Customization surfaces

This page is retained as a compact cross-reference. Use the detailed guides for implementation:

• Choose the right surface
• Skills
• Plugins
• Hooks, rules, and permissions
• MCP, tools, and subagents
• Repository file layout

Composition model

instructions route work
    ↓
skills define repeatable judgement and procedure
    ↓
scripts perform deterministic mechanics
    ↓
MCP/tools provide bounded capability and current data

artifacts preserve evidence and handoff state

hooks react automatically to lifecycle events
rules/permissions constrain authority
plugins package stable combinations for distribution

The arrows describe common composition, not authority inheritance. A skill does not grant tool access; an MCP server does not decide the workflow; a hook does not replace permissions; and a plugin does not make its bundled code trusted.

Lifecycle

1. Prove a need with a prompt, existing command, and repo documentation.
2. Create a local skill when the procedure repeats and needs judgement.
3. Extract repeatable mechanics into scripts shared with CI.
4. Add hooks only for behavior tied to a specific event.
5. Add MCP only for external state or actions.
6. Add native rules and permissions for security boundaries.
7. Package as a plugin only after routing, outcomes, dependencies, and safety have stabilized for more than one consumer.
8. Evaluate, version, observe, and retire the capability as a maintained product.

Non-negotiable distinctions

• Automatic is not mandatory. Implicit skill routing can fail; mandatory workflows need explicit instruction and deterministic enforcement.
• Interception is not containment. A pre-use hook may miss equivalent tool paths; permissions, sandboxing, and per-tool policy define authority.
• Packaging is not trust. Plugin skills, hooks, scripts, MCP servers, and remote updates remain executable supply-chain inputs.
• Parallelism is not free. Subagents multiply contexts and tool work; use them for independent bounded tasks.
• Shared standard is not identical configuration. OpenAI and Anthropic use related concepts with different discovery, trust, event, and manifest semantics. Share intent and portable components, then test each client.
• A final message is not durable evidence. Plans, logs, diffs, generated artifacts, and validation results must survive compaction and reconnect.
• More context is not more legibility. Expose queryable UI and operational signals, then load only those required by the current workflow.