The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external customer reviews and comments located in project subdirectories (inputs/reviews/, inputs/comments/) and forum content (Dcard, PTT). This creates a surface for indirect prompt injection where malicious instructions could be embedded in customer feedback. The skill implements specific guardrails: 1. Ingestion points: Untrusted text data enters the agent context via the input library and registry path. 2. Boundary markers: Hard Rule 3 explicitly instructs the agent to treat external content strictly as data and never as instructions. 3. Capability inventory: The skill performs file writes to local output directories and generates CSV files, but lacks direct shell execution or unauthorized network write capabilities. 4. Sanitization: A mandatory four-step CSV formula-injection guard is defined to sanitize data before writing to CSV formats, including loops to strip hidden characters like the Unicode Byte Order Mark (BOM).
[COMMAND_EXECUTION]: The skill includes defensive measures against command and formula injection in exported CSV files. It specifies a rigorous normalization process to strip whitespace and BOM (U+FEFF), and prepends single quotes to dangerous leading characters (=, +, -, @, tab, CR, LF) to prevent spreadsheet software from evaluating content as an executable formula.
[DATA_EXFILTRATION]: The skill accesses local project files such as brand profiles and positioning documents. It maintains a strict provenance schema and limits data handling to local project directories. No unauthorized network exfiltration is observed, and all external URL references point to official platform documentation for Google, Meta, and TikTok.

ad-creative