Skills with MCP, Tools, and Subagents

These pieces are complementary:

• Tools: atomic actions (read file, run command, call API)
• MCP: external system/data integration layer
• Skills: repeatable workflow procedure
• Subagents: isolated delegated workers

Working Mental Model

• Tools are verbs.
• MCP is the connector fabric.
• Skills are playbooks.
• Subagents are workers.

End-to-End Pattern

1. MCP exposes external data/tools.
2. Tools execute concrete operations.
3. Skills define step order and decision criteria.
4. Subagents run specialized tasks in isolation or parallel.
5. Parent agent consolidates final output.

Example: Data to Branded Presentation

1. Pull metrics through an MCP-connected datastore.
2. Run analysis skill for summarization/scoring.
3. Run brand-guidelines skill for style constraints.
4. Run presentation/document skill for final artifact.

This pattern keeps capability access modular and output behavior consistent.

Context-Window Discipline

• Keep tool definitions lean.
• Load only skill metadata initially.
• Load full skill files and references only when triggered.
• Use subagents to prevent main-thread context bloat.