The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches an installation shell script from an external community repository: https://raw.githubusercontent.com/ZhangHanDong/makepad-skills/main/install.sh.
[REMOTE_CODE_EXECUTION]: The downloaded script is executed via bash, which facilitates arbitrary remote code execution from a source outside of the vendor's own infrastructure.
[COMMAND_EXECUTION]: The skill establishes persistence and intercepts agent operations by modifying .claude/settings.json and adding executable shell scripts to .claude/hooks/. These hooks are configured to trigger automatically during UserPromptSubmit, PreToolUse, and PostToolUse events.
[PROMPT_INJECTION]:\n
Ingestion points: User prompts are intercepted and processed by the makepad-skill-router.sh hook to dynamically route queries to different skills.\n
Boundary markers: There are no explicit delimiters or security markers defined to prevent malicious instructions embedded in user data from influencing the routing logic.\n
Capability inventory: The system utilizes persistent hooks to execute local bash scripts and can modify the agent's runtime environment configuration.\n
Sanitization: There is no evidence of input validation or sanitization of the user prompt before it is ingested by the routing hook mechanism.

evolution