Jane Street Open-Sources agent_skills With a Socratic Tutor SKILL.md

Jane Street — the quantitative trading firm known for OCaml-heavy engineering — published agent_skills on GitHub in early November 2026. The repo crossed roughly 2,100 stars within days, per the repository README and GitHub counter on November 10.

It is not a model release. It is a reference implementation of the same Agent Skills folder pattern AWS Quick adopted in its November skills guide and Anthropic popularized for Claude agents. The first skill, socratic-tutor, turns a host model into a strict personal tutor that teaches through questions, diagnostics, and leveled mastery checks.

TL;DR

Why a trading firm shipped a tutor skill

Jane Street's public engineering brand centers on disciplined learning — interns famously work through Socratic problem sets. Publishing a SKILL.md that encodes that pedagogy for external agents extends the same idea to any model host that reads the Agent Skills spec.

For explainx.ai readers, the interesting part is structure, not mystique about Jane Street's internal stack. The skill is portable text you can audit, diff in git, and reuse in Claude Code skills folders, Quick, or custom harnesses like Executor 2's MCP gateway.

Repository layout

The GitHub tree on November 10 listed:

• socratic-tutor/SKILL.md — primary instructions
• CONTRIBUTING.md, LICENSE.md, .ocamlformat, .gitignore

README copy: "Agent skills from Jane Street" — "A collection of skills for AI agents and chatbots. Each skill is a directory containing a SKILL.md file with instructions for the agent."

No Python package, no npm install — intentionally host-agnostic.

Inside socratic-tutor/SKILL.md

The published SKILL.md opens with YAML metadata:

name: "socratic-tutor"
description: "Structured Socratic tutoring when the user explicitly asks to be taught, quizzed, or drilled to mastery on a topic..."

Guard: do not hijack casual questions

The skill mandates a guard before the teaching loop. If the user only asks "what is X" or "explain X", answer normally. If intent is ambiguous, ask one confirmation: quick answer versus full tutoring. That prevents accidental quiz mode — a common failure when skills descriptions are too broad.

Session flow

1. Decompose the topic into 4–10 subskills (fewer for narrow subjects).
2. Run a 3–5 question diagnostic spanning easy to hard before teaching.
3. Render an ASCII skill map with levels 0–5 per subskill and an overall percentage bar.
4. Enter the teaching loop: one question per message, wait, no trailing hints.

Levels are defined plainly: 0 unaware through 5 can teach it / handle edge cases.

Strict evaluation rules

Jane Street's copy is unusually explicit about anti-sycophancy:

• No level bumps without two independent checks, one unprompted recall.
• Vague answers count as wrong — probe until understanding is demonstrated.
• Spaced review can lower levels if the user forgets earlier material.
• Level 5 requires an edge-case or adversarial question and teach-back quality.
• Do not flatter — prefer "Not quite — why would X happen if Y were true?"

User commands

The skill documents escape hatches:

• show progress — redraw the map
• harder / easier
• just tell me — one-question lecture override
• save state / resume from: — portable session blobs

That last pair is valuable for long-running mastery paths across sessions — a pattern more skills should copy for education use cases.

How to try it today

With a SKILL.md-capable agent (Claude Code, Cursor agents, other spec hosts):

1. Clone https://github.com/janestreet/agent_skills.
2. Copy socratic-tutor/ into your host's skills directory per that product's docs.
3. Prompt: "Use socratic-tutor — teach me how KV cache eviction works."

With a plain chatbot:

1. Paste the body of SKILL.md after the frontmatter into a custom instruction field.
2. Start with an explicit trigger: "Tutor me on backpropagation."

Compare results with AWS Quick's approach to progressive disclosure in skills publishing: Jane Street loads everything when triggered; Quick loads metadata first. Host behavior still matters.

How this compares to other team skills repos

Star velocity and maintenance signals

The candidate feed logged 2,124 stars with repo created November 3, 2026 — fast uptake for a single-skill repo. Commit history showed two commits on November 10 fetch — early but not abandoned.

