The Agent Skills Directory

Dynamic Execution (MEDIUM): The skill demonstrates how to configure Promptfoo to load and execute custom Python scripts from the local filesystem for assertions (e.g., file://scripts/metrics.py:custom_assert). This allows for arbitrary code execution in the evaluation environment.\n- External Downloads & Remote Code Execution (LOW): The instructions suggest using npx promptfoo@latest, which downloads and executes the package directly from the npm registry.\n- Data Exposure (LOW): Documentation in SKILL.md reveals a specific local path (/Users/tiansheng/Workspace/prompts/tiaogaoren/), exposing a system username.\n- Indirect Prompt Injection (LOW): The skill identifies a pattern where data is loaded from files and interpolated into LLM prompts without explicit sanitization or boundary markers.\n
Ingestion points: promptfooconfig.yaml and tests/cases.yaml use the file:// scheme to import external content.\n
Boundary markers: Absent in the provided prompt templates.\n
Capability inventory: Promptfoo executes shell commands (npx), runs Python scripts, and makes network requests to external providers.\n
Sanitization: Absent.

promptfoo-evaluation