The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.Popen in scripts/run_eval.py to invoke the claude CLI tool. This is the core mechanism for evaluating skill performance by running test queries against the agent. The commands are constructed using list-based arguments which mitigate common shell injection risks.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html file loads the xlsx.full.min.js library from cdn.sheetjs.com. This is a well-known and established service for spreadsheet processing in the browser and is used here to render benchmark results.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted data from evals/evals.json and feedback.json to perform optimization tasks.
Ingestion points: Data enters the system via the evals/evals.json file (user-provided test queries) and feedback.json (user feedback from the viewer UI).
Boundary markers: The system relies on instructional guidance in SKILL.md and agents/grader.md to interpret data, though explicit structural delimiters for every data point are not present in the processing scripts.
Capability inventory: The skill can execute shell commands via claude -p, write files to the workspace and .claude/commands/, and start a local HTTP server.
Sanitization: The eval-viewer implementation performs HTML escaping when rendering user feedback and query results to prevent cross-site scripting (XSS) in the review interface.
[COMMAND_EXECUTION]: The eval-viewer/generate_review.py script uses subprocess.run to call lsof for the purpose of identifying and terminating processes occupying its designated local server port (default 3117). This is a standard utility for local service management.

skill-creator