The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use the caliper tool to run evaluations. These evaluations are defined in .eval.yaml files containing setup, cleanup, and assert blocks. The agent executes arbitrary shell scripts and Python code found in these blocks to manage test environments and verify results. This functionality represents an arbitrary command execution surface if the agent is directed to run an untrusted evaluation specification.
[PROMPT_INJECTION]: The summarize reference skill included in the package processes content from external URLs and YouTube links. The skill instructions do not mandate the use of delimiters or other boundary markers when interpolating this external data into the agent's context, which could allow malicious instructions embedded in web content to influence the agent (Category 8).
Ingestion points: External URLs and YouTube links in the summarize skill.
Boundary markers: None specified in the instructions.
Capability inventory: Execution of the summarize CLI and general shell access.
Sanitization: No specific sanitization or filtering logic is provided.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of caliper-eval via pipx. It also references repositories from the vercel-labs organization on GitHub, which is recognized as a well-known and trusted source for development tools. These dependencies are documented for operational setup.
[COMMAND_EXECUTION]: The screenshot reference skill utilizes powershell -ExecutionPolicy Bypass to run helper scripts on Windows systems. While this is a common method for executing local automation scripts, it involves bypassing system-level execution policies.
[REMOTE_CODE_EXECUTION]: The assert: blocks in evaluation files are executed as Python code during the evaluation process. This enables dynamic verification of system state and artifacts after a skill has been run.

evaluate-skill