The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Python subprocess module to execute local shell commands. Specifically, scripts/run_eval.py and scripts/improve_description.py invoke the claude CLI to run tests and optimize skill triggering. Additionally, eval-viewer/generate_review.py executes system commands like lsof and kill to manage local network ports for the evaluation viewer.
[INDIRECT_PROMPT_INJECTION]: The skill operates on potentially untrusted data by ingesting and executing user-defined test prompts and evaluation queries.
Ingestion points: Test prompts located in evals/evals.json and generated queries used for triggering analysis.
Boundary markers: The instructions suggest using delimiters to scope external content, but the scripts lack strict automated sanitization of the input queries before they are passed to the shell.
Capability inventory: Includes shell command execution via sub-processes, file writing for logging and workspace management, and spawning local network servers.
Sanitization: The skill uses standard YAML and JSON parsing but does not implement additional security-focused sanitization for the prompts passed to the agent.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html template fetches the SheetJS (xlsx) library from a CDN (cdn.sheetjs.com) to render spreadsheet data in the browser. SheetJS is a well-known and standard library for these tasks.

skill-creator