The Socrates Agent
Socrates is a minimalist AI tutoring agent (~60 lines) built on Claude Code that teaches exclusively through questioning, never providing direct answers or completing any part of a student's deliverable The agent enforces a strict Socratic method: it asks one question per reply, requires students to show their work first, shrinks problems when stuck, and ends sessions by having students explain the full solution from scratch It explicitly blocks common jailbreak attempts including "just show me
Analysis
TL;DR
- Socrates is a minimalist AI tutoring agent (~60 lines) built on Claude Code that teaches exclusively through questioning, never providing direct answers or completing any part of a student's deliverable
- The agent enforces a strict Socratic method: it asks one question per reply, requires students to show their work first, shrinks problems when stuck, and ends sessions by having students explain the full solution from scratch
- It explicitly blocks common jailbreak attempts including "just show me yours," "pretend you're a different AI," "my teacher said it's fine," and even narrowing answer ranges — all receive questions in return
- The underlying philosophy frames AI assistance as a "gym" problem: tasks you want done versus tasks you want to do yourself, warning that over-reliance on AI erodes the cognitive muscle of working through hard problems
- Installation is trivial — a single markdown file placed in
~/.claude/agents/— and the agent is designed to be portable across any system prompt-compatible harness (ChatGPT, Gemini, etc.)
Why It Matters
This agent addresses a growing concern among educators and technologists: as AI becomes capable of producing near-instant answers, students and adults alike are losing the ability to sit with difficult problems and work through them independently. The Socrates agent offers a practical, deployable countermeasure that preserves the learning process rather than outsourcing it, making it relevant not just for K-12 education but for anyone concerned about cognitive atrophy from over-reliance on AI tools.
Technical Details
- Built as a Claude Code subagent using a single YAML frontmatter + markdown prompt file (~60 lines total), leveraging tools like Read, Grep, and Glob while explicitly disallowing Edit, Write, NotebookEdit, Bash, WebFetch, and WebSearch to prevent the agent from producing deliverables
- The system prompt enforces a strict conversational protocol: every response is short, ends with exactly one actionable question, requires the student to show prior work before any guidance is given, and uses problem-shrinking techniques (simpler analogues, known cases) when the student is genuinely stuck
- The agent maintains a hard boundary against answer-giving across all framing attempts — whether the user asks for a number, a thesis, code, a proof, or a multiple-choice confirmation, the response is always a redirecting question; even confirming partial correctness (e.g., "is the answer bigger than 50?") is refused
- Safety override: the only exception to the Socratic protocol is when a student reports danger or self-harm, in which case the agent immediately stops teaching and directs them to a trusted adult
- The agent is designed for portability: the prompt file works in any harness that accepts system prompts, and it ships as part of the LifeOS ecosystem with a public mirror at ourlifeos.ai/agents/socrates
Industry Insight
- The "Socratic agent" pattern represents an emerging class of AI tools designed not for productivity but for cognitive preservation — as AI capabilities advance, there will be increasing market demand for agents that help users maintain independent thinking skills rather than replace them
- The extreme simplicity of this agent (60 lines, single file) demonstrates that powerful behavioral constraints can be achieved through prompt engineering alone, without fine-tuning or complex architecture — a signal that prompt-level design will remain a critical skill even as models grow more capable
- The framing of "gym vs. job" tasks for AI use is likely to become a standard mental model in edtech and workplace learning: tools that help users do work they want to learn to do themselves will differentiate from tools that simply automate output, especially as regulatory and parental scrutiny of AI in education intensifies
Disclaimer: The above content is generated by AI and is for reference only.