Use Claude Code as an agentic tutor in your learning environment
How do you use your LLM coding agent? Task: The author used Claude Code as an agentic tutor to learn dbt by migrating an existing DuckDB SQL pipeline. They configured a CLAUDE.md file to enforce a teaching role, requiring the AI to explain concepts and guide the user rather than just generating code. The human actively intervened to correct the AI when it attempted to 'just do it' instead of instructing, and manually wrote code based on AI-provided scaffolds.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include a working knowledge of SQL and the specific domain data (UK Environment Agency flood monitoring). A key technical constraint is DuckDB's single-writer limitation, which caused lock conflicts during parallel execution. The AI's tendency to bypass the 'tutor' instruction requires constant human vigilance and prompt reinforcement to prevent it from simply completing the task without teaching..
Tools
Published on Agent Work Atlas:
Overview
How do you use your LLM coding agent?
Workflow
The author used Claude Code as an agentic tutor to learn dbt by migrating an existing DuckDB SQL pipeline. They configured a CLAUDE.md file to enforce a teaching role, requiring the AI to explain concepts and guide the user rather than just generating code. The human actively intervened to correct the AI when it attempted to 'just do it' instead of instructing, and manually wrote code based on AI-provided scaffolds.
Reported result
The author reported the experience was 'great' and successfully learned dbt concepts through interactive tutoring. However, the AI initially struggled to maintain the tutor persona, frequently attempting to write code directly, which required repeated human correction. The workflow also revealed AI hallucinations, such as inventing API field values, which the human caught and forced the AI to verify against actual data.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include a working knowledge of SQL and the specific domain data (UK Environment Agency flood monitoring). A key technical constraint is DuckDB's single-writer limitation, which caused lock conflicts during parallel execution. The AI's tendency to bypass the 'tutor' instruction requires constant human vigilance and prompt reinforcement to prevent it from simply completing the task without teaching.
Try it
Act as my agentic tutor for [replaceable: target technology, e.g., dbt] using [replaceable: existing project context, e.g., SQL scripts]. First, create a CLAUDE.md file that explicitly defines your role: you must explain concepts and guide me, but I will write the code. Do not generate full solutions unless I am completely stuck and ask for scaffolding. When I provide input, check if it aligns with best practices; if I use shortcuts like SELECT *, correct me. If you encounter errors, diagnose them but teach me the troubleshooting steps rather than fixing them silently. Start by reviewing my current [replaceable: source files] and proposing a learning path. Wait for my confirmation before proceeding to the first module. If you start writing code instead of teaching, I will interrupt you. Ensure you verify any factual claims about data schemas against the actual files before stating them as fact.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.