Chapter C65 · Interactive Course
Guide Claude Code with context and extensions
Project guidance, diagnosing a rule that did not apply, skills, narrow connections, delegation and hooks.
Core Learning Concept
Configure the fictional session by selecting relevant guidance and capabilities. Inspect the applicable CLAUDE.md files, use a skill only for a matching recurring job, and keep connected access aligned with the task. Give any delegated investigation a bounded question and review executable hooks as code.
What You’ll Learn in This Chapter
-
CLAUDE.md holds project guidance
Choose concise conventions the session should read.
CLAUDE.md provides project guidance loaded by Claude Code for local conventions, architecture, and verified commands. Keep instructions specific and current, but remember rules guide the model without replacing permissions or test checks.
-
Inspect loaded guidance
Diagnose an instruction that did not apply.
When guidance seems surprising, inspect the files and scopes the product loads, including applicable higher-level or more specific instructions. The visible project file may not be the whole input. Compare the actual convention with current documentation before editing a rule that may not be the one taking effect.
-
Use a skill for a repeated job
Choose an appropriate installed workflow.
Use a skill to package a recurring procedure with clear selection criteria and supporting resources. In Claude Code, discovery and invocation follow its documented skill conventions. Confirm that the chosen skill fits the task, read its procedure, and inspect the output. Reuse does not remove the need for verification.
-
Connect a service only when needed
Choose a narrow MCP capability for a task.
Connect an external service when the task needs a capability it provides. Inspect the server or integration identity, requested access, and intended data flow. A broad connection is not automatically useful. Start with the smallest suitable operation and verify its result before relying on a larger workflow.
-
Delegate one bounded investigation
Use a subagent with a specific evidence request.
Delegation works best with a concrete question, bounded scope, and expected results. A subagent investigating a parser returns findings while another task continues independently. Verify context and tool access deliberately under the host.
-
Hooks need execution review
Inspect an event-triggered action before enabling it.
Hooks react to configured events and may run commands or other behavior. Inspect their trigger, executable contents, dependencies, and permissions before enabling them. A hook that formats files has different effects from one that contacts a service. Automatic invocation makes understanding the behavior more important, not less.
-
Configure the toy session
Choose one project rule, one applicable skill and an optional narrow connection, and identify an extension the task does not need.