Hah, I think about this all the time. I think we subtly desire LLMs to be more and more deterministic and efficient. This is why one of the main uses of LLMs is building tools to make their job easier.

I made my own project, with one of the goals being discounting tokens, but found that the real goal was just ensuring quality and making things more programmatic.

https://ktext.dev

Basically ends up being agents.md in a schema driven yaml file. Thinking about extending it to also generate or replace skill.md.

I think the proliferation of markdown is cool, and lowers the barrier for entry, but it’s also very unpredictable and loose. I think over time we will drive these to be more like config files instead of free text.