Jane Street open-sourced a tutor: it doesn't give you answers, it only asks you questions

Jane Street has put a skill called socratic-tutor on GitHub. The repository is called agent_skills and is MIT-licensed. It doesn't train a model, nor does it call some mysterious API. The core is a SKILL.md that constrains an LLM into a strict personal tutor: first diagnose what you already know, then produce a 0-to-5 skill map, ask only one question at a time, force you to reason through questions, and finally schedule spaced review and a final test.

For developers, don't treat it as an agent framework. It is a behavior-constraint file loaded through the host tool's skill mechanism. The value is in the process design, not the amount of code. For ordinary users, the barrier is also low: copy, upload, start chatting.

Three ways to install it:

Claude: Customize -> Skills -> Upload a skill. Package the socratic-tutor folder together with SKILL.md into a zip, upload it, and enable it. After that, use Teach me... or /socratic-tutor.
ChatGPT: Create a new Project and paste the body of SKILL.md into the project instructions. Every new conversation in that Project will start in tutor mode.
GrokBot: In the comments of the original post, use @bot to post Crée un bot et charge le skill.

It can also take in materials. Send it a balance sheet, a chart, a supply-chain diagram, or just drop in a PDF or course content, and let the tutor ask questions around the material. The original post says you don't need the strongest model; a medium reasoning setting is enough, because the key is asking the right questions, not doing research.

The poster listed a set of topics good for trying it out, leaning toward finance and semiconductors: reading a 10-K, understanding free cash flow, PER and multiples, DCF, options and time value, the chip chain from wafer to CoWoS, why HBM became AI's bottleneck, data-center optical networking, second-order effects of AI capex, and the valuation controversy around Ams-Osram. Trying these topics will show you the skill's temperament much better than asking it about today's weather.

The most interesting part is that the barrier is low to the point of being a little funny. A quantitative trading firm wrote down how to teach as Markdown, put it on GitHub, and now it can become a tutor inside several chat tools. It doesn't solve model hallucination, nor does it build a knowledge system for you. What it does is process: diagnose, ask, review, assess. It's useful for self-learners, and possibly annoying for people who want answers directly.

Someone in the comments asked whether it remembers where you left off last time. The original post doesn't give a clear answer. Based on the usual way Projects and skills are used, memory depends on the tool itself and how you organize the session; don't assume it automatically has long-term memory. Others worried about the safety of downloading from GitHub. This repository is a public Jane Street repository, but before downloading any skill you should still read through the file contents, especially anything you're going to paste into a system prompt.

The repository is here: janestreet/agent_skills.