Watch for:

• Additional skills directories beyond socratic-tutor
• CONTRIBUTING.md norms for external PRs
• Whether Jane Street pins model-specific addenda

Treat star count as interest in agent skills as a format, not proof the pedagogy beats human tutors.

Skill map example (what the agent should render)

Jane Street specifies an ASCII template the host should print after diagnostics. The pattern looks like this in SKILL.md:

Topic: overall: ######.... 58%
-----------------------------------------------------------------
subskill A [########..] 4/5 * independent
subskill B [######....] 3/5 <- current focus
subskill C [##........] 1/5
-----------------------------------------------------------------

The file asks for solid block U+2588 and light block U+2591 characters when the environment supports them. That visual discipline matters for long-running tutoring sessions where users lose track of progress. It also makes save state blocks easier to parse when pasted into a new chat.

Teaching loop rules in plain language

After the map appears, the agent should:

• Ask one question, then stop the message — no trailing hints, no example answers in the same turn.
• Prefer questions that force prediction, error finding, or transfer to a new case.
• On wrong answers, narrow the question twice before explaining.
• On right answers, occasionally devil's advocate to test robustness.
• Run spaced review on earlier subskills without announcing it.

The skill explicitly forbids attributing insights the user did not state. That line fights hallucinated progress reports, a common failure mode when models want to be encouraging.

Final assessment gate

Completion requires every subskill at the target level (default 4/5, or 5/5 if the user asked for full mastery). Then the agent runs a mixed assessment with at least one question per subskill plus one integrative problem. Failures send the user back into the loop rather than issuing a participation trophy.

Porting socratic-tutor to your agent harness

If you maintain skills in git, add a CI check that SKILL.md frontmatter name matches the directory name — the same lint AWS Quick recommends for production skills in its November guide.

Jane Street and the wider skills arms race

November 2026 is crowded with skills-shaped releases:

• AWS Quick productized publish/share for business users.
• Jane Street open-sourced pedagogy.
• Trending repos such as Stack Diagram target diagram generation.

Jane Street's contribution is normative text: it tells the model how not to teach. That is harder to copy than a script because it depends on turn-by-turn discipline from the host runtime.

For education startups, the skill is a free baseline to compare against proprietary tutors. Run A/B tests: default system prompt versus socratic-tutor on the same calculus or ML syllabus and measure human-rated mastery, not token count alone.

OCaml culture and .ocamlformat in the repo

The repository root includes .ocamlformat, a wink at Jane Street's OCaml ecosystem, even though the skill itself is Markdown. Contributors editing future skills should keep CONTRIBUTE.md norms in mind if the firm accepts external skills — today the repo is tiny, but the file signals reviewed contributions rather than drive-by prompts.

Limitations and risks

• Token burn: Socratic loops are long; pair with output-token awareness when using reasoning models.
• No sandbox: The skill assumes the host enforces tool policies; it does not ship execution sandboxing — irrelevant for pure Q&A tutoring but critical if you add code-exec exercises.
• Not financial advice: Jane Street's brand does not make this a trading signal.

When not to use this skill

Skip socratic-tutor for:

• Incident response or on-call pages where the user needs an answer in one shot.
• Compliance answers that must quote regulation verbatim — use retrieval with citations instead.
• Creative brainstorming where leading questions constrain ideation.
• Children's homework without oversight — the skill is strict, not a substitute for teacher supervision.

For security training, pair tutoring with factual runbooks and restrict tool access in the host so a student agent cannot exfiltrate data while "learning."

Related reading

Jane Street agent skills library card pulled from shelf illustrating SKILL.md collections

• What are agent skills — complete guide
• AWS Quick skills and agents guide
• Amazon Quick agentic primitives
• Executor 2 MCP gateway
• Stack Diagram agent skill trending repo
• Medium TK AI notebook with MCP and skills

Star counts and commit totals are accurate as of November 10, 2026 per GitHub.