Skills & Agents — On-demand context

🌐 日本語

NOTE

Skills and Agents are loaded only when the LLM (or the user) explicitly invokes them. They are the right home for instructions that are large, specialized, or only occasionally relevant.

How it works

• Skills: declarative capability packets the LLM can call by name (skill: "pdf").
• Agents: separate execution contexts that run a focused task and return a result.
• Both keep their bodies out of the main context until activation.

Related chapters

• Part 5 — On-Demand Context (overview)
• Skills Design
• Agents Design
• Skills vs Agents — decision criteria
• Configuration Reference

Structural problems it addresses

• Context Rot — specialized knowledge stays out of context until needed.
• Knowledge Boundary — Skills inject up-to-date procedural knowledge.
• Hallucination — Agents isolate risky reasoning into a fresh context.

When a single Agent isn't enough

For a one-shot delegation, an Agent (sub-agent) is the right primitive. When the work needs to survive across many sessions, let children talk to each other, or avoid filling the parent's context with summaries, the next step is Agent Teams — peer-to-peer, persistent sessions covered in Part 10.

See also

• Topic: CLAUDE.md — for always-loaded context.
• Topic: Rules — for context triggered by file globs.
• Topic: MCP — for tool-based external context.
• Topic: Agent Teams — multi-session coordination when one Subagent isn't enough.
• Feature Index