Extending Claude’s capabilities with skills and MCP servers
December 19, 2025 at 12:00 AM
Summary
TL;DR
MCP connects Claude to external systems; Skills teach Claude how to use those connections for your team’s workflows. The combo targets predictable orchestration and consistent outputs.
What actually happened
Anthropic explains how Skills and MCP servers complement each other in Claude-based agents.
MCP is positioned as secure, standardized connectivity to tools and data sources.
Skills are positioned as reusable procedural knowledge: sequencing, standards, and edge-case handling.
The post gives two worked examples: financial valuations using market-data providers and meeting prep in Notion.
It provides a decision guide for when to build a Skill vs. when to add an MCP server.
Key numbers
Reading time: 5 min
Date: December 19, 2025
Why this was hard
Raw tool access doesn’t encode “where to look” or “what done looks like,” so the model guesses.
Multi-step workflows can run in the wrong order without explicit sequencing.
Teams need outputs that match internal standards and compliance checkpoints, not generic formatting.
Instructions can conflict between tool-side hints (MCP) and workflow-side requirements (Skills).
How they solved it
Separate responsibilities: MCP for connectivity; Skills for workflow logic and domain expertise.
Keep MCP instructions generic (how to use the tool correctly), push process/presentation into Skills.
Use Skills to encode discovery rules (which sources to query first, then next).
Use Skills to define orchestration steps so the same workflow runs the same way each time.
Use Skills to enforce performance criteria like formatting standards and compliance validation.
Compose freely: one Skill can coordinate multiple MCP servers; one MCP server can back many Skills.
Why this matters beyond this company
Treat “integration” and “procedure” as different artifacts to keep agent systems composable.
Put tool-specific usage details near the tool, but keep workflows reusable across tools.
Explicitly resolve instruction conflicts (e.g., JSON vs. markdown) by assigning ownership per layer.
Stealable ideas
Write Skills as playbooks: discovery order, step sequencing, output spec, and edge-case rules.
Keep MCP server guidance generic; avoid embedding team-specific formatting in server prompts.
Design for reuse: Skills that orchestrate multiple MCP servers, and multiple Skills per server.
Add a validation step in the Skill when outputs must meet compliance or internal standards.