The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the use of custom TypeScript/JavaScript providers and assertions via file:// URIs and inline code blocks (e.g., transform scripts and javascript/python assertion types). This is a standard extensibility feature of the Promptfoo tool for complex evaluation logic.
[INDIRECT_PROMPT_INJECTION]: As a tool designed to evaluate LLM outputs, there is an inherent risk surface where malicious payloads in the text being evaluated could attempt to influence the evaluation logic. The skill includes examples of red teaming to detect such issues in target models, but the evaluation framework itself processes this untrusted data.
Ingestion points: Evaluation results (LLM output), variable interpolation in promptfooconfig.yaml, and test data loaded from external YAML/CSV files (e.g., tests/cases.csv).
Boundary markers: Not explicitly demonstrated in the provided configuration examples.
Capability inventory: Ability to execute shell commands (promptfoo eval), run JavaScript code (transform, javascript assertion), execute Python code (python assertion), and perform network operations via provider configurations.
Sanitization: The skill relies on user-defined configurations and does not explicitly demonstrate sanitization of external variables interpolated into prompts.
[DATA_EXFILTRATION]: The documentation references the --share flag and the promptfoo share command, which upload results to a remote server. The skill includes a clear warning against using this feature with sensitive data unless self-hosting the results server.
[COMMAND_EXECUTION]: The skill utilizes the npx command to run the evaluation framework. This is a standard development practice for executing CLI tools from the npm registry.

ai-observability-promptfoo