After creativity comes code, and for a simple reason. A poem can't be right or wrong. Code either runs or it doesn't. That makes code a language you can check automatically, and models trained with that feedback get very good at it very fast.
The consequence is that natural language now works as a programming language. You can describe an app to Claude or ChatGPT and watch it appear. Tools like Cursor and Claude Code are built for this. This is vibe coding, and it is where I now spend most of my time. I wanted to code for most of my life and never had the patience for it. Now I build tools every week.
What I tell people who start:
- A non-developer can build a complete application by working with a language model.
- You need to understand about 20 percent of what is going on. The model provides the other 80.
- Be specific and build in very small steps, one addition at a time.
- Use the model for everything: design, code, naming, opinions.
- Even when you get 99 percent of the way, the last 1 percent may need an experienced developer.
I call the method chunking. First ask the model for the big picture: what you want to build, which tools, which steps. Then build it one small piece at a time: make this button blue, then make it move. Ask for two things at once and it will forget one.
For most people this is the highest-return way to spend time with AI. A coding tool costs about $20 a month. The best exercise is to pick a problem you have and solve it.
Try this
Pick a problem you have, at work or at home, and build a first version that works. It takes about 30 minutes.
I wrote more about teaching this to around 500 people in Teaching 500 Non-Coders to Build Software.
Since 2024
Coding agents now write, run and fix code on their own for minutes or hours at a time. The method still holds: big picture first, then small steps. Since 2025 I have taught it to around 500 people who had never written code.
Sources
- Andrej Karpathy, the hottest new programming language is English (2023) and vibe coding (2025)
- GitHub, quantifying Copilot's impact on developer productivity (2022)