Study notes · 1% of the exam

Claude Hooks

Use Claude Code and Agent SDK hooks as deterministic guardrails: pick the right event and matcher, block correctly, place the configuration at the right scope, and know how hooks interact with permission rules.

Key points

1
Hooks run in the harness at fixed lifecycle points, so they fire every time. Use them for rules that must hold, such as blocking rm -rf, force-pushes or edits to protected files, rather than relying on CLAUDE.md or prompt instructions.
2
PreToolUse runs before a tool executes and is the event that can prevent it. PostToolUse runs after success, which suits linters, formatters and feedback, but the action has already happened.
3
Blocking from PreToolUse: exit code 2 (stderr becomes the reason shown to Claude), or exit 0 with JSON hookSpecificOutput.permissionDecision: "deny" plus permissionDecisionReason. Exit code 1 is a non-blocking error and the call proceeds.
4
Command hooks receive JSON on stdin, including tool_name and tool_input (for Bash, the full command; for Edit and Write, file_path), so the hook can inspect arguments.
5
Matchers filter by tool name for tool events: "Bash" or "Edit|Write" match exactly, a pattern with other characters is a regex (such as mcp__.*), and "*" or an empty matcher matches everything.
6
UserPromptSubmit runs before Claude processes a prompt and can reject it, for example a prompt that contains a pasted secret. It sees user prompts, not commands Claude generates.
7
Stop fires when Claude finishes responding. Returning decision: "block" with a reason makes Claude keep working, for example until tests pass. Check the stop_hook_active input so the hook does not keep blocking on a condition Claude cannot fix.
8
Scope comes from where the hook is defined: ~/.claude/settings.json (you, all projects), .claude/settings.json (committed and shared with the team), .claude/settings.local.json (personal, gitignored), managed policy settings (organization-wide), plugins, and Skill or agent frontmatter.
9
Hooks merge across levels, and disableAllHooks set in user, project or local settings cannot disable managed hooks. Admins can use allowManagedHooksOnly to block all non-managed hooks.
10
Hooks and permission rules layer. A hook "allow" skips a prompt but does not override matching deny or ask rules, while a hook that exits 2 blocks even when an allow rule matches.
11
A hook matched to Edit and Write does not see Bash. If the shell can reach the same files, add sandbox filesystem restrictions or also cover Bash.
12
In the Agent SDK, hooks are callback functions passed in the options (for example on PreToolUse) that can deny or allow calls. They run before permission rules and the canUseTool callback.

Read the source

Test yourself on Claude Hooks

Ten questions, with the answer and explanation after each one.