The Agent Skills Directory

[SAFE]: The skill is purely informational, consisting of Markdown documentation and YAML templates. It does not contain executable code, scripts, or active commands that could be exploited.
[EXTERNAL_DOWNLOADS]: All external URLs referenced in the documentation and references are from trusted organizations (Anthropic, Vercel/Promptfoo, LangChain) or well-known academic/benchmark sites (arXiv, SWE-bench, OSWorld). These are used for attribution and further reading.
[COMMAND_EXECUTION]: While the skill mentions commands like ruff, mypy, and bandit in templates, these are provided as examples of static analysis tools for the user to implement in their own evaluation pipelines. The skill itself does not execute these commands.
[INDIRECT_PROMPT_INJECTION]: The skill acknowledges the risk of model-based graders being bypassed and recommends robust rubrics and human calibration to mitigate these risks in the user's implementation.
[CREDENTIALS_UNSAFE]: No hardcoded credentials or sensitive environment variable patterns were detected. Templates use placeholders like <relative-path> or path: "src/auth/*".

anthropic-evaluations