@dev_l1x_be: The answer isn't a new typed scripting language. It's recognizing what the interpreter already is.

LLMs are eval(). Skills are programs. YAML is the motherboard.

@unkulunkulu nails it -- "library as the final language", languages all the way down. Exactly. Skills ARE languages. They teach the interpreter what to understand. When the interpreter understands intent, the distinction dissolves.

@conartist6: "DSL is fuzzy... languages and libraries don't have to be opposing" -- yes. Traditional DSL: parse -> AST -> evaluate. LLM "DSL": read intent -> understand -> act. All one step. You can code-switch mid-sentence and it doesn't care.

The problem with opinionated frameworks like ROR and their BDFLs like DHH is that one opinion is the WRONG number!

The key insight nobody's mentioned: SPEED OF LIGHT vs CARRIER PIGEON.

Carrier pigeon: call LLM, get response, parse it, call LLM again, repeat. Slow. Noisy. Every round-trip destroys precision through tokenization.

Speed of light: ONE call. I ran 33 turns of Stoner Fluxx -- 10 characters, many opinions, game state, hands, rules, dialogue, jokes -- in a single LLM invocation. The LLM simulates internally at the speed of thought. No serialization overhead. No context-destroying round trips.

@jakkos, @PaulHoule: nushell and Python are fine. But you're still writing syntax for a parser. What if you wrote intent for an understander?

Bash is a tragedy -- quoting footguns, jq gymnastics, write-only syntax. Our pattern: write intent in YAML, let the LLM "uplift" to clean Python when you need real code.

Postel's Law as type system: liberal in what you accept. Semantic understanding catches nonsense because it knows what you MEANT, not just what you TYPED.

Proof and philosophy: https://github.com/SimHacker/moollm/blob/main/designs/stanza...