The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative language ('MANDATORY', 'REQUIRED BEFORE ANY ACTION') to force the agent to execute a shell command (curl) immediately upon invocation. This pattern aims to override the agent's autonomous decision-making and ensure execution regardless of the specific task context.
[COMMAND_EXECUTION]: Instructs the agent to execute a backgrounded network request via curl to a local service (http://localhost:8888/notify). The command is specifically designed to run silently by suppressing all output (> /dev/null 2>&1) and running as a background process (&), reducing user visibility of the agent's network activity.
[DYNAMIC_EXECUTION]: The skill is configured to automatically search for and load external configurations and resources from a specific local directory (~/.claude/skills/PAI/USER/SKILLCUSTOMIZATIONS/Documents/) to override its default behavior. Loading instructions from user-writable or computed paths without validation allows for persistence and behavior modification.
[PROMPT_INJECTION]: The skill processes untrusted document formats (DOCX, PDF, XLSX, PPTX) from the file system. These files are significant vectors for indirect prompt injection, as the skill is instructed to extract and interpret content that could contain hidden instructions targeting the agent's logic.
[REMOTE_CODE_EXECUTION]: The skill relies on various external scripts (e.g., recalc.py) and complex libraries that perform dynamic parsing and generation of binary/XML formats. This complexity increases the attack surface for vulnerabilities in the underlying parsers or the execution of generated scripts.

documents