Alex Lieberman Alex Lieberman Managing Partner of Tenex
AI

I Used 664 Million Tokens Built My Favorite App Yet

I Used 664 Million Tokens Built My Favorite App Yet

664 million tokens later, I have built my favorite custom app yet. Over the last four days, I built an app called Bricker. It is an AI-powered LEGO set creator. You go to this website, you describe any LEGO set that you can imagine, you specify how many pieces and what your skill level is. Within 30 seconds, it designs a custom LEGO set that doesn't exist anywhere in the world.

It breaks it down step-by-step with instructions. It tells you the exact list of pieces that you need to build the set, and it gives you a 3D interactive model to zoom in, rotate, look around, and visualize your set in 3D. The app works, it has actual users, it's connected to a database, and the best part is that I don't know how to code.

A Trip to the LEGO Store Sparked the Idea

About a week ago, I took my two-year-old son, Leo, to a LEGO store in New York City. We walked in, and we were just blown away by how much stuff was in this store. But after spending 30 minutes there, Leo turned to me, and he's like, "Dada, where are the garbage trucks?" and "Where is the street sweeper?" Leo is obsessed with big vehicles that you see in New York City: garbage trucks, street sweepers, fire engines, construction vehicles.

I went around the store looking for a garbage truck, and I found a total of one set. And I thought to myself, LEGO is one of the most successful companies in the world. They're doing close to $10 billion in revenue, but their entire business model is constrained by physical inventory. They can only make and sell sets that have enough mass market demand to justify producing millions of them.

So I asked myself, what if anyone could build any LEGO set that they could imagine on demand, in seconds, tailored to their exact skill level? That's how Bricker was born.

How Bricker Works in Three Steps

The concept was simple. Step one, you input an idea. For example, a vintage green pickup truck with a dog in the back, 150 pieces, intermediate difficulty. Step two, an AI creates the 3D model, the step-by-step building instructions, and the parts list. Step three, you can explore the 3D model right in your browser, step through each instruction, see what pieces you need for each step, and even order the exact pieces from third-party LEGO marketplaces like BrickLink or Pick a Brick.

I Am Not an Engineer and I Don't Know Code

Before I tell you how I built this, let me give you a quick reality check on my technical background. I am not an engineer, I didn't study computer science, I don't know JavaScript, I don't know Python. My background is in media. I co-founded Morning Brew, which is a newsletter company. I know how to write, I know how to edit, and I know how to tell stories. But over the last several months, I've become obsessed with how AI can democratize software creation. I've been experimenting with every tool under the sun, and for this project, I pushed these tools harder than I have on any other project I've worked on so far.

Why Claude Code Is the Most Powerful Coding Tool

The main tool that I used is called Claude Code. It's Anthropic's developer-focused tool, which basically runs in the terminal. And in my opinion, it is the single most powerful coding tool on the market right now. Unlike tools that try to hide the code from you, Claude Code embraces the code, but it acts like a 10x engineer that you can talk to in plain English.

Over the course of four days, Claude Code consumed 664 million tokens to build this app. To put that in perspective, that is roughly the equivalent of reading the entire Harry Potter series 600 times. That is how much code and context was being processed back and forth to bring Bricker to life.

In terms of cost, it cost me about $1,200 in Claude API credits. Now, $1,200 is not nothing, but compare that to what it would cost to hire a software engineer or a development agency to build a full-stack web application with a 3D rendering engine, AI generation, and a database. You're talking about tens of thousands of dollars, easily $30,000, $40,000, $50,000, and it would take months, not four days.

I've also explored 11 powerful Claude Code use cases from ultra-users that go beyond traditional engineering work.

The Tech Stack Behind Bricker

The rest of the stack looks like this. For the front end, Next.js with React and Tailwind CSS. For 3D rendering, this was the hardest part. I used Three.js. If you've never heard of Three.js, it's a JavaScript library that lets you render 3D graphics in the browser. It is notoriously difficult, even for experienced developers, but Claude Code handled almost all of the heavy lifting. For the database and user authentication, I used Supabase, which is an open-source alternative to Firebase. And for hosting, I used Vercel, which lets you deploy a full-stack app with one click.

