The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses a subagent to analyze the conversation history (last 10-30 messages) to identify unwanted behaviors. This history is untrusted data that could contain adversarial instructions intended to influence the subagent's findings or the resulting rule generation.
Ingestion points: Conversation messages processed in Step 1 via the subagent prompt.
Boundary markers: The task instructions for the subagent define the scope, though specific delimiters for user messages are not explicitly defined.
Capability inventory: The skill can write files (write_to_file) to the .claude/ directory and execute local commands (mkdir, python3, ls) via run_command.
Sanitization: No specific sanitization or escaping is mentioned for the patterns extracted from the conversation, relying instead on user confirmation in Step 2.
[COMMAND_EXECUTION]: The skill instructs the agent to perform local system operations, such as creating directories (mkdir -p .claude) and running diagnostics in the troubleshooting section (python3 -c "import re; ...", ls). These are standard administrative tasks for the skill's functionality.
[EXTERNAL_DOWNLOADS]: The instructions require loading an external dependency (hookify:writing-rules) using the Skill tool to understand rule syntax. This indicates a modular dependency on another skill from the same provider.

hookify