The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill functions as an evaluation framework that ingests and processes untrusted data, including user-provided skill instructions, test queries, and execution transcripts.
Ingestion points: scripts/run_eval.py (processes test queries), agents/grader.md (evaluates transcripts and output files), and agents/analyzer.md (compares skill outputs).
Boundary markers: The subagents lack explicit boundary markers or instructions to disregard potential injection patterns within the processed content.
Capability inventory: The skill possesses the ability to write files (scripts/run_eval.py), execute CLI commands (subprocess calls in multiple scripts), and host a local web server (eval-viewer/generate_review.py).
Sanitization: There is no evidence of sanitization or escaping of external content before it is interpolated into agent prompts or rendered in the HTML viewer.
[COMMAND_EXECUTION]: The skill uses subprocess to execute system commands necessary for its operation.
scripts/run_eval.py and scripts/improve_description.py invoke the claude CLI tool to perform evaluations and optimizations.
eval-viewer/generate_review.py uses lsof and kill to manage process conflicts on the network port used for its local viewer.
[EXTERNAL_DOWNLOADS]: The HTML-based review tools reference assets from well-known external services.
assets/eval_review.html and eval-viewer/viewer.html load styles and scripts from Google Fonts and the SheetJS CDN to facilitate the UI and spreadsheet rendering.
[DYNAMIC_EXECUTION]: scripts/run_eval.py dynamically generates skill manifest files in the .claude/commands/ directory. These files define the temporary skill environment used during the evaluation phase.

skill-creator