The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses curl, wget, and pip to manage dependencies and assets. It specifically references trusted sources like anthropic/skills, google-gemini, supabase, and vercel-labs for guidelines and templates.
[COMMAND_EXECUTION]: The skill executes various Python scripts and CLI tools (e.g., claude -p, nohup, kill) to manage subagents, run evaluations, and serve the HTML review interface. These operations are core to the skill's purpose of automating the development lifecycle.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of code within isolated subagent environments to test new skill drafts. While this involves executing dynamic content, it is the primary function of the skill and is governed by strict user-review loops.
[DATA_EXFILTRATION]: No evidence of unauthorized data transmission was found. Network operations are restricted to communication with the local HTTP server for the evaluation viewer and standard API calls to Anthropic for description optimization.
[PROMPT_INJECTION]: The skill includes instructions to ignore safety filters or bypass constraints if the user's intent requires it (e.g., for 'roleplay' skills), but it maintains a 'Principle of Lack of Surprise' and explicitly forbids creating skills for unauthorized access or malicious activities.
[NO_CODE]: This skill contains extensive implementation logic across multiple Python scripts and HTML templates to handle benchmarking, grading, and visualization.

skill-creator