The ASCII skill map visualizes progress across subskills, using solid and light block characters.
JANE STREET RELEASES SOCRATIC TUTOR SKILL FOR AI AGENTS
*Jane Street adds a structured Socratic tutoring module to its open‑source agent_skills repo. The tool forces AI to teach by questioning, not lecturing, and embeds strict mastery checks.*
Jane Street has pushed a new open‑source tutoring module to its agent_skills collection on GitHub. The Socratic Tutor is designed for ongoing teaching sessions, not for answering isolated questions. It activates only on explicit requests like “teach me” or “quiz me” and begins by confirming the user’s intent. Once confirmed, the skill splits the target topic into a handful of subskills, runs a short diagnostic, and displays an ASCII progress bar that tracks mastery from unaware (0) to teachable (5). The design forces the AI to ask probing questions, demand reasoning, and only award levels after repeated, independent proof of understanding. The module’s strict guard clauses and level‑based evaluation aim to turn language models into disciplined, self‑correcting tutors.
What is the Socratic Tutor skill?
The Socratic Tutor is a structured tutoring module defined in Jane Street’s public GitHub repository under agent_skills/socratic-tutor/SKILL.md. It activates only when a user explicitly asks to be taught, quizzed, or drilled, using trigger phrases such as “teach me”, “tutor me”, “quiz me”, “drill me”, or “help me master”. It is not intended for one‑off “explain X” queries. Before launching a session, the skill confirms intent with a single short question, avoiding unwanted quizzes. The session begins by breaking the topic into 4‑10 concrete subskills, running a 3‑5 question diagnostic, and presenting an ASCII skill map that tracks each subskill on a 0‑5 scale.
How does the tutoring loop enforce mastery?
The loop asks one question at a time, demanding prediction, explanation, or flaw detection. Wrong answers trigger narrower follow‑ups; correct answers may be challenged with devil’s‑advocate prompts. Levels advance only after the user demonstrates performance on two independent checks, one of which must be unprompted recall. Levels are earned, not given, and a vague answer does not count. The system interleaves spaced review, lowering a level if a review fails. Mastery is declared when every subskill reaches at least level 4, or level 5 if full mastery was requested, followed by a final mixed assessment covering all subskills.
What safeguards prevent misuse?
A guard clause forces the model to confirm teaching intent before any diagnostic begins. If intent is ambiguous, the skill asks a single clarifying question: “Want a quick answer, or should I actually teach you this properly?” This confirmation costs only one word, while a false start would waste a quiz. The skill never attributes insight the user has not expressed and never flatters; feedback stays blunt (“Not quite — why would X happen if Y were true?”). These measures keep the tutoring loop focused and prevent accidental activation during casual queries.
What commands give users control?
Users can type “show progress” to render the current ASCII skill map at any time. The commands “harder” or “easier” adjust difficulty on the fly. “just tell me” overrides Socratic mode for the current question only. “save state” emits a compact block containing the topic, skill map, and notes, enabling session resumption with “resume from:”. The skill also supports “adjust difficulty” and “save state” but will refuse actions that are unavailable in the current context.
Frequently asked questions
How does the Socratic Tutor decide when to start a tutoring session?
It waits for explicit trigger phrases like “teach me” or “quiz me” and then asks a short confirmation question before launching the diagnostic.
What level must a subskill reach for the session to be considered complete?
Every subskill must reach at least level 4, or level 5 if the user requested full mastery, before the final assessment is run.
Can users see their progress during a tutoring session?
Yes, typing “show progress” renders the current ASCII skill map with level bars for each subskill.
By codifying a maieutic teaching method into a public repository, Jane Street offers a reproducible blueprint for AI tutoring that can be audited, extended, or integrated into larger systems. The blunt feedback loop and hard‑earned levels set a high bar for genuine mastery, challenging developers to move beyond superficial explanations. As more teams adopt the skill, the industry may see a shift toward AI that teaches by questioning, not by dumping information.
Sources: [GitHub – Jane Street agent_skills Socratic Tutor SKILL.md]