Jane Street's Open-Source AI Tutor: The First Thing It Does Is Test You

Jane Street has put a skill package on GitHub. It is not a trading strategy, nor a quantitative model; it is a SKILL.md that turns a chatbot into a strict Socratic tutor.

The repo is here: janestreet/agent_skills, and for now it contains only one skill, called socratic-tutor.

How It Actually Works

According to the repo's instructions and the original tweet, the process is a fixed five steps:

Diagnose first, to figure out what you already know and what you don't
Build a skill map, breaking what needs to be learned into levels 0 through 5
Ask only one question at a time, without giving answers directly
Use spaced review to bring old material back
Finish with a comprehensive assessment

Of these five steps, step 3 most determines the experience. A normal chatbot, when you say "I don't understand X," gives you an explanation. This skill does the opposite: it asks you questions, waits for your answer, and then looks for holes in your answer. You learn slowly, but blind spots cannot hide.

Someone in the comments mentioned that what really makes content stick is the spaced review in step 4. That judgment is easy to understand: if the one-question-at-a-time rhythm is not paired with review, what you learn scatters as soon as you finish.

How to Use It

There are two ways, and neither requires writing code.

In a coding agent that supports the SKILL.md format: copy the skill directory into the agent's skills folder, then say "teach me [any topic]".

In a regular chatbot: paste the body of SKILL.md into the conversation, or into a custom system prompt, and then add what you want to learn.

Someone in the comments said they tried it in Claude Code, asking Claude to teach them everything about Claude, from zero to mastery, and their feedback was "currently works very well." This is a personal experience, for reference only.

An Unavoidable Question

The most valuable comment is not praise. A commenter asked directly: if "saving state" means copying text into a new conversation, how do you ensure the skill map and review plan are not lost?

This question hits not just Jane Street's one prompt, but the common boundary of the entire "interview-style prompt" category. At the same time, Thomas from NTNU in Norway posted in the comments the repo for their research grant application mentor: research-grant-mentors, with the system prompt public, and they are looking for other universities to discuss this approach. That document states its limitations more honestly than most product descriptions:

It only knows its own instructions and what you write; it does not go online to look things up
It has no memory across sessions; a new conversation starts from zero
If the conversation gets too long, it exceeds the model's readable range; the earliest messages fall out or are replaced by summaries, and it will not tell you
Its tone is equally certain when it is right and when it is wrong
Ask the same question twice, and the answer may differ
Language models tend to agree with the conversation partner; you have to actively ask it to find faults
It may fabricate references

Apply these back to socratic-tutor, and the conclusion is simple: the actual maintainer of that 0-to-5 skill map and review plan is you, not the model. If you want spaced review to actually happen, you have to save the map yourself and paste it back when you start a new conversation.

A Few Thoughts

This skill's value is not in "teaching better." What it does is push the cost of learning back to you: it does not give you answers; it forces you to say them yourself. For people used to having AI directly produce results, this counts as a kind of offense. A commenter put it plainly: "Asking only one question at a time is great, until you just want the answer; then it feels like being in class."

There is also an interesting reverse need: a commenter said what they wanted was the opposite, with themselves as Socrates and the AI as the student.

The same prompt can be read with completely different expectations. This probably also shows that the question is not whether the model can teach, but whether you are willing to be asked.