How can a non-technical founder build software with AI?

I'm not a developer. I code with my voice. Over about 18 months I've built three connected software products on my own, plus subscriber forms, a media kit builder, dashboards and an advertiser CRM. Local media brands pay about $400 a month for the newsletter software. Here's what actually works.

What does it really take?

Honest answer: a lot of hours and a real AI bill. I've been working 14-hour days and spending $600 to $800 a month on AI for 18 months, using both Claude and Codex. But the individual builds are shockingly fast:

• An admin backend for a survey and quiz tool: about 15 minutes, built against an existing design system.
• A media kit builder: four or five AI coding sessions over a couple of hours. I built an earlier media kit generator on my phone at a restaurant.
• A Kanban advertiser CRM inside my newsletter builder: about 30 minutes. Advertiser submissions from my forms flow onto the board.
• A dashboard pulling Meta ads data from nine ad accounts plus newsletter stats: Codex did it in one attempt after Claude Code failed at the same task.
• A post-sales-call automation that reads the call transcript, builds a proposal site, makes a payment link and sends onboarding once paid: under 48 hours.
• greetform, with lead pages, offers, integrations, multiple brands and custom domains: a month to a month and a half on the calendar while I ran everything else, but I'd guess about seven days of focused effort.

Another tool took me about 20 hours total, and most of that was talking to it, hitting enter and waiting. My team builds features about five times faster and cheaper than before. If you aren't paying $20 for Claude Code right now, get on it.

How do you get started if you've never touched code?

The thing nobody told me at the start: coding agents work from folders on your own computer, not in a cloud chat. Step one is boring and it matters:

1. Make a home folder. In your user folder, create one called ai-projects, all lowercase. Inside it, make a folder for each thing you're building, and point the tool at that folder.
2. Skip the terminal. The terminal version scared me. The Claude desktop app has a Code button now, so you just chat with it.
3. Ask regular Claude to walk you through setup. Tell it you're completely non-technical, you don't understand GitHub or coding, and you want every little step to get a working product. Sign up for a free GitHub account when it asks.
4. Make it interview you. My best prompt: interview me and ask me up to 20 questions about the product before you build it. It ends up with a much better grasp of what you want.

Don't be afraid of the word code. It's just an agent.

Being non-technical might even be an advantage. When I see something cool online, I screenshot it or paste the link and say: help me do this, I don't know how. The last three things I built started exactly that way. If you've never built anything at all, a hosted builder is a gentler first step: which AI app builder to start with.

How do you talk to a coding agent if you can't code?

Plainly. I don't fully understand pull requests or GitHub, and I don't need to. Supabase, Convex, GitHub and Railway are all confusing to me, so I just tell the agent to do them. When it starts talking in code, I've trained my brain to turn away and wait for the end, where it tells me in English what it did or what it needs.

The habits that make that work:

1. Let the agent lead. My early mistake was jumping straight to telling agents what to do. They're primed to accept what you say, so they just build your ideas. The agent is the one who sees the code. Have it ask you questions about the short-term MVP and the long-term vision, then let it tell you what to do.
2. Ask it to do things, not tell you how. Agents still tell me to go grab an API key. I reply: no, you do it. If you hit a wall, ask "is this something you're capable of?"
3. Ask it to spell out anything only you can do. When a step truly needs me, I want exact instructions, not jargon.
4. Don't let approvals stall the build. If something needs my sign-off, the agent sets it aside, keeps building around it and brings the list to me later.
5. Ask for the bro version. At the end I want what went well, what went poorly, any questions for me and what we should do next.
6. Ignore AI time estimates. When Codex or Claude says a task takes many hours, that's human developer hours. Just run it.

Copying is a superpower, too. To launch a new market, I tell Codex in plain English to copy an existing market's survey, offers and lead flow into the new brand, changing only the city name, logos and landing page. I don't even open the software.

What goes in a good brief for a coding agent?

When I hand an agent a real fix, the brief has four parts:

• Confirm the diagnosis first. Find the code and confirm what's wrong before doing anything else.
• Hard limits. Database changes must be additive, never destructive. Don't touch anything outside the system being fixed.
• One reason to stop. Ask me only if something costs real money, like a paid API or new infrastructure over $50 a month. Naming, file structure, libraries and UI details are the agent's call.
• Proof before done. Show me the real fixed example, test the edge case, and confirm nothing broke on the other sites that share the code.

What bites non-technical founders?

1. Parallel chats on one codebase. Sometimes they screw each other up. One build at a time per project.
2. "Fixed" doesn't mean live. The tool fixes things locally and doesn't say whether they're deployed. Even a GitHub "merged" badge doesn't prove production updated. See how to know an AI fix really works.
3. Made-up problems. AI tools report bugs that don't exist. I spent about 40 hours over four months tuning my agent instruction files and skills to cut that down. Worth every hour.
4. Too much code. Agents add rules, restrictions and code that don't need to exist. Ask for deletions: why AI agents write too much code.
5. The last 10%. AI gets you 90% done in an hour on work that used to take months. Easy apps still work great. Complex multi-step workflows still need heavy iteration.

What should your job be once it works?

I want to be the idea guy: brainstorming, feature planning and verifying at the end when it's supposedly all good and checked. I don't want to chat about code in any way. The goal is a system every chat already knows, where checks close the loop on their own. That saves me a ton of time, and it's likely much better technically too, since I'm not a dev.

When I bring on an operator, I clone my whole setup onto their machine: instruction files, brand guidelines and project context, set up to work with either Codex or Claude Code. They become an extension of me with fewer permissions. My real limit isn't capability, it's focus. After about an hour I tend to drop threads, so their job is to hold execution so nothing drops.

"This is way easier than you think... if you don't know something, which you won't, there will be a lot of things you don't know. You just ask."