The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage files and system state.
Uses ln -s (POSIX) and mklink (Windows) to create directory symbolic links directly inside agent configuration folders (~/.claude/skills, ~/.codex/skills, ~/.agents/skills).
Executes git commands (add, commit, push) to publish local skill data to a remote repository.
Performs folder movement (mv) and removal (rm, rmdir) during renaming operations.
[EXTERNAL_DOWNLOADS]: The skill uses npx @younndai/yon-parser, which downloads an external package from the public npm registry at runtime.
[REMOTE_CODE_EXECUTION]: The use of npx to download and run a package from an unverified source (@younndai) represents a remote code execution vector.
[DATA_EXFILTRATION]: While intended for publishing skills, the capability to perform git push on directories containing potential configuration data could be abused to exfiltrate information if the destination repository is compromised or redirected.
[PROMPT_INJECTION]: The skill creates an indirect prompt injection surface by generating new instructions (SKILL.md, protocol.yon) based on unvalidated user input. These generated files are then automatically linked into the agent's execution environment, potentially allowing the creation of malicious skills through the scaffold tool.
[COMMAND_EXECUTION]: The skill provides a Python snippet using shutil.rmtree to remove directories on Windows, which involves executing arbitrary script logic to bypass shell restrictions.

new-skill-creator