I’m experimenting with this variation of the Ralph Loop.

• Explain the overall need to the machine. It doesn’t matter if it is a new system or an existing system. Write it down in markdown. Save it and version control it so that the dog doesn’t eat it.
• Ask the machine to iterate on the docs (if there are any, or create them) and the code, and ping you back after each iteration.
• You’re done when 1) the system that solves the problem; 2) you have full understanding of the system.

Can LLMs simplify, not just add? In February 2026, I am suspecting it. I’m excited about the possibilities. What if LLMs could help me get something through my thick skull and let me actually understand what I’m building?

Glossary:

• Machine: SOTA LLM coding agents