Short answer: Socratic Tutor is a free agent skill from the trading firm Jane Street that turns an AI chatbot or coding agent into a strict personal tutor. Instead of explaining a topic, it asks you questions one at a time: first a short diagnostic, then a skill map that tracks each part of the topic on a 0 to 5 scale, spaced review of what you already covered, and a final assessment. It's a single SKILL.md file in Jane Street's agent_skills repository, released under the MIT license. You can drop it into an agent's skills folder or paste it into any chatbot.
The skill took off fast. The Hacker News post from November 10, 2026 passed 700 points (Hacker News), and the repository had 10,106 stars on GitHub when we checked on November 18, 2026. Below you'll see what the skill does, how a session feels, how to install it, and when it's the wrong tool.
What is the Socratic Tutor skill?
An agent skill is a folder of instructions an AI agent loads when a task calls for it. The common format is a SKILL.md file: a short description up top that tells the agent when to use it, then the instructions themselves. If that format is new to you, our explainer on Agent Plugins covers how skills get packaged and shared.
Jane Street's skill is one of those files. Its job is to stop the AI from lecturing. The skill tells the model to "teach primarily by asking questions that force the user to reason, predict, and explain." That's the Socratic method, named after the Greek philosopher who taught by asking questions. Jane Street's README is upfront that the idea is much older than LLMs and that other tutoring prompts exist. It says its version was written internally and isn't knowingly derived from any of them.
How a session works
The skill gives the AI a clear script. Here's the order, straight from the SKILL.md:
- It checks that you want a lesson. If you only asked "what is X," it answers normally. If it can't tell, it asks one short question first.
- It splits the topic. The AI breaks your topic into 4 to 10 concrete subskills, fewer for narrow topics.
- It runs a diagnostic. You get 3 to 5 questions, from easy to hard, one at a time. It doesn't teach anything until the diagnostic is done.
- It shows your skill map. Each subskill gets a level from 0 to 5, drawn as a little progress bar, and the AI suggests where to start. You can pick a different starting point.
- It teaches by asking. One question per message, and then it stops and waits. When you're wrong, it asks a narrower question that exposes the gap instead of giving you the answer. It only explains directly after two or three failed tries.
- It reviews without warning. Earlier subskills come back as surprise questions. If you miss one, the level drops.
- It ends with a final assessment. Once every subskill reaches the target level, 4 out of 5 by default or 5 if you asked for full mastery, you get a mixed test with at least one question per subskill.
The levels mean something specific. 0 is "unaware," 3 is "can do with help," 4 is "can do independently," and 5 is "can teach it" and handle edge cases. The skill says levels are "earned, not given": you need to prove a skill at least twice, once from memory. It also tells the AI not to flatter you.
What it looks like in practice
Say you're learning Python and you type "teach me Python loops." Based on the skill's rules, the opening could go like this. This is our illustration, not a recorded session.
The AI first splits the topic into a handful of subskills, something like for loops over a list, range(), while loops, break and continue, and nested loops. Then it says it will start with a few questions to see what you know, and asks just one:
What do you think this prints?
for n in range(3): print(n)
If you answer "1, 2, 3," it won't correct you. It will ask something narrower, like what range(3) gives you when you turn it into a list. After the diagnostic, it shows your map, maybe range() at 2 and nested loops at 0, and suggests where to begin.
That's slower than reading a tutorial. That's the point. You leave each subskill able to do it, not just having read about it.
Commands you can use anytime
The skill gives you five commands to steer the session:
- "show progress" draws your skill map.
- "harder" or "easier" changes the difficulty.
- "just tell me" turns off the questions for the current question only.
- "save state" gives you a short block with the topic, your levels and notes on your mistakes, so you can continue in a new chat.
- "resume from:" followed by that block picks up where you left off.
Save state matters more than it sounds. Long tutoring sessions fill up a chat, and this lets you start fresh tomorrow without losing your progress.
How to install it
Jane Street's README gives two ways.
With a coding agent. If your agent supports the common SKILL.md format, copy the socratic-tutor folder from the repository into your agent's skills folder. Then ask it to "teach me" a topic. The skill's own description lists the phrases that trigger it: "teach me," "tutor me," "quiz me," "drill me" and "help me master."
With a regular chatbot. Open the SKILL.md file, copy the instructions below the header, and paste them into a chat, or into the custom instructions or system prompt if your chatbot has one. Then say what you want to learn.
The skill doesn't need any special software. It's plain text, so it works with whichever AI you already use. How strictly the model follows it can vary from one model to another.
When it's the wrong tool
The skill itself says it's "not for one-off 'explain X' or 'what is X' questions." If you need a quick answer to get unblocked at work, a tutor that quizzes you first will just slow you down.
It's also not a course. It doesn't pick what you should learn next or give you projects. And it's only as accurate as the AI running it, so check anything that matters against a trusted source.
A few more things to know about the repository:
- It's MIT licensed, so you can use, copy and adapt the skill.
- Jane Street isn't taking pull requests that change the existing skills or add new ones. If you improve it, the README asks you to keep your own copy. Issues are welcome if something doesn't work as described.
- It's shared, not sponsored. Jane Street says it publishes skills it uses day to day internally. 4Geeks has no connection to Jane Street, and nothing in this guide means Jane Street endorses any course, including ours.
Go deeper
- What are Agent Plugins: the open standard for packaging Agent Skills and MCP servers.
- What is Hermes Agent: an agent that turns finished tasks into reusable skills.
- Learn about artificial intelligence: a practical starter path for 2026.
- Best AI courses for beginners: how to choose a first course that fits your goal.
- AI tools for developers: the 2026 tools, sorted by what they do.
Learning with AI, on your own schedule
A tutor that asks before it tells is a good way to learn anything technical. If you like learning this way and want a structured path through coding and AI at your own pace, see how AI Flex works.
