Jane Street Agent Skills: the socratic-tutor SKILL.md
Jane Street's agent_skills repo (MIT) ships one skill: socratic-tutor, a Socratic tutor you copy into Claude Code, dsh or a plain chatbot. What it does.
Last updated: 2026-11-15
Quick answer
In November 2026 the GitHub organization Jane Street published janestreet/agent_skills under the MIT license. Its README is one sentence of scope — "a collection of skills for AI agents and chatbots. Each skill is a directory containing a SKILL.md file with instructions for the agent" — and the repository currently holds a single skill: socratic-tutor, which turns a model into a strict one-question-at-a-time tutor with a 0–5 skill map.
That is the whole thing. There is no CLI, no installer, no plugin manifest and no model of its own. If you came here expecting a Jane Street trading stack, or a dsh plugin written by Jane Street, this is not it — see the section on the name mix-up below.
Everything on this page was checked against the repository and its files on 2026-11-15; quotes are verbatim and the star count is a snapshot from that day, not a fixed number.
What is actually in the repository
The full tree, read from the GitHub API on 2026-11-15:
| Path | What it is |
|---|---|
README.md | Scope, the single skill, and how to use it |
socratic-tutor/SKILL.md | The skill itself — instructions for the agent |
CONTRIBUTING.md | Pull requests plus a Developer Certificate of Origin sign-off requirement |
LICENSE.md | MIT, "Copyright (c) 2026 Jane Street Group, LLC" |
.ocamlformat | OCaml formatting config, a house habit of the org |
.gitignore | — |
Two facts settle the "is this official" question. The owner is the janestreet Organization (446 public repos, blog opensource.janestreet.com), and CONTRIBUTING.md opens with "This repository contains open source software that is developed and maintained by Jane Street". The license file names Jane Street Group, LLC.
The absence of a build is also a fact worth keeping: .ocamlformat and no OCaml source means the repo is not a library. It is documentation-shaped, which is exactly what an agent skill is.
What the socratic-tutor skill does
The frontmatter of socratic-tutor/SKILL.md is the clearest summary, and it is also the gate:
"Structured Socratic tutoring when the user explicitly asks to be taught, quizzed, or drilled to mastery on a topic: diagnostic placement, a skill map with levels 0-5, spaced review, and a final assessment. Trigger phrases: 'teach me', 'tutor me', 'quiz me', 'drill me', 'help me master'. Not for one-off 'explain X' or 'what is X' questions."
That last sentence matters in practice. The skill starts with a guard that refuses to run for a single factual question, and if intent is ambiguous it asks one short question — the file's own reasoning is "a false start costs the user an unwanted quiz; a confirmation costs one word."
Once a session starts, the loop is prescribed in detail:
- Break the topic into 4–10 subskills (two-level tree for broad topics) and say the tutor will ask a few questions first.
- Diagnostic placement: 3–5 probing questions, easy to hard, with no teaching until the diagnostic is done.
- Show the skill map and propose a starting point, letting the user override.
- Teach by questioning, one question per message — "End the message immediately after the question and wait for the answer".
- Final mixed assessment once every subskill reaches the target level (default 4/5, or 5/5 if the user asked for full mastery).
The skill map is a five-level scale, rendered as ASCII bars after every level change:
0 unaware . 1 heard of it . 2 can follow along . 3 can do with help
4 can do independently . 5 can teach it / handle edge cases
The evaluation rules are the part that separates this from a "be a tutor" prompt. Levels are "earned, not given" — a level goes up only after 2+ independent checks, one of them unprompted recall. Vague answers do not count. Spaced reviews are interleaved without warning, and failing one lowers the level. Level 5 requires handling an adversarial question and explaining the idea well enough to teach it.
Install it in your agent
The README's instruction, in full:
"With a coding agent that supports the common
SKILL.mdformat, copy the skill directory into your agent's skills folder, then ask the agent to use it ('teach me X'). With a plain chatbot, paste the body of the skill'sSKILL.mdinto the chat (or into a custom/system prompt), followed by what you want it to do."
Three routes, from most automatic to most manual:
Claude Code and other SKILL.md agents. Either put the folder where that agent reads skills, or paste the file body into a custom prompt. Skills repositories rarely get simpler than this — there is nothing to build.
DeepSeek Harness. dsh reads project skills from .dsh/skills/<name>/, so one copy per workspace is enough:
git clone --depth 1 https://github.com/janestreet/agent_skills /tmp/jane-street-agent-skills
mkdir -p .dsh/skills/socratic-tutor
cp /tmp/jane-street-agent-skills/socratic-tutor/SKILL.md .dsh/skills/socratic-tutor/SKILL.md
Then say teach me <topic> in your next message. The install-skills walkthrough covers where the folder lives, and what dsh plugins are explains why a skill needs no plugin entry at all: plugins and skills are two different things.
