shareAI-lab/learn-claude-code

• **Archetype**: harness-framework (educational harness architecture)
• **Stars**: 52,165
• **Last pushed**: 2026-04-07
• **License**: MIT
• **Discovered**: 2026-04-12
• **Source**: gh-search (topic: claude-code)
• **Skills found**: 4 (pdf, code-review, mcp-builder, agent-builder)

Summary

A "nano claude code-like agent harness" built from scratch in Python/Bash for educational purposes. Includes skills for PDF processing, code review, MCP server building, and agent building. The agent-builder skill contains a comprehensive agent design methodology: core philosophy ("the model already knows how to be an agent"), three elements (capabilities, knowledge, context), and design thinking framework.

Assessment

HIGH VALUE. The agent-builder skill is a standout -- it contains a well-articulated agent design methodology that emphasizes simplicity and the model's inherent capability. The code-review skill and MCP-builder skill contain procedural workflows. The educational framing provides clear, extractable process definitions.

Extraction Priority

HIGH -- The agent-builder methodology is directly extractable:

• Agent design thinking -> methodologies/ (agent architecture methodology)
• Code review skill -> already have code-review but may contain unique patterns
• MCP-builder -> specializations/shared/ (cross-domain MCP development process)

Processes

1. **agent-design-thinking** -- Purpose -> Domain -> Capabilities (3-5 essential) -> Knowledge -> Context analysis 2. **agent-build-loop** -- The core agent loop: context + capabilities -> decide -> act or respond 3. **mcp-server-development** -- Process for building MCP servers from scratch 4. **code-review-workflow** -- Structured code review process

Plugin Ideas

None specific -- methodology content, not tool integration.

Harness Integration Ideas

Educational Architecture Patterns

• **Current limitation**: Limited documentation of harness architecture principles
• **Integration approach**: Use educational patterns to improve babysitter harness documentation and architecture understanding
• **Implementation scope**: Internal documentation, architecture refinement, team education

Agent Loop Optimization

• **Current limitation**: Could benefit from clearer agent loop architecture documentation
• **Integration approach**: Apply educational agent loop patterns to babysitter runtime documentation
• **Implementation scope**: packages/babysitter-sdk/src/runtime/, improved loop architecture

Permission System Enhancement

• **Current limitation**: Limited safety validation patterns for model intent execution
• **Integration approach**: Study permission layer patterns for agent execution safety
• **Implementation scope**: packages/babysitter-sdk/src/security/, model intent validation

Implicit Procedural Knowledge

• "Start with 3-5 capabilities, add more only when agent consistently fails" -- capability minimalism principle
• "Make knowledge available, not mandatory. Load when relevant, not upfront" -- lazy knowledge loading
• "Context is precious. Isolate noisy subtasks. Truncate verbose outputs. Protect clarity" -- context hygiene
• The agent loop distilled to its simplest form
• "The model does the reasoning. The harness gives the model a working environment" -- fundamental harness philosophy

Library Mapping

Plugin Marketplace Mapping