The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to configure and execute local Python scripts through the Promptfoo framework using the file:// URI scheme.
The SKILL.md file provides examples of loading custom assertions from scripts/metrics.py (e.g., value: file://scripts/metrics.py:custom_assert).
The references/promptfoo_api.md file describes dynamic test generation where Promptfoo executes a local script to return test cases (e.g., tests: file://tests/generate.py:get_tests).
This mechanism allows the runtime loading and execution of code from computed or specified local paths.
[INDIRECT_PROMPT_INJECTION]: The framework is designed to ingest and evaluate untrusted data in the form of LLM outputs.
Ingestion points: Untrusted data enters the context when Promptfoo captures the response from configured LLM providers (Anthropic, OpenAI, etc.).
Boundary markers: No explicit boundary markers or "ignore" instructions for the model output are mentioned for the assertion logic.
Capability inventory: The skill has the capability to execute local scripts and perform network operations via the Promptfoo CLI.
Sanitization: The skill documentation in SKILL.md and scripts/metrics.py suggests basic sanitization (stripping HTML tags) to improve metric accuracy, but not to prevent instruction injection.
[EXTERNAL_DOWNLOADS]: The skill uses the npx promptfoo@latest command to download and execute the Promptfoo CLI tool from the public npm registry.

promptfoo-evaluation