Hook Development

Tool usage validation in Claude Code is handled by PreToolUse hooks, which intercept input before execution. You can implement prompt-based LLM reasoning or command-based deterministic checks to enforce policies and block invalid operations.

Claude Code plugin workflows support PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart, and SessionEnd events. These events enable automated context loading, policy enforcement, and reactions to tool results across active sessions.

SessionStart hooks load environment context when a Claude Code session initializes. You reference environment variables using the ${CLAUDE_ENV_FILE} token within your plugin's hooks.json configuration to inject session-specific data.

Claude Code plugin hooks should be structured in a hooks.json file using ${CLAUDE_PLUGIN_ROOT} for portable path references. Include YAML frontmatter in SKILL.md for discovery and define both prompt-based and command-based hooks for your events.

Claude Code hooks support both prompt-based and command-based formats. Prompt-based hooks apply LLM-driven reasoning for complex validation, while command-based hooks execute deterministic scripts for strict policy enforcement and automation.

Claude Code hooks fail to access plugin resources when paths are hardcoded instead of using the ${CLAUDE_PLUGIN_ROOT} variable. Reference plugin resources via ${CLAUDE_PLUGIN_ROOT} in your hooks.json to ensure portable path resolution across environments.