The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-controlled markdown files from an Obsidian vault, which creates a surface for indirect prompt injection if notes contain instructions intended to influence the agent.
Ingestion points: Notes are read from the local filesystem using read_file and search_files based on the OBSIDIAN_VAULT_PATH environment variable.
Boundary markers: The skill contains