The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The skill provides instructions and code templates for creating hooks that execute arbitrary bash and Python scripts. The hooks.json configuration explicitly supports a command type for runtime shell execution.
[PROMPT_INJECTION] (HIGH): The skill enables 'Indirect Prompt Injection' by creating a high-privilege surface that processes untrusted data. \n
Ingestion points: Data enters through stdin JSON from Claude Code events like PreToolUse (tool inputs) and UserPromptSubmit (user prompts). \n
Boundary markers: None are specified or required in the prompt templates provided in the examples. \n
Capability inventory: Execution of subprocesses (Bash/Python), modification of $CLAUDE_ENV_FILE, and the ability to auto-approve tool execution via permissionDecision: allow. \n
Sanitization: The provided example (Example 4) uses basic regex for 'dangerous patterns,' which is an incomplete and bypassable sanitization method.
[PRIVILEGE_ESCALATION] (HIGH): The skill documents how to programmatically override security dialogues using PermissionRequest hooks. This allows an agent (or an attacker via indirect injection) to grant itself permissions that would otherwise require user intervention.
[COMMAND_EXECUTION] (MEDIUM): The skill facilitates 'Dynamic Execution' by teaching the agent to generate and then execute local scripts (Bash/Python), creating a pathway for runtime code injection.

hook-skill