85% of My AI Projects Are in the Trash — And That's Exactly Why the Rest Work

85% of my AI projects are in the trash.

And that is good :D

It's exactly why my current projects are better and built faster.

The prompt that started it all

A year ago, when I prompted a coding agent, it looked like this:

"Create a NextJS webapp with a React frontend and Supabase backend to do X"

A weak prompt fueled with hope that the AI gets it right, rather than filled with value for the coding agent. But I was missing the knowledge to ask for the correct things.

I didn't know what a proper database schema looked like. I didn't know why certain patterns break on serverless. I didn't know when to use an API route vs a Server Action. I didn't even know what questions to ask.

So I'd prompt, watch it build something, run into a wall, try to fix it, hit another wall, and eventually throw the whole thing away.

Project after project. Into the trash.

What the trash pile taught me

Now? I spend most of my time creating a /specs folder before I write a single line of code.

That folder contains:

• Project overview
• Technical stack decisions (and why each choice was made)
• Frontend architecture
• Backend API specs
• Database schemas
• Design system
• Error handling patterns
• Testing strategy

Detailed markdown files, filled with everything the AI needs to build exactly what I want.

Where did I learn what goes in those files?

From 85% of failed projects that went straight to the trash. Every dumb idea taught me something:

• Why certain patterns break at scale
• How to structure context for better output
• What to specify vs what to leave open for the AI to decide
• When to use certain APIs and when they'll bite you
• How to handle edge cases before they become production bugs

For example, now I don't ask AI to create something that won't work on a serverless environment. But I had to learn that somewhere :D

FAFO engineering

None of this knowledge came from tutorials or courses. It came from FAFO — F*** Around and Find Out.

Me and my stupid ideas and experiments. Some succeed. Most don't.

But there's always one win from each project: the learning. And that learning compounds over time.

The specs folder I use today is the fruit of every failed experiment from the past year. Each failure deposited a small piece of knowledge:

• That failed e-commerce app taught me about database schema design
• That broken automation taught me about error handling at the edges
• That overengineered dashboard taught me that simple wins
• That half-finished CLI tool taught me about argument parsing and user experience

Individually, each failure felt like wasted time. Looking back, each one was a prerequisite for what I build today.

The compound effect of failure

Here's the thing about AI-assisted building that nobody talks about: the learning curve is not linear. It's exponential.

Project 1: You don't know what you don't know. Everything breaks.

Project 5: You start recognizing patterns. "Oh, I need to specify the auth flow upfront."

Project 15: Your specs are getting detailed. The AI produces cleaner code on the first pass.

Project 30: You're spending 80% of your time on specs and 20% on execution. The code almost works on the first try.

Each project — even the failed ones — adds to your internal library of patterns, anti-patterns, and hard-won knowledge. The AI gets better because you get better at telling it what to build.

Why you should start today

This is why you should just start building. Little projects. Big projects. Stupid projects. Projects that will definitely fail.

Six months from now, when you start a new project, you'll have something most people don't: a brain full of lessons that no tutorial can teach you.

The more projects you do, the better the next one will be.

The trash pile isn't waste. It's the foundation.

Start FAFOing.