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."

My first experience with strong LLM like Claude made me want a lot of tokens for a lot of good ideas, then later I realized that I needed more than software for my project to take off, and I did not have social contacts, marketing, capital, etc. So, now I see AI like just a game. Perhaps this is just one data point that might generalize to solo founders.

The sore pain is here, LLM agents, NPM ecosystem. The timing looks perfect. There are mature solutions like Chromebooks. Some ideas: two computers in one; or two sudo levels, the higher just for system recovery and diagnosis.

The problem is not so much technical as of providing an easy path to adoption.