Ask HN: Onboarding LLMs or Students?

[sinsudo]

As a professor trying to build an onboarding and orchestrating systems for LLMs, the strong idea that is rising in my mind is the need to have a mental model to transmit through language.

There are two fundamentally different regimens depending of what is the goal: learning or just passing grades. If the goal is learning, the system should help you find the final answer, wherever when it is passing grades the system should gives you something like: this is the minimum viable structure to solve problems of type X, something similar to recipes.

Regimen-dependent prompt architecture

If regimen = LEARNING:

    LLM receives: "Never provide a final answer first. Always ask a question that moves thinking forward. If the student asks for a recipe, explain why that would undermine the goal. Reward reasoning traces over correct conclusions."

    Outputs look like: "What do you think happens at the boundary here?"

If regimen = PASSING GRADES:

    LLM receives: "Provide minimal viable structures. Assume the user needs to complete a known problem type efficiently. Prioritize clarity, completeness, and reproducibility over depth. Do not ask exploratory questions unless the user fails twice."

    Outputs look like: "For problem type X, the recipe is: (1) identify variables A,B,C; (2) apply formula D; (3) check condition E. Example attached."

[tstansel]

Is the goal really learning here? If so what prevents the students from just go straight to a separate LLM?

I think you need to get the students to buy in and then provide them a claude skill or something more integrated that teaches the why behind the llm answer. Today getting the what is incredibly easy and thats not going away, helping the student learn how to learn from the LLM, how to pull out the why and building tools to help with that seem like a better path to me