When does the system need a language transform?

LLMs can do so many things with language that it’s tempting to look for somewhere to put one in every product. But “the model can do this” doesn’t tell me whether the system needs it to.

Take a training scene (like in SkillSaige). The system has an objective, a current state, and rules about what can happen next. If a timer runs out, or the learner chooses an offered option, ordinary software can handle that. There’s no relationship in language to interpret. Asking a model what should happen would just make a known rule less reliable.

But suppose the learner answers in their own words. They might ask a question, refuse, agree reluctantly, or say something the designer never thought to put on a menu. The system still needs to respond in a way that serves the scene’s objective, but we also don’t want it to put up an invisible wall and FORCE the user back on track. And that’s a plausible place for a semiotic transform: take the learner’s expression and produce a representation the system knows how to use.

Or perhaps the system already knows what a character needs to communicate, but the line itself should vary naturally. That’s a different transform, with a different input and output. In neither case does the model need to decide what the scene is for or what state the learner is in. Those belong to the system.

So before I choose a model or write a prompt, I want to be able to say: here is the meaning that needs to move, here is the form it arrives in, and here is the form the system can use next.

If I can’t identify the next step that consumes the result, why am I asking a processor that’s stuck in the domain of only language to do it for me? If I don’t know what happens when its interpretation is ambiguous, I haven’t finished designing the system. And if there’s no meaningful transformation to make, I don’t need an LLM for that part at all (usually it’s that one, let’s be honest).