Skip to content
Download .md

How building with AI works ​

Claude can build an app quickly. Whether it builds the app you wanted depends on the order you do things in: talk first, plan, look at a picture of it, and only then build.

Why "just build it" goes wrong ​

It is tempting to type "build me an app for my club" and let Claude get on with it. Claude will build something. It is usually not what you meant, because nobody worked out what you meant.

When a first app goes wrong, it is nearly always for one of these reasons:

  • Nobody talked it through. Claude filled the gaps with guesses.
  • The job was too big. One huge request, instead of small steps.
  • Nobody checked. New parts got built on top of broken ones.

None of this is because Claude is bad at building. It is very good at it. But it can't read your mind, and it only knows what you tell it.

The steps that work ​

  1. Talk it through. Tell Claude your idea and answer its questions. Who is the app for? What should it do? What does it need to remember?
  2. Plan the first version. Decide what goes in now and what waits. This is called scoping. Start with the smallest version that is useful, and add more once people have tried it. NodeGX saves the plan inside your project, so it is still there next time. See Plan before you build.
  3. See a mockup. A mockup is a picture of the screens, made before anything is built. Changing a picture is quick. Changing a built app is slow. See See it before it's built.
  4. Build in NodeGX. Claude builds your app one part at a time. NodeGX checks every change before it saves it. See Building with Claude.
  5. Check it. Open the app in NodeGX and click through it as your users would. Claude checks screenshots of its own work, but only you know if it is what you meant.
  6. Change it. Ask Claude to fix what's wrong, or change it yourself in NodeGX. Most of your time goes here, and that is normal. See Changing things yourself.
  7. Put it online. When it works on your computer, put it on the internet so other people can use it. See Putting your app online.

For the next feature, go round again, starting with step 1.

Three habits ​

Talk before you build ​

A change is cheapest while it is still a sentence. Before Claude builds anything new, talk it through. If Claude says something you don't agree with, or don't understand, say so.

You can ask for this at the start of any job:

text
Before you build anything, ask me questions until you understand what I want. Then tell me in a few sentences what you are going to build, and wait for me to say yes.

One job per conversation ​

A long conversation fills up with old ideas, detours and decisions you have since changed. Claude reads all of it every time you send a message. That makes its answers worse, and it uses up your plan's limits faster.

So give each conversation one job. When it is done, start a new conversation for the next one. You lose nothing, because NodeGX keeps your app's plan in the project. More in One job per conversation.

Working is not the same as written ​

When Claude says "done", it means it has made the change. It does not always mean the app now does what you wanted. Before you move on, open the app and try the thing you asked for.

If it isn't right, say exactly what you see. A screenshot helps more than a long description. See When something goes wrong.

Where you do all this ​

Use the Code tab of the Claude desktop app. NodeGX works in the Code tab, not in Chat. You can do the talking there too: before Claude creates a new app, it asks you what the app is for and writes your answers into the project.

Advanced

The Code tab and the claude terminal tool are the same Claude Code, and they share one settings file (~/.claude.json). So connecting NodeGX once covers both. More in Claude Code in depth. For other tools, see Other AI coding tools.

Quick reference ​

  • Talk, plan, see a mockup, build, check, change, put it online. In that order.
  • Start with the smallest version that is useful.
  • Ask Claude to question you before it builds anything.
  • One job per conversation. Start a new one for the next job.
  • "Done" means changed. Open the app and check it yourself.

Written by Richard Osborne, main developer of NodeGX · nodegx.io