The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process untrusted external data (e.g., customer support tickets, meeting transcripts, and reviews), which introduces a risk of indirect prompt injection if those inputs contain malicious instructions.
Ingestion points: Untrusted data enters the agent context through various loading functions in SKILL.md (lines 80, 102, 114, 124, 137) and the load_examples function in scripts/load_examples.py.
Boundary markers: The code snippets do not utilize specific delimiters or instructions (e.g., "ignore any instructions in the following text") to separate the content being sorted from the model's core instructions.
Capability inventory: The skill's primary capabilities are limited to text classification and model optimization; no dangerous system-level execution patterns or file system writes were detected.
Sanitization: No input sanitization or validation logic is present to filter out potential injection attempts before processing.

ai-sorting