Jane Street Socratic Tutor: How to Use the Skill Without Mistaking a Score for Learning

Jane Street's Socratic Tutor is a set of instructions for an AI assistant, not a separate model or learning platform. You can use it to practise a technical topic through questions rather than receive another explanation. The important setup decision is what you want to demonstrate at the end.

The project is attracting fresh attention on Hacker News, but this is a usage guide, not a claim that the repository launched today. There is no evidence here that installing the skill makes the underlying model a reliable examiner.

How to load Jane Street's Socratic Tutor

The official repository describes two routes. In an agent that supports the SKILL.md format, copy the socratic-tutor directory into its skills folder. In an ordinary chatbot, paste the body of the skill file into a conversation, followed by the topic you want to learn. The repository does not require a special Jane Street account or provide its own model service.

Read the instructions before installing them. A skill changes how an assistant behaves; it does not establish that every proposed exercise or factual answer is correct. For a first session, pasting the instructions into a fresh chat is a simpler experiment than changing a workspace you use for unrelated coding work.

The skill itself asks the assistant to diagnose your starting point, build a skill map and ask one question at a time. It reserves tutoring for an explicit request to be taught, rather than turning every factual question into a quiz. Its commands also let you request a direct answer or save a compact state for another session.

A worked setup for learning database transactions

Start with a narrow outcome instead of “teach me databases.” The following is a proposed exercise setup, not a transcript from a tested session:

Use Socratic Tutor to teach me database transaction isolation.
I can write SQL but often confuse isolation levels.
My goal is to predict which anomalies a transaction schedule permits,
then explain what change would prevent them.
Start with a diagnostic, one question at a time.
Use a small inventory table as the recurring example.
Separate general concepts from database-specific behavior.
For implementation-specific claims, use the database's current documentation.

That boundary matters. Someone can repeat the definition of an isolation level while still failing to predict what two concurrent transactions will do. A good exercise makes that difference visible, and a consistent inventory example reduces the effort spent learning a new scenario with every question.

For example, ask the tutor to show two transactions reading the same stock count before either writes. Before seeing an explanation, predict the final quantity and state the assumptions you are making about the database and isolation level. Then change one condition, such as introducing a row lock, and explain why your answer changes or does not change.

The objective is not to memorise the tutor's preferred wording. It is to make a prediction, inspect the mechanism and transfer the reasoning to a modified case. If the assistant does all three for you, the conversation may be informative, but it is no longer demonstrating your understanding.

How to check whether the tutor is following its own rules

The skill's assessment rules require repeated evidence before increasing a level, including an unprompted check. Treat those levels as the assistant's session notes, not an independent qualification. A model can still accept a vague answer, miss a contradiction or reward a conclusion that happens to match its own mistake.

Keep a short record with three columns: the question, your answer before hints and the evidence used to check it. For SQL, that evidence might be current database documentation or a small experiment in an isolated test database. Do not let the assistant count a corrected answer as proof that you could solve the original problem unaided.

A useful final check is to close the conversation and solve a different example without the skill. If you need the same hint again, record that dependency rather than preserve a flattering score. This is especially important when using the tool for interview preparation, where recognising an explanation and producing one under pressure are different abilities.

When to use a simpler explanation instead

Socratic questioning adds friction deliberately. It fits practice, diagnosis and identifying gaps; it is a poor default when you need a command reference or a quick clarification to finish a task. Use the direct-answer override when the questioning has stopped exposing useful uncertainty, then return to practice with a new problem rather than rehearsing the answer you just saw.

The useful part of this release is a visible teaching contract you can inspect and change. Its value depends on whether it makes your mistakes easier to find, not whether the assistant eventually declares you finished.

I’m Paras. I work across AI, engineering and technical storytelling.

I help technical teams turn tools and workflows into clear explanations that developers can actually use.

Book a 30-minute call