MCP Refactoring Skill

Guidelines for refactoring Raindrop MCP tools for better LLM integration, usability, and following MCP best practices.

Capabilities: Resources, Sampling, Elicitation

Resources

Expose all major Raindrop entities as MCP resources:

• collections://all, collections://{id}
• bookmarks://{id}, tags://all
• highlights://all, user://info

Implement resource discovery and navigation (list, get, search, children, etc.) using standard MCP resource URIs and methods. Ensure each resource supports GET (read), and where appropriate, CREATE, UPDATE, DELETE (write) actions.

Sampling

For large collections/bookmarks/tags/highlights, implement sampling endpoints:

• e.g., bookmarks://collection/{id}?sample=10 returns a random or recent sample.
• Add tool parameters for limit, offset, and sample to all list/search tools.
• Use MCP's sampling capability to advertise this in the server manifest.

Elicitation

Implement elicitation tools for:

• Confirming destructive actions (delete, merge, etc.)
• Requesting missing parameters (e.g., if a required field is omitted, prompt the LLM/user)
• Use the MCP elicitation capability to allow the server to ask clarifying questions or confirmations.

Streamlining Tools for LLMs

Hierarchical, Predictable Naming

• Use a consistent {resource}_{action} pattern (e.g., collection_list, bookmark_create, tag_manage)
• Group related actions under a single tool with an operation parameter where possible (e.g., collection_manage for create, update, delete)

Reduce Redundancy

Collapse similar tools:

• Merge collection_create, collection_update, collection_delete into collection_manage with an operation parameter
• Do the same for bookmarks, tags, highlights
• For read-only actions, keep list, get, and search as separate, simple tools

LLM-Friendly Descriptions

• Ensure every tool and parameter has a clear, concise description
• Use Zod schemas for validation and documentation

Example: Refactored Tool Set

LLM/AI-Optimized Features

• Resource URIs: Support direct resource access via URIs (e.g., collections://all)
• Streaming: For large lists, support streaming or pagination
• Sampling: Add sample and limit parameters to all list/search tools
• Elicitation: Use MCP's elicitation to prompt for missing/ambiguous info and confirmations
• Consistent Error Handling: Always return structured, descriptive errors