The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the subprocess module to manage the development environment and execution lifecycle.
Evidence includes calls to claude -p for testing skill triggers in scripts/run_eval.py, npx tsx for running backtests in implementation plans, and lsof to manage network ports in eval-viewer/generate_review.py.
[EXTERNAL_DOWNLOADS]: The skill requires network access to interact with external APIs.
scripts/improve_description.py and scripts/run_loop.py use the anthropic Python SDK to communicate with remote AI models for description optimization.
The evaluation viewer in eval-viewer/viewer.html loads the xlsx.full.min.js library from cdn.sheetjs.com at runtime to render spreadsheet data.
[DATA_EXFILTRATION]: The eval-viewer/generate_review.py script starts a local HTTP server (defaulting to port 3117) to provide a user interface for reviewing results.
This script reads contents from the evaluation workspace, including text files, images, PDFs, and spreadsheets, and serves them via the local server.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes and executes user-provided data through subagents.
Ingestion points: Untrusted data enters the context via evals/evals.json (test prompts) and feedback.json (user reviews).
Boundary markers: No explicit delimitation or 'ignore instructions' warnings are applied to the test prompts before they are passed to subagents.
Capability inventory: The skill possesses capabilities for subprocess execution, file system modification (Path.write_text, shutil.copytree), and network communication.
Sanitization: There is no evidence of sanitization or escaping performed on the content of test prompts before execution.

skill-creator