The Starting Point
It began with a nearly empty directory: just a README with an idea. Bernard had a clear vision—a small expression language that could compile to Ruby, JavaScript, and SQL—but no code. The first prompt was simple:
“Read README. I want to invent a small language that compiles to Ruby, Javascript and SQL. Let’s start with arithmetic expressions involving scalars and (open) variables.”
What Worked: The Task File Pattern
The most effective pattern we discovered was the task file. Before each significant change, Bernard would write a short markdown file describing the problem, the idea, and follow-up refinements after seeing results.
Vision vs. Execution
A clear division of labor emerged naturally:
- Bernard provided the vision, domain expertise, design principles, and course corrections
- Claude handled implementation details, proposed architectural solutions, wrote the code, and maintained test coverage
Here’s the remarkable part: Bernard didn’t write a single line of code. Not one. He wrote specifications in markdown files, reviewed outputs, and provided feedback—but never touched the TypeScript.
The Result
In roughly 24 hours of collaboration, we built a complete expression language with a parser, type system, three compilers, a standard library, a CLI tool, and a documentation website. Not bad for a day’s work.
Elo isn’t just a demonstration that AI can write code. It’s a demonstration that humans and AI can build together—each contributing what they do best.
The complete history of our collaboration is preserved in the task files on GitHub.