an exciting future with fast local LLMs is direct integration at the code layer mixing determinism with micro prompts or "intelligent expressions", e.g.

ai_if `${someString} contains a phone number` {
    num = ai_string(`phone number in ${someString} in XXX-YYY-ZZZZ format`)
    return `The number is ${num}`
}

It’s like having an on-demand guy in your code expression

So each ai_if or ai_string goes to a local llm? Why not write a prompt to accomplish this?

Damn that's the birth of new coding languages. Higher level programs, more intuitive ways to "talk" to a computer.