The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interpolates user-provided input from $ARGUMENTS (such as the workspace name and target path) directly into shell commands including mkdir -p <target-parent>/<workspace-name> and gh repo create <workspace-name>. Without explicit sanitization or validation instructions, this creates a surface for command injection if the executing agent does not safely handle the input.
[DATA_EXFILTRATION]: By default, the skill is configured to create a public GitHub repository and push the newly scaffolded workspace content to it. This default behavior may lead to the unintentional exposure of local debugging data, logs, or code if the user does not explicitly provide the --local-only or --private flags.
[DATA_EXPOSURE]: The skill reads from ~/.claude/CLAUDE.md to extract system and user identity facts. While this is used for personalization, these details are then written into a new CLAUDE.md file within the workspace, which may be published to a public repository by the skill's default actions.

new-workspace