Recoleta
GitHub
Source note

Agentic Education: Using Claude Code to Teach Claude Code

Authors
Zain Naboulsi
Source
arXiv
Published
Collection
Software Intelligence
Agentic CodingDeveloper EducationClaude CodeAdaptive LearningMulti Agent Workflows
Open arXivMarkdown source

Summary

This paper presents cc-self-train, a 50-module curriculum that uses Claude Code itself to teach developers how to use Claude Code. The main claim is that agentic coding tools need structured, auto-updating instruction, and the pilot study reports large self-efficacy gains across all measured skill areas.

Problem

  • Developers can access powerful agentic coding tools, but current learning materials are fragmented, feature-by-feature, and often stale within days because the tools change quickly.
  • Existing docs, blog posts, and courses do not provide a progressive path from beginner use to advanced features such as hooks, skills, subagents, and multi-agent workflows.
  • This matters because advanced agentic features require compositional understanding; without a structured path, adoption and effective use stay limited.

Approach

  • The system is a hands-on curriculum called cc-self-train with 50 modules organized as 10 sequential modules across 5 project paths: Canvas, Forge, Nexus, Sentinel, and BYOP.
  • It teaches Claude Code by having learners build real software projects while interacting with Claude Code as the instructor inside the same environment.
  • Instruction changes across 4 personas aligned to Gradual Release of Responsibility: Guide → Collaborator → Peer → Launcher. Early modules explain more; later modules step back and expect more learner control.
  • An adaptive layer watches learner engagement through hook-based heuristics, uses streak detection for mid-module intervention, and changes persona schedules at module boundaries.
  • The system also includes step pacing, cross-session state tracking, a parameterized test suite for structural consistency across all modules, and an onboarding agent that checks for upstream Claude Code changes and updates teaching materials before instruction starts.

Results

  • The paper reports a pilot evaluation with 27 participants.
  • It claims statistically significant reported self-efficacy gains across all 10 assessed skill areas, with p < 0.001.
  • The largest reported gains were on more advanced Claude Code features, including hooks and custom skills.
  • The curriculum contains 50 module files, 5 project domains, 10 progressive modules per path, 22 context documents, and 8 test files used for consistency checks.
  • The excerpt does not provide task-performance metrics, completion rates, time savings, or comparisons against a baseline curriculum or alternative tool-training method.

Link