When you let an AI agent write Clojure code, you expect it to leverage the language's superpowers—the REPL's interactivity, structural editing, format-preserving code manipulation, and the rich ecosystem of wrapper libraries. Instead, what you typically see is mediocre code written slowly, as the agent makes the same mistakes every developer learns to avoid.

I discovered this the hard way.

I keep seeing people share vibe-coded apps built on TypeScript/React + Supabase — seemingly the default recommendation from Lovable or Cursor. As a Clojure programmer, I can't stay quiet about this. In an era where AI agents are deeply embedded in the development workflow, that choice carries structural hidden costs that almost nobody is talking about.

Context Window Is the Bottleneck, and Framework Design Determines Burn Rate

When I told others that I am a Clojure programmer, they responded apathetically. Why so many people in Taiwan never heard of this great programming language? One day, an idea occurred to me that how about teaching some students?

The advertisement

The background of this post

I began to use Datomic seriously in my project at work from February 2019. I encountered certain performance issues and I solved them through disk cache and LU cache.

Background

I needed to do eight Excel ETLs at my project. At the beginning, I just implemented some of the ETLs without any design. I did not even implement schema validation, and then I felt the pain soon. After several re-writing, I abstracted out some idioms and patterns for ETL.