The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.Popen and subprocess.run within scripts/run_eval.py and scripts/improve_description.py to invoke the claude CLI and other internal scripts for evaluation and description optimization.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data such as user feedback and subagent execution results, creating an attack surface where malicious input could influence the agent's behavior during skill improvement.
Ingestion points: Feedback provided via the feedback.json file (generated by the viewer) and outputs from claude -p runs are used to refine skill instructions.
Boundary markers: The instructions do not define strict delimiters when interpolating untrusted feedback into prompts for the improvement model.
Capability inventory: The skill possesses the ability to execute shell commands via subprocess, perform local file writes, and start a local web server.
Sanitization: There is no evidence of explicit sanitization or escaping of feedback data before it is included in model prompts.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html file loads the SheetJS library (xlsx.full.min.js) from cdn.sheetjs.com to enable spreadsheet rendering in the review interface. This is a reference to a well-known service for document processing.
[DYNAMIC_EXECUTION]: scripts/run_eval.py dynamically generates skill definition files in the .claude/commands/ directory. These files are then interpreted and executed by the claude CLI to simulate and test triggering behavior.
[DATA_EXFILTRATION]: eval-viewer/generate_review.py implements a local HTTP server that binds to 127.0.0.1. This server is used to host the evaluation results viewer and capture user feedback locally.

skill-creator