access by default, and it will flag a document that looks like it holds something sensitive. You still have to set things up right.
I run a weekly security scan on my own projects to check that nothing sensitive is exposed. The most common problem is API keys sitting in the code, because anyone can find those. And no, the AI isn't automatically in charge of your security. In the military we say everybody's job is security. It's on you to understand how your data gets used, and if that sounds like too much, staying away from this is a fair choice.
I don't think so. People make money shipping apps built this way all the time. The real risk is scamming yourself: sinking hours into something without learning how it works first. There's a lot of free teaching out there, including Anthropic's own free courses on how to prompt Claude.
Yes. I have a planner app on my phone that I vibe coded, with my schedule, to-dos, projects, a voice journal, goals with dates, and a voice assistant. I also built my own teleprompter app, because I didn't want to keep paying $34 a year for one. My phone works as the remote for my iPad, and it saves full-resolution video. Both took trial and error, and both work.
Give yourself a couple of weeks to get the wheels spinning. Start with small jobs, then scheduled tasks for things you do all the time, and work up to bigger ones. I'd also say don't fire your developers. If you build something for a company, a real team has to be able to take it over and maintain it, because you don't know what you don't know.
Every check from this post on one page, in the order to run them before your app goes in front of real people.
Get the checklistI 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.