A chat window with no agent loop. Paste the SKILL.md body as a system prompt, then state the topic. You lose the ability to have the model read files or run exercises, and every session restarts from the diagnostic unless you use the skill's own "save state" command.
The user commands worth knowing
socratic-tutor defines six phrases the learner can say at any time — the escape hatches are what make the strictness tolerable:
| Say | What happens |
|---|---|
| "show progress" | Renders the skill map again |
| "harder" / "easier" | Adjusts difficulty for the next question |
| "just tell me" | Overrides Socratic mode for the current question only |
| "save state" | Emits a compact block (topic, map, misconceptions) to reuse later |
| "resume from:" | Restores a pasted state block and continues |
"Save state" is the only persistence the skill has. There is no database, no file output, no memory across sessions except that block — which is why it is worth using it before you close the window.
Where it stops
- One skill, one job. No coding, no research, no tooling beyond exercises it invents. Do not expect a Jane Street-flavoured agent stack.
- No install machinery. No package, no plugin manifest, no update path. Keeping current means pulling the file again.
- The quality ceiling is the model, not the skill. The file tells the agent how to be strict; whether the questions are good depends on the model you plug in.
- Markdown-only means prompt-injection surface is low but not zero. You are still trusting a ~200-line instruction file with your conversation; read it first, it is short.
The "Jane Street" name on dsh quant plugins
This is the mix-up worth clearing up, because the search results blend the two. Nothing in the DeepSeek Harness plugin catalog is published by Jane Street. What exists is third-party work that uses the firm's name as a style reference:
| Repository | What its own description says | Affiliation |
|---|---|---|
pengpengyi92/dsh-quant | "Dsh-Quant: The Everything-Plugin Ai native Quant OS"; its README describes a "Jane Street-style UI demo" | Third-party, not official |
pengpengyi92/dsh-quant-ui | "Jane Street-style quant workbench for dsh-quant — K-lines, equity curves, fund simulation cards, metric selector" | Third-party, not official |
janestreet/agent_skills | "A collection of skills for AI agents and chatbots" — one skill, socratic-tutor | Official Jane Street repo, MIT |
A "Jane Street-style" chart layout is a design compliment, not a distribution. The quant plugins are community packages for trading research; the official repository is a single tutoring skill. They share three words and nothing else — no code, no authors, no endorsement.
FAQ
What are the Jane Street agent skills?
An open-source repository, janestreet/agent_skills, published by the Jane Street GitHub organization under the MIT license on 2026-11-03. Its README calls it "a collection of skills for AI agents and chatbots", where each skill is a directory holding a SKILL.md file with the instructions for the agent. At the time of writing it contains exactly one skill: socratic-tutor.
Do I need a Claude Code plugin or a package manager to install it?
No. The README describes a copy: put the skill directory into your agent's skills folder, then ask the agent to use it ("teach me X"). For a plain chatbot, paste the body of the SKILL.md into the chat or a custom/system prompt. There is no installer script, no npm package and no marketplace entry in the repo.
Does it work in DeepSeek Harness (dsh)?
Yes, by copying. dsh reads project skills from .dsh/skills/<name>/ inside the workspace, so a copy of the socratic-tutor folder at .dsh/skills/socratic-tutor/SKILL.md gives you the same behaviour. The skill is a single Markdown file with no runtime dependencies, which is why the copy route works across agents.
Is the Jane Street quant plugin for dsh made by Jane Street?
No. Repositories such as pengpengyi92/dsh-quant and pengpengyi92/dsh-quant-ui only borrow the name for a UI style — their descriptions say "Jane Street-style". They are third-party DeepSeek Harness plugins and are not affiliated with Jane Street. The only official repository in this article is github.com/janestreet/agent_skills.
Sources
Read on 2026-11-15:
- github.com/janestreet/agent_skills — README (scope sentence, the skill list, the copy-based install instructions),
socratic-tutor/SKILL.md(frontmatter trigger phrases and the "not for one-off questions" exclusion, the 4–10 subskill rule, the 3–5 question diagnostic, the 0–5 level scale, the evaluation rules, the six user commands),CONTRIBUTING.md(Jane Street maintains it; DCO sign-off),LICENSE.md(MIT, Jane Street Group LLC). - GitHub API, snapshot taken 2026-11-15: owner is the
janestreetOrganization, MIT license, created 2026-11-03, last push 2026-11-03, default branchmaster, about 8,700 stars (a snapshot, not a live figure), full file tree as listed above. - github.com/pengpengyi92/dsh-quant and pengpengyi92/dsh-quant-ui — the "Jane Street-style" wording, taken from the repository description and README. Both are third-party projects with no stated affiliation; nothing distinguishes them as endorsed by Jane Street and this page does not suggest otherwise.
- github.com/janestreet — the organization profile: "Jane Street", 446 public repositories,
opensource.janestreet.com.
Related reading: the best DeepSeek Harness skills, installing skills, plugins vs skills, humanizer skills.
