Build server-side reference implementation for skills-as-resources #41

Background

The Feb 26 office hours focused on building reference implementations for skills-as-resources. This issue covers the server side — a canonical example showing how to expose skills using existing MCP primitives (resources, tools, resource templates) without protocol changes.

This is distinct from #24 (NimbleBrain-specific approach) — this issue is about a minimal, framework-agnostic reference implementation that demonstrates the convention.

Scope

Build a reference MCP server that:

1. Exposes skills as resources using a consistent URI scheme (e.g., skill://server-name/skill-name)
2. Supports resources/list returning skill metadata (name, description, version)
3. Supports resources/read returning full skill content (SKILL.md body)
4. Optionally supports resources/subscribe and notifications/resources/updated for push-based updates
5. Optionally exposes a list_skills tool as a model-controlled alternative
6. Includes at least 2 example skills (one simple, one complex multi-step workflow)
7. Uses server instructions to guide the model toward skill discovery

Acceptance Criteria

• Working MCP server (Python or TypeScript preferred for broad accessibility)
• At least 2 example skills with different complexity levels
• Demonstrates both application-controlled (resources) and model-controlled (tool) access
• Includes a test demonstrating skill discovery and loading from a client
• Added to examples/ folder in this repo (per Add examples to common shared skills folder for testing implementation(s) #27)
• Documented in docs/experimental-findings.md

References

• Feb 26 office hours notes
• Approach 3 in docs/approaches.md
• Related to Add reference implementation for NimbleBrain skills-as-resources approach #24 and Add examples to common shared skills folder for testing implementation(s) #27