Claude Code Extensions: Skills, Hooks, Subagents, and MCP

Skills provide workflow knowledge; hooks react to events; subagents handle separate work with their own context and configured access; MCP connects tools or data. A plugin can package capabilities. Installing an extension does not by itself make every action appropriate.

Write the repeated problem first. Review any extension source, script, server provenance, and requested permissions. This exercise chooses a design; it does not install an MCP server or add an event hook.

If you repeatedly forget the same label and keyboard checks, start with a written checklist or skill. If you need a separate read-only review of a finished diff, a bounded subagent may help. A hook that executes a command after edits introduces automatic execution. An MCP connector is unnecessary unless the task genuinely needs an external capability.

• Knowledge need: a repeatable “inspect labels, tab order, and narrow viewport” checklist.
• Independent work: a reviewer returns findings without editing the implementation files.
• External need: none for this local form example; do not connect a customer ticket system just because it is available.

1. State the desired input and output without naming a technology.
2. Classify the need as reusable knowledge, event-triggered execution, bounded side work, or an external tool/data connection.
3. Inspect the proposed extension’s capabilities and failure behavior. For code-bearing extensions, review what actually executes.
4. Trial the smallest option on a disposable example and compare its output with a manual checklist.

The choice should reduce a known source of error without adding unnecessary access. For the sample form, a reusable review checklist is enough until a real need for automation or external data appears.

Stop if a hook runs commands you cannot explain, a connector asks for write access for a read-only task, or two workers can edit the same file without coordination. Remove the capability from the plan until its purpose and boundary are clear.

• Can you explain why this extension is necessary?
• Are side effects visible and reversible where possible?
• Does a failure leave a clear record instead of silently skipping the review?