Skills
A Skill is reusable knowledge or a workflow you teach the Employee once and reuse across runs. A Skill might be a few lines describing a formatting preference, a multi-page methodology document, or a callable workflow you trigger by typing /skill-name. Skills can carry reference files (playbooks, compliance frameworks, pricing matrices) that the Agent loads when relevant. They accept inputs, so /close-package North America and /close-package Europe run the same workflow with different scope. Use Skills when you have institutional knowledge or a structured process you want consistently applied.
Sub-agents
A sub-agent is a fresh execution context the Employee can dispatch to handle a specific piece of work in parallel or in isolation. Each sub-agent gets its own context window so it doesn't pollute the parent run. Use sub-agents when you have independent work items (research three companies in parallel, scan five repos, summarize four documents) or when you need a clean scratch space for a focused side-task. Skills tell the Employee what to do; sub-agents are the where the work runs.
MCP tools
MCP tools are external integrations exposed through the Model Context Protocol — connectors to databases, SaaS apps, internal APIs, and your own services. Where Skills are knowledge and sub-agents are execution capacity, MCP tools are how the Employee reaches the outside world. Wire an MCP tool when the Agent needs to read or write to a system that isn't already on disk.
Hooks
Hooks are deterministic callbacks that fire on lifecycle events — before a tool runs, after a file change, on session start. Use hooks for guardrails (block destructive commands, require approval before pushing to main) and observability (log every dispatched run to your audit system). Skills shape behavior; hooks enforce policy.
Slash commands
Slash commands are short shortcuts that expand into full prompts at the moment you invoke them. They're the lightest primitive — perfect for "run this exact phrase that I keep typing" without packaging it as a full Skill.
Picking the right primitive
- Reusable knowledge or methodology → Skill
- Parallel or isolated execution → sub-agent
- External system access → MCP tool
- Policy enforcement or audit → hook
- Phrase you type often → slash command
The primitives compose. A Skill can call MCP tools, dispatch sub-agents, and rely on hooks for guardrails — the right design usually combines two or three.