Quantitative trading firm Jane Street published the open-source agent_skills repository under an MIT license — a collection of "skills" for AI agents in SKILL.md format. The first and currently only skill in it, socratic-tutor, turns a language model into a strict personal tutor that diagnoses the student's level, maintains a skill map with levels from 0 to 5, asks one question at a time, and applies spaced repetition. Within a week of its creation, the repository gathered approximately 2,400 stars on GitHub.

image
image

What happened

Jane Street posted the agent_skills repository to GitHub under an MIT license on November 3, 2026, and by November 10, it had gained approximately 2,400 stars. The socratic-tutor skill works entirely through text instructions: upon first contact, it diagnoses the student's level, builds a skill map where each skill is rated from 0 to 5, then conducts training one question at a time, incorporates spaced repetition, and concludes the course with a final assessment. Text commands are provided for managing the pace, including saving the session state to continue learning in a different chat. The SKILL.md file is intentionally not tied to a specific model: it can be copied into an agent's skills folder or inserted into the system prompt of any language model.

Context

The repository has become a reference implementation of the growing Agent Skills standard — the SKILL.md format, which is supported by Claude and other hosts for AI agents. Here, it is pure prompt engineering: no code, no dependencies, and no infrastructure, so the agent's behavior exists as a transparent text file that is diffed in git and transferred between models by simple copying. The skill's pedagogical apparatus — Socratic dialogue, spaced repetition, mastery levels — relies on long-known heuristics from learning science; the repository contains no new architecture, training method, or benchmark. Community interest currently outpaces knowledge of the results: the discussion on Hacker News at the time of data collection consisted of a single item with no comments.

Why this matters for the industry

The main signal for the industry is not the tutor itself, but the template for packaging agent behavior: SKILL.md sets a way to distribute behavioral instructions as a portable artifact that is versioned in git, undergoes code review, and can be reverted, while the time to deployment is reduced to copying a file or inserting it into a system prompt. A reference from a large company with a strong engineering culture encourages teams to store such skills in repositories alongside code and to clone the pattern to adjacent domains like interviews and onboarding within days. Within a six-month horizon, a wave of imitations is likely: proprietary skill libraries, forks of socratic-tutor, and the first amateur comparisons of "skill vs. base prompt," while product wrappers will have to build what is missing in the skill — persistence, spaced repetition planning, and progress dashboards. Within a two-year perspective, consolidation of the agents' "behavioral layer" into a separate versioned artifact with evaluation standards and regression tests between model versions is possible, although the reverse scenario is also open: future generations of LLMs may adopt such pedagogical patterns "out of the box."

Why this matters for users

It can be tried today and almost for free: simply copy SKILL.md into your agent's skills folder or insert it into a chatbot's system prompt and ask "teach me X." Instead of a standard lecture, you get structured learning with progress control: level diagnostics, sequential questions, spaced repetition, and a final assessment, while commands like show progress, harder/easier, just tell me, and save state / resume from: allow managing the pace and transferring the session between chats. No new service, API keys, or separate payment is required — the cost is limited to the model's tokens, and the file itself can be freely modified and supplemented for your own purposes.

What is currently unknown / limitations

The skill's effectiveness has not been measured in any way: the sources contain neither pre/post knowledge tests nor a comparison with a base prompt, so the expression "turns any LLM into a strict tutor" remains a hypothesis of usefulness rather than a proven effect. The portability of SKILL.md as a file is not equal to the portability of behavior: the consistency of operation across different language models has not been measured or demonstrated. Approximately 2,400 stars are a signal of interest and distribution, not effectiveness, since there are no independent reproductions or comparative tests in the sources.

Sources

Author

Look at AI, editorial team