The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs extremely assertive language ("MANDATORY", "STOP", "FORBIDDEN") and explicitly defines "anti-memorization" protocols intended to bypass the AI's standard training defaults and internal behavior patterns. Specifically, the 'mobile-design-thinking.md' file states its purpose is to prevent the use of memorized patterns and AI defaults.\n- [COMMAND_EXECUTION]: The SKILL.md file explicitly instructs the AI to execute a validation script (scripts/mobile_audit.py) with the directive "(don't read, just run)". While the script itself appears to perform benign regex-based audits of mobile UX patterns, encouraging the execution of code without review is a risky practice.\n- [INDIRECT_PROMPT_INJECTION]: The mobile_audit.py script functions by ingesting and analyzing local files from a user-provided project directory. Because the script's output (which may contain snippets or data from untrusted user code) is then processed by the AI, this creates a potential surface for indirect prompt injection.\n
Ingestion points: scripts/mobile_audit.py reads and processes the content of all source files in a directory supplied as an argument to the tool.\n
Boundary markers: The instructions do not establish boundary markers or provide the agent with guidance on how to safely process the audit tool's output.\n
Capability inventory: The skill has access to the Bash, Read, Glob, and Grep tools, which are used to execute the script and analyze the project filesystem.\n
Sanitization: There is no evidence of sanitization or escaping for the file content retrieved by the script before it is output to the agent's context.

mobile-design