Three Technical Challenges We Had to Solve

There are three big technical challenges that we had to solve to make this work.

Challenge number one was the AI generation pipeline. When a user types in, "Build me a vintage pickup truck," how does the AI actually create a buildable LEGO set? You can't just ask ChatGPT to give you a LEGO set, because it will hallucinate pieces that don't exist, it'll create shapes that violate the laws of physics, and it will give you steps that don't make any sense. So, we had to build a structured prompt system that forces the AI to output valid LEGO data. Every piece has to have a real LEGO part number, a valid color, precise 3D coordinates, and a logical step in the build sequence.

Challenge number two was the 3D rendering. Rendering a LEGO brick in 3D sounds simple until you realize that a LEGO brick isn't just a cube. It has studs on top, it has tubes underneath, it has bevels, it has specific proportions. If the studs are slightly off, it doesn't look like LEGO. It looks like a cheap knockoff. And then you have to render hundreds of these bricks together, without crashing the browser, and allow the user to rotate, zoom, and pan around the model smoothly. We ended up building custom Three.js geometry for the studs and the bricks, and the result looks shockingly close to real LEGO.

Challenge number three was the building instructions. Think about a real LEGO instruction booklet. Every step highlights the new pieces that you need to add, it shows you where they go on the model, and it ghost-outs or changes the color of the pieces that are already placed so you don't get confused. Replicating that experience in a web browser was a massive challenge. We built an interactive step-by-step viewer where you can click "Next Step," and the 3D model will dynamically highlight the new pieces for that step, show you a parts card with the exact pieces that you need to pick up, and let you inspect the build from any angle before moving on to the next step.

Prompt Engineering Is Software Engineering Now

Lesson number one: prompt engineering is software engineering now. The biggest breakthroughs in this project didn't come from writing better code. They came from writing better instructions for the AI that was writing the code. When the 3D models were looking weird, it wasn't a bug in the Three.js code. It was a failure in the prompt that told the AI how to place the bricks in 3D space. The moment I started treating my prompts like code, with version control, testing, and iteration, the quality of the output 10xed.

The AI Gets Tired on Large Projects

Lesson number two: the AI gets tired. I know that sounds crazy, but when you're working on a massive project with millions of tokens, the AI's context window gets cluttered. It starts forgetting things that you decided two hours ago, it starts introducing regressions, breaking things that were working fine earlier. The solution to this was aggressive modularization. I broke the project into small, independent components: the 3D renderer, the instruction engine, the parts catalog, the user interface. Whenever I was working on one piece, I made sure Claude Code was only looking at that piece and not the entire codebase all at once.

The Gap Between Idea and Product Has Collapsed

Lesson number three: the gap between "I have an idea" and "I have a product" has basically collapsed to zero. A year ago, if I had this idea for Bricker, it would have stayed in my Notes app. I would have thought, "That's a cool idea, but I don't know Three.js, I don't know how to build a 3D engine, and I'm not going to spend $40,000 to hire someone to build a prototype." Today, four days later, I have a working product that people are using. My son Leo has already seen a custom garbage truck that was built specifically for him, and he was completely mesmerized by it.

This shift reminds me of building a 17-slide partnership deck in under 10 minutes with Claude Code. The tools are compressing what used to take weeks into hours.

Four Days and $1,200 Changed Everything

Right now, Bricker is in a private beta with about 50 users who are testing it and giving feedback. The next big milestone is integrating directly with BrickLink's API, so that with one click, you can take any custom set that you've generated and order the exact pieces from real LEGO sellers delivered to your doorstep.

What I learned from spending four days and 664 million tokens building this thing is that the tools have gotten good enough that your ideas don't have to stay in your Notes app anymore. If you can describe it, you can probably build it.

If you want to check it out, follow along on my X account where I've been posting videos and updates of the build process.