Ask a chatbot to teach you something and it will usually do the opposite: hand you a polished summary, you nod along, and a week later you remember nothing, because you never had to think. Jane Street, the quantitative trading firm, has published an open-source collection of agent skills whose first and only skill so far takes the opposite approach — a Socratic tutor that refuses to lecture and makes you answer questions instead. This guide shows how to learn anything with Jane Street's Socratic tutor skill, from installing it to running a full session.
What the Socratic Tutor skill is
The repository is janestreet/agent_skills on GitHub, under the MIT license. Checked on 2026-11-10, it had 2,593 stars and 151 forks, was created on 2026-11-03, and had 2 commits. The README says it plainly: "A collection of skills for AI agents and chatbots. Each skill is a directory containing a SKILL.md file with instructions for the agent." Right now the collection holds exactly one skill: socratic-tutor.
The skill turns an LLM into a strict personal tutor that teaches a topic of your choice using the Socratic method: a diagnostic, a skill map with levels 0-5, one question at a time, spaced review, and a final assessment. It works on any topic you pick — the README's framing is "teach me X" — whether the subject is statistics, a programming language, or music theory. The tutor teaches by asking questions that force you to reason, predict, and explain, not by lecturing.
Know this up front: the repository is seven days old, with one skill and two commits. That is not a flaw — the SKILL.md is a complete, carefully written specification — but you are adopting something new, not a mature library. There is no marketplace listing and no install command in the README; installation is a manual copy, covered below.
Install and use
The steps below follow the README. The clone command is the standard git form for the repository URL — the README itself gives no command — and nothing here was executed in this run (unverified):
git clone https://github.com/janestreet/agent_skills.gitWith a coding agent that supports the common SKILL.md format, copy the skill directory into your agent's skills folder, then ask the agent to use it — replace X with your topic:
teach me XThe skill's trigger phrases are "teach me", "tutor me", "quiz me", "drill me", and "help me master".
With a plain chatbot, there is no installation: paste the body of socratic-tutor/SKILL.md into the chat — or into a custom or system prompt — followed by what you want it to do. The same file works both ways: the instructions are plain text the model reads.
One guard is worth knowing first. The skill is built for an ongoing teaching session, not for answering questions: ask a one-off question ("explain X", "what is X") and it answers normally without starting the loop. If your intent is ambiguous, it asks one short question first ("Want a quick answer, or should I actually teach you this properly?"). A false start costs you an unwanted quiz; a confirmation costs one word.
How a session works
First, the diagnostic. The tutor splits your topic into 4-10 concrete subskills — fewer for narrow topics, a two-level tree for broad ones — and asks 3-5 probing questions, easy to hard, to place your starting level. It teaches nothing until the diagnostic is done, then shows the initial skill map, proposes where to start, and lets you override.
Second, the skill map. Each subskill is tracked at levels 0-5: 0 unaware, 1 heard of it, 2 can follow along, 3 can do with help, 4 can do independently, 5 can teach it or handle edge cases. The map renders as ASCII after every level change and whenever you ask.
Third, the teaching loop, repeated per subskill. The tutor asks one question at a time and stops the message there — no hints, no example answers, no teaching after the question. It favors questions that make you predict an outcome, explain why, spot the flaw in a plausible-but-wrong claim, apply an idea to a new case, or connect two ideas. A wrong answer gets a narrower question that exposes the contradiction, not the answer; only after 2-3 failed attempts does it explain directly, then immediately re-tests from a different angle. A right answer sometimes meets a devil's advocate ("are you sure? what about X?"). Short explanations appear only for genuinely new material, and where tools are available it sets practical exercises and checks your actual output.
Fourth, strict evaluation. Levels are earned, not given: a level rises only on demonstrated performance across two or more independent checks, at least one of them unprompted recall. Vague answers do not count, and the tutor never credits you with an insight you did not express. Spaced review re-tests earlier subskills without warning; failing lowers the level. Level 5 requires handling an edge case or adversarial question and explaining the concept well enough to teach it. The skill is explicit about tone: do not flatter — "Not quite — why would X happen if Y were true?" beats "Great try!".
The course ends only when every subskill reaches the target level — 4/5 by default, 5/5 for full mastery — followed by a final mixed assessment with at least one question per subskill and one integrative problem. Pass, and you get the final map and a summary of your weakest areas; fail, and it is back to the loop. During a session, "show progress" renders the map, "harder" or "easier" adjusts difficulty, "just tell me" overrides Socratic mode for the current question only, and "save state" / "resume from:" let you pause and pick up a course in a new session.
Where it fits, and where it does not
It fits when you want to learn something rather than get an answer: preparing for an exam, picking up a new programming language properly, or building durable understanding of a subject you will use again. The diagnostic and skill map show what you know and what you do not, and save-state lets a long course survive across sessions.
It does not fit one-off questions — the skill says so itself — or work you want the agent to do rather than teach you: this skill teaches, it does not implement. Because the repository is a week old with one skill and two commits, expect change: more skills may arrive, and the SKILL.md may be revised. Treat your copy as a snapshot and check the repository before starting a long course.
A Socratic tutor is slower than a summary by design: a session asks one question at a time and waits. That is the method, not a limitation. Want a quick answer? Ask for one. Want to learn the topic? Let the tutor ask the questions.
FAQ
How do I install Jane Street's Socratic Tutor skill? Clone the repository with git clone https://github.com/janestreet/agent_skills.git, then copy the socratic-tutor directory into your coding agent's skills folder and say "teach me X". The README gives no plugin or marketplace install command; the clone command is the standard git form for the repository URL, not a README quote, and nothing was executed in this run.
How do I use the Socratic Tutor skill in Claude Code? Copy the socratic-tutor directory into your agent's skills folder and say "teach me X" (or "tutor me", "quiz me", "drill me", "help me master"). The README names no agent-specific step — any coding agent that supports the common SKILL.md format works. The skill then runs its diagnostic, shows the skill map, and teaches one question at a time.
Can I use the Socratic Tutor skill with a plain chatbot? Yes. Paste the body of socratic-tutor/SKILL.md into the chat, or into a custom or system prompt, followed by what you want it to do. No installation is needed. The same instructions drive the tutoring loop, including the diagnostic, the skill map, and the final assessment.
How does the Socratic Tutor skill track progress? It keeps a skill map with levels 0-5 per subskill — from 0 unaware to 5 able to teach it or handle edge cases — rendered as ASCII after every level change and on request. Levels rise only on demonstrated performance across two or more independent checks, including unprompted recall. Spaced review re-tests earlier subskills without warning, and failing one lowers the level. Say "show progress" to see the map.
Is this an official Jane Street product? The repository is published under the janestreet organization on GitHub as "Agent skills from Jane Street" under the MIT license. Nothing in it describes the collection as a supported product, and this article claims nothing beyond what the repository itself states.
See the skill's directory entry for install, platforms, and quick start: Jane Street Agent Skills on aiskills.one.