The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the llm CLI tool via subprocess.run in scripts/llm_select.py. The implementation uses a list-based argument structure rather than a shell string, which effectively mitigates shell injection vulnerabilities while enabling intelligent selection and term extraction.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, presenting a theoretical injection surface.
Ingestion points: Content is ingested from Google Search result metadata (titles and snippets) and local Markdown files (Obsidian notes) via scripts/llm_select.py and scripts/obsidian.py.
Boundary markers: The system prompts in llm_select.py define strict JSON output formats and instructions, though they lack explicit delimiters or 'ignore' instructions for the interpolated external content.
Capability inventory: The skill has the ability to perform network requests for image downloads (urllib.request) and write to the local file system to update notes and save images.
Sanitization: Filenames are sanitized using a slugify function, and markdown alt-text is escaped to prevent formatting breakages.
[DATA_EXPOSURE]: The skill manages sensitive API keys for Google Search and OpenRouter. These are handled via .env file resolution in scripts/config.py, which is a standard and secure practice. Additionally, the skill's access to Obsidian vault settings (app.json) and note content is restricted to its primary purpose of media integration.

google-image-search