Using AI to learn

AI tools can be useful study partners, but they can also make it easy to skip the thinking that produces learning. The goal is not to obtain a finished answer faster. It is to improve your understanding, judgment, and ability to solve the problem yourself.

A good rule is to attempt the task first, then ask AI to help you think, not to think for you. Three reusable skills support this approach.

A simple workflow

• Try first. Work on the problem or concept independently. Write down what you understand, what you tried, and where you became stuck.
• Choose a mode. Ask the AI to act as a Socratic tutor, as a novice whom you will teach, or, after your attempt, as a study partner that compares your work with a released solution.
• Work one step at a time. Avoid broad requests such as "solve this problem" or "explain everything."
• Verify. Check every claim against the course material, your own calculations, and small examples or counterexamples.
• Reconstruct. Close the chat and explain the concept or solve the problem again without AI. If you cannot do this, you have not finished learning it.

Install the skills

The ai-teaching-skills repository contains three skills in the open Agent Skills format. They contain no instructions specific to one AI agent. If you use a skills-compatible agent, install all of them globally with:

npx skills add bstellato/ai-teaching-skills --skill '*' -g

The installer will ask which supported agent should receive them. You can inspect the skill files and individual installation commands before installing them.

You do not need to complete a prompt template. Invoke a skill by name or clearly request its learning mode. The AI will infer what you have already provided and ask for missing context one question at a time.

Mode 1: AI as a Socratic tutor

The ai-socratic-tutor skill helps you move forward without taking over the reasoning. It asks for the course or topic, your learning goal, and your current attempt when these are not already clear. It then asks one focused question, gives one small hint, or briefly explains one prerequisite at a time.

A concrete request could be:

Request

Use ai-socratic-tutor to help me understand the first-order optimality condition for a constrained convex problem.

The worked convex optimization example shows the tutor guiding the learner through the necessary direction without writing the proof.

Mode 2: AI as a novice

Explaining an idea is one of the best ways to discover whether you understand it. With the ai-novice skill, you are the teacher. The AI asks what you will teach, then poses one short, non-leading question at a time. It does not supply missing definitions, facts, corrections, or answers.

A concrete request could be:

Request

Use ai-novice while I teach you the KKT conditions.

The worked convex optimization example shows the AI asking the learner to test an overbroad claim without revealing the correction.

Mode 3: study a released solution

Homework solutions are released with the problems. Reading a solution before attempting the problem teaches little. With the ai-solution-study skill, you make your attempt first, then give the AI the problem, the released solution, and your work, as pasted text, attached PDFs, a photo, or a notebook. It names only the first substantive gap between your attempt and the solution, asks you to repair it, and ends with a variation whose answer it withholds. It declines to walk through a solution when there is no attempt.

A concrete request could be:

Request

Use ai-solution-study on Problem 2 of the homework. The problem and the solutions PDF are attached; here is my attempt.

The worked example shows the study partner naming one gap in a linear-program reformulation and leaving the repair to the learner.

How to check what the AI says

No mode guarantees correct feedback. Before trusting an AI response, ask yourself:

• Assumptions. Did I state the relevant assumptions, dimensions, domains, and sign conventions?
• Logic. Can I justify every equation and implication myself?
• Examples. Does the claim survive a small example, edge case, or attempted counterexample?
• Sources. Does it agree with the lecture notes and course references?
• Independence. Can I reproduce the explanation or solution after closing the chat?

The AI can help expose gaps, but deciding whether an argument is valid remains your responsibility.

Keep the human part

AI is an additional study tool. It should not replace conversations with instructors, course staff, or classmates. Human discussions provide context, judgment, and alternative ways of thinking that a chatbot cannot reliably reproduce.

The same principle applies when working alone. Preserve the productive struggle that builds understanding. Use AI to extend your curiosity and test your thinking, not to automate the intellectual task.

Further reading

• Stanford MS&E 125, Working with AI, distinguishes using AI to produce work from using it to learn. For learning, it recommends trying first and then asking AI to teach, hint, critique, or quiz.
• The MIT report on AI use in teaching, learning, and research training emphasizes productive struggle and using AI to augment learning rather than automate it.
• UC Berkeley's role-reversal classroom AI example has the student teach a novice AI and learn by explaining.