Jane Street's Socratic Tutor Skill: Setup for Claude Code

Jane Street doesn't publish much. The quant trading firm is famous for being quiet about almost everything except its internship puzzles and its OCaml obsession. So when janestreet/agent_skills showed up on GitHub on November 3, 2026, with a single file inside it, people noticed. By November 10 it was sitting at #4 on Trendshift's trending repos, and ten days in it had pulled past 7,900 stars and nearly 500 forks — all for one Markdown file and zero lines of executable code.

The file is socratic-tutor/SKILL.md. It's a Claude Code Skill (one of those SKILL.md instruction files that an agent reads and follows as a system prompt). This one turns Claude Code, ChatGPT, or any SKILL.md-compatible agent into a tutor that refuses to just tell you the answer.

What it actually does

Say "teach me recursion" or "quiz me on options Greeks" and the skill kicks off a fixed sequence instead of launching into an explanation:

1. Diagnostic. It decomposes the topic into 4–10 concrete subskills, then asks 3–5 probing questions across difficulty levels before teaching anything. You don't get content until it knows what you already know.
2. Skill map, levels 0–5. Every subskill gets tracked on a scale from "unaware" to "can teach it and handle edge cases." You see the map and can override your starting point.
3. One question at a time. No lecturing. Questions ask you to predict, explain, spot the error, or apply a concept — and the skill is told explicitly not to drop hints or example answers right after asking.
4. Spaced review. It periodically re-tests subskills you supposedly already passed, without warning. Get one wrong and your level drops. Nothing is permanent just because you got it right once.
5. Final assessment. Once every subskill hits the target level (4/5 by default, 5/5 if you asked for full mastery), it runs a mixed exam plus one integrative problem. Fail it, and you're back in the teaching loop.

The rigor bar is the part that separates this from a generic "be a tutor" prompt: a level only goes up after two or more independent checks, at least one of them unprompted recall. Partial or vague answers get probed further instead of waved through, and even correct answers sometimes draw a devil's-advocate follow-up to check you're not just pattern-matching.

There are escape hatches built in, because a strict tutor that can't be interrupted is just annoying: "just tell me" suspends Socratic mode for one question, "harder" or "easier" retunes difficulty, "show progress" redraws the skill map, and "save state" / "resume from:" lets you pick a session back up later.

Why it's one file, not an app

The skill map in the repo README is explicit about the design choice: this ships as plain instructions, no code, because that's what makes it portable. We've looked at that same skill-vs-MCP trade-off before in our skill vs. MCP comparison. The same SKILL.md body works as a Claude Code skill, as a custom instruction in ChatGPT Projects, or pasted directly into any chat window as a system prompt. Jane Street didn't build a tutoring app: it wrote down what a good tutor does and let the model execute it.

That's also the skill's most obvious limitation, and reviewers have pointed it out: a prompt tuned against one model generation is a moving target. As newer models get better at task decomposition and self-assessment on their own, instructions this explicit can start feeling redundant rather than load-bearing. For now, though, it's a clean demonstration of how much behavior you can lock down with careful prompting alone — no RAG, no fine-tuning, no agent harness beyond "follow these rules."

Using it with Claude Code

Installing it is copy-and-paste, which is the whole point of the SKILL.md format:

git clone https://github.com/janestreet/agent_skills.git
# make it available in every project:
cp -r agent_skills/socratic-tutor ~/.claude/skills/
# or scope it to one project instead:
cp -r agent_skills/socratic-tutor .claude/skills/

Restart Claude Code (or start a new session) and ask it to "teach me [topic]," "quiz me on [topic]," or "drill me on [topic]." Those are the exact trigger phrases in the skill's frontmatter — a plain "explain X" or "what is X" deliberately skips the tutoring loop and gets you a direct answer instead, so the skill doesn't hijack ordinary questions.

If you're not on an agent that reads SKILL.md files, you lose nothing: open the file, copy everything below the frontmatter, and paste it as a system prompt or custom instruction wherever you're working. The repo's README calls this out directly as the fallback path.

Should you install it

If you're using Claude Code to actually learn something — a new language, a math topic, how a protocol works — rather than to ship code, this is worth having around. It's the opposite of how most people use an LLM, which is to get answers fast. The entire point here is friction: it won't let you fake understanding by nodding along to an explanation, because it keeps circling back to check if you can reproduce the reasoning cold. For code-focused work it's less obviously useful, since you're usually not trying to be quizzed on your own codebase. But as a free, single-file way to make an agent stop being a search engine and start being a teacher, it's hard to argue with the price.

FAQ

What is Jane Street's socratic-tutor skill? It's a single SKILL.md file in the janestreet/agent_skills GitHub repo that instructs an AI agent to teach a topic using the Socratic method: diagnosing what you know, tracking mastery on a 0–5 skill map, asking one question at a time, running spaced review, and ending with a final assessment.

Is it made by Jane Street the trading firm? Yes. It's published under the janestreet GitHub organization, released November 3, 2026, under the MIT license. It's the firm's first public AI agent skill.

How do I install it in Claude Code? Clone the agent_skills repo and copy the socratic-tutor folder into ~/.claude/skills/ to make it available across all projects, or into .claude/skills/ inside a single project to scope it there. Then ask Claude Code to "teach me," "tutor me," "quiz me," or "drill me" on a topic.

Does it work outside Claude Code? Yes. Any agent that reads SKILL.md files (ChatGPT Projects and similar setups) can use it the same way. For plain chatbots without skill support, you paste the body of the file in as a system prompt or custom instruction.

Can I make it just answer my question instead of quizzing me? Say "just tell me" to suspend Socratic mode for a single question, or say "explain X" / "what is X" instead of "teach me X". Those one-off phrasings skip the tutoring loop entirely by design.

Does my progress carry over between sessions? Say "save state" to export your skill map and session data, then "resume from:" with that saved state in a later session to pick up where you left off. There's no built-in persistence beyond what you save and paste back in yourself.

Why did this blow up on GitHub? It went from a fresh repo on November 3, 2026, to roughly 7,900 stars and nearly 500 forks by November 13, reaching #4 on Trendshift's trending list on November 10. That's partly because Jane Street rarely publishes anything publicly, and partly because the "it refuses to give you the answer" framing traveled well on social media.