The Agent Skills Directory

[SAFE]: The skill documents the integration and configuration of Promptfoo, a well-known open-source evaluation framework. The provided patterns follow industry best practices for LLM testing.
[COMMAND_EXECUTION]: The skill suggests using npx promptfoo@latest to initialize and run evaluations. This is the standard method for executing the Promptfoo CLI from the official NPM registry and does not involve suspicious or obfuscated commands.
[DYNAMIC_EXECUTION]: The skill utilizes Promptfoo's native capability to run local Python scripts for custom assertions (scripts/metrics.py). The included script performs benign text processing, such as word count validation and HTML tag removal, using standard library modules like re.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for ingesting external data via file:// references in vars and tests configurations. While this is an ingestion point for untrusted content, the risk is mitigated by the skill's purpose (evaluation) and the use of standard templating delimiters ({{var}}). Mandatory evidence chain:
Ingestion points: Data is loaded from local files such as data/context.txt and tests/cases.yaml.
Boundary markers: The skill uses Promptfoo's default double-curly-brace interpolation.
Capability inventory: Capabilities are limited to text processing and LLM calls, with no unsafe interpolation into system shell commands or network exfiltration operations.
Sanitization: The documented Python metrics include examples of HTML tag stripping and text normalization to sanitize model outputs before scoring.

promptfoo-evaluation