The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to read local artifacts, such as code, plans, and designs, and transmit them to external services including Vercel AI Gateway and the TypeSafe API for processing. While these are well-known services, the transmission of potentially proprietary artifacts constitutes a data exposure surface.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute local Python scripts (scripts/grill-jev.py, scripts/jev-budget-check.py) using shell commands or subprocess.run. This execution model relies on the proper sanitization of arguments like file paths and context strings provided by the user or extracted from artifacts.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted artifacts to generate interrogation questions, which creates a vulnerability to indirect prompt injection where instructions embedded in the artifact could influence the LLM's analysis.
Ingestion points: The skill reads artifacts via the Read tool and passes them as input to the grill-jev.py script using --file or --text flags.
Boundary markers: The instructions for generating the question battery do not specify the use of delimiters or clear separation between the artifact content and the system instructions.
Capability inventory: The skill has access to Read, Write, Edit, and Bash tools, and performs network operations via Python scripts to external APIs.
Sanitization: There is no evidence of sanitization or filtering of the artifact content before it is interpolated into the question generation prompt.
[DYNAMIC_EXECUTION]: The skill generates a JSON-formatted question battery at runtime and writes it to a temporary file (/tmp/grill-questions.json), which is then loaded and executed by the script to drive the interrogation process.

grill-jev