The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes built-in mitigations for indirect prompt injection risks that occur when analyzing external websites. It provides clear instructions to the agent to distinguish between structural site data and potentially malicious instructions embedded in web content.
Evidence: SKILL.md contains a specific 'Security: indirect prompt injection risk' section instructing the agent to treat captured content as raw data and ignore directives.
[COMMAND_EXECUTION]: The skill describes a workflow for generating project tasks and scripts, including a local installation script for new capabilities. This process is transparent and requires manual user execution, preventing silent system modification.
Evidence: SKILL.md suggests generating a skills/install.sh script to move files from a development directory to the agent's configuration folder.
[EXTERNAL_DOWNLOADS]: The skill recommends the installation of well-known, trusted tools for media and browser analysis to support its primary functions.
Evidence: SKILL.md prompts for the installation of @anthropic-ai/agent-browser and ffmpeg using standard package managers.

ralph-kage-bunshin-start