"You need to learn to code before you can build an app" is only half true now. I'm a working software engineer at a US tech company, and these days AI writes most of the code in my side projects. A web studio, a directory site, several iOS apps — all built and shipped that way.
This post is the map. Each big topic gets its own deep-dive; here we draw the full picture.
What's actually possible
As of 2026, AI coding tools can genuinely deliver:
- Websites / landing pages — the easy case. Idea to deployed in a day.
- Web apps (SaaS) — sign-ups, payments, a database. Takes longer, but one person can do it.
- iOS / Android apps — all the way through App Store review. I've done it multiple times.
- Automation — scripts and bots that do repetitive work. Often better ROI than building an app at all. (Covered separately in the AI automation map.)
Some things still don't work: high-scale architecture, serious security requirements, and deciding what to build. Those remain human jobs.
You need fewer tools than you think
My actual stack is simple:
- An AI coding tool — I use Claude Code. See Claude Code vs Cursor vs Codex for an honest comparison.
- Deployment — Vercel. Where the website lives. The free tier covers personal projects.
- A database — serverless Postgres like Neon or Supabase. Also free to start.
Curious about money? I published the actual costs from my own projects. Short version: beyond a monthly AI subscription, it's almost nothing.
The order matters more than the tools
The process I keep reusing:
- Validate the idea — check search demand and competition before building. Killing bad ideas here is where you save months.
- Define the smallest version — this step is about removing features, not adding them.
- Write a CLAUDE.md — a document that tells the AI your project's rules and scope. The difference with and without one is dramatic.
- Build → ship → iterate — don't finish then launch; launch then finish.
The full 10-step process, with the exact prompts I use at every step, is packaged as a free kit — including a version Claude can run for you as a skill.
Where people actually get stuck
For non-developers, the blocker is rarely the code itself:
- Environment setup — terminal, Git, connecting a deploy account. The first time is the hardest time.
- Hitting an error — pasting the error message into the AI solves most of them. The real problem is not knowing what to ask.
- Deciding direction — AI executes well, but it won't choose your priorities for you.
I'll unpack each of these in upcoming posts. New posts go out via the newsletter.
Posts in this series
- What is vibe coding? An honest explainer
- Getting started with Claude Code — the 30-minute guide
- How to validate an app idea — 5 checks before building
- Build a landing page with AI in one day
- What it actually costs to build an app with AI
- Claude Code vs Cursor vs Codex
- ChatGPT vs Claude for building apps
- No-code vs AI coding — which to learn
- What are Claude Skills? 4 free ones