The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to use the Bash tool to execute shell commands (pdftotext and pandoc) using variables derived from user input ($0). While the examples provide quoting, passing user-supplied file paths to a shell environment carries inherent risks.
Evidence: TMP=$(mktemp -t evals).txt && pdftotext "$0" "$TMP" (SKILL.md).
Evidence: TMP=$(mktemp -t evals).txt && pandoc "$0" -t plain -o "$TMP" (SKILL.md).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process free-text student evaluations, which are untrusted external data sources. This content could contain malicious instructions intended to manipulate the agent's analysis or the resulting teaching improvement plan.
Ingestion points: The primary input ($0) contains raw student comments (SKILL.md).
Boundary markers: The instructions lack explicit requirements for using delimiters or "ignore embedded instructions" warnings when presenting the evaluation content to the model for clustering and classification.
Capability inventory: The skill utilizes Read, Write, and Bash tools to process the data and save the final report (SKILL.md).
Sanitization: The skill mentions anonymizing quotes by stripping identifying details, but does not provide a mechanism for sanitizing or validating the content against prompt injection attempts designed to influence the teaching plan recommendations.

respond-to-eval