Guide · Building
How to build apps with AI
Building with AI means describing what you want in ordinary language and having a model write, run and correct the code, while you remain the person who decides what is correct. You still read the result, run it, and deploy it. That judgement is the skill being learned, and it is the part no tool replaces.
How long does it take to learn?
A first working app takes an afternoon. Enough judgement to build something you would let strangers use takes roughly six to eight weeks at a few hours a week. The distance between those two numbers is the whole subject, and it is the part most tutorials never mention.
The fastest path, in order
- 01
Set up a real workspace, not a chat window
An editor with an agent inside it (Claude Code, Cursor or GitHub Copilot), a terminal, and a Git repository from the first line you write. Version control is the cheapest insurance in software and the first thing beginners skip.
- 02
Learn what the model actually is
Not a search engine and not a compiler. A next-token predictor working inside a fixed context window. That one fact explains almost every strange behaviour you will hit, including why it confidently imports a library that has never existed.
- 03
Ship one small thing end to end
One page, one form, one thing it saves, live on a public URL in week one rather than week eight. A project that never leaves your laptop has taught you half the job and the easier half.
- 04
Shift from writing code to reading it
Read every change before you accept it. Ask the model why it chose an approach, and push back when the answer is thin. The first time you say no, do it the other way, you have stopped being a passenger.
- 05
Take the data layer seriously
Where things are stored, who may read them, what happens when two people write at once. Models are fluent at interface work and careless about data, and data mistakes are the ones that follow you into production.
- 06
Learn to debug without the model
When it loops on the same broken fix, stop prompting and read the actual error. Being able to work when the AI cannot help is what separates building from being built for.
Tools worth your time
| Tool | What it is for |
|---|---|
| Claude Code | Agent that works in your terminal and repository, runs commands and edits files directly. |
| Cursor | Editor built around AI, best if you want a familiar VS Code surface with an agent in it. |
| GitHub Copilot | Inline completion and chat inside most editors, the gentlest entry point. |
| Next.js | The React framework most AI assistants write best, because it is what they have seen most. |
| Vercel | Deploy from a Git push, free tier included. The fastest way to get a public URL. |
| Supabase | Hosted Postgres with authentication attached, so the data layer is real from day one. |
Prompt-to-app builders such as Lovable, Bolt or v0 are genuinely useful for a first draft and a real trap as a destination: they get you a demo quickly and leave you unable to change it. Use one to see the shape of a thing, then rebuild it where you can read every file.
Mistakes that cost people weeks
Accepting code you have not read
It works, so it goes in. Three weeks later nothing can be changed because nobody, including you, knows what any of it does. Read the diff every single time, even when it is boring.
Starting with the ambitious project
The marketplace with payments and messaging is not a first project, it is four first projects stacked. Ship something trivial and public, then add one hard thing at a time.
Prompting harder instead of reading the error
The stack trace usually names the file and the line. Ten seconds of reading beats four rounds of asking the model to try again, and it is how you learn the system rather than the assistant.