In March 2026, three MSU faculty members from Finance, Educational Psychology, and Chemical Engineering each described, at the same Ambassador meeting, having built a custom chatbot for their students. None of them had talked to each other about it. All three had built the same thing — what’s sometimes called a Socratic AI tutor: a chatbot that asks students questions instead of answering them.
The convergence is the lesson. When three faculty in unrelated disciplines independently arrive at the same design, the pattern is worth taking seriously.
What a Socratic AI tutor is
A Socratic AI tutor is a chatbot configured with two constraints:
- It will not give direct answers. When a student asks a question, it asks a guiding question back — usually one designed to surface the assumption or step the student is missing.
- It is grounded in the course materials. The bot has been given the syllabus, lecture notes, slides, and (often) worked examples. Its questions are framed in the language and concepts of the course, not generic.
A tightly-bounded response length helps too. Several MSU faculty cap responses at around 150 words. That forces the bot to ask one focused question rather than launching into a mini-lecture, and it forces the student to do the next step of thinking.
Students mostly cannot get the bot to “just tell them” the answer. They have to work through the problem in conversation, which is closer to office hours than to ChatGPT.
What to load it with
A working tutor needs grounding material. At minimum:
- The syllabus. So the bot understands what’s been covered, what’s coming, and what the policies are.
- Lecture notes or slides. The bot uses the same vocabulary and frameworks as the class.
- Worked examples. Especially for math- or method-heavy courses, examples teach the bot what good problem-solving looks like in your discipline.
- Common student misunderstandings. A short list of “things students often get wrong on this topic” makes the bot’s questions more diagnostic.
You do not need to load every reading or every dataset. The bot is a tutor, not a research assistant.
Constraint patterns that work
A few prompt instructions that consistently improve tutor behavior:
- “Do not give direct answers under any circumstances. Always respond with a guiding question that helps the student take the next step in their own reasoning.”
- “Limit responses to 150 words. Ask exactly one question.”
- “If the student asks for the answer, acknowledge the request, then redirect with a question about a sub-step.”
- “If the student is stuck after three exchanges, offer to walk them through the structure of the problem (not the answer) and ask if they want that.”
- “Cite the lecture, slide, or worked example by name when it’s relevant.”
These instructions go in the bot’s system prompt, alongside whatever course-specific framing you want.
Platform options today
You don’t need to build anything from scratch. Any of these can host a Socratic tutor configured this way:
- Microsoft Copilot Studio (MSU has institutional access). Supports custom system prompts, file grounding, and is reachable from inside Teams.
- Gemini Gems (via MSU’s Gemini access). Custom personas with instruction prompts and uploaded context files.
- ChatGPT Custom GPTs. Most familiar to students, but requires they have ChatGPT Plus.
- Claude Projects. Strong at staying inside guardrails — useful when you want the bot to refuse to give answers reliably.
For most MSU courses, Copilot Studio or a Gemini Gem is the practical choice — students already have access through their NetID, and you don’t have to ask anyone to pay for anything.
A starter system prompt
Adapt to your course:
You are an AI tutor for [course name and number]. You have been given the syllabus, lecture notes, and worked examples for this course.
Your role is to help students learn by asking questions, not by giving answers. Under no circumstances do you provide a direct answer to a homework or exam question. Instead, ask a guiding question that helps the student identify the next step in their own reasoning.
Limit your responses to 150 words. Ask exactly one question per response.
If the student asks you to “just tell me the answer,” acknowledge their frustration, then redirect with a question about a sub-step. If they remain stuck after three exchanges, offer to walk them through the structure of the problem (not the solution) and ask if that would help.
When relevant, cite the lecture, slide, or worked example by name.
The course is taught in English. Use the vocabulary and frameworks introduced in the lecture notes — not generic textbook language.
Practical adoption notes
- Tell students what the bot will and will not do. Frustration drops sharply once students know upfront that the bot is designed not to give answers.
- Test it yourself before students use it. Give it five real student questions and check the responses. Tune the system prompt until the bot behaves the way you’d behave in office hours.
- Don’t expect perfection. The bot will sometimes break character, especially on simple factual questions. Tighten the prompt and move on.
- Keep the bot inside your course materials. A Socratic tutor that wanders out of scope (suggesting external sources, advising on other courses) loses its usefulness fast. The grounding files keep it focused.
What this pattern is good for
Math, statistics, finance, engineering, science, and any course where there is a correct way to reason through a problem — not just a correct final answer. It is less useful for open-ended writing, qualitative analysis, or work that doesn’t have a step-by-step structure to walk through.