The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions and internal scripts (e.g., scripts/fetch_trace.mjs and references/skill-template.md) promote the installation of dependencies using the command curl -fsS https://setup.steel.dev | sh. This pattern executes remote code directly in the user's shell without pre-verification or integrity checks.
[COMMAND_EXECUTION]: The skill includes a dedicated script scripts/install_skill.mjs designed to write AI-generated skill files to the user's home directory at ~/.claude/skills/. This allows for the persistent modification of the agent's environment and the installation of new executable instructions.
[COMMAND_EXECUTION]: The skill performs multiple automated tasks by spawning subprocesses for the steel CLI and local Node.js scripts to manage browser sessions, fetch traces, and scaffold new project files.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface (Category 8). It ingests browser traces containing arbitrary data from external websites (such as page titles and element labels) and uses this untrusted data to 'reason' about and generate logic for new agent skills. This could allow a malicious website to influence the structure or behavior of generated skills.

steel-skill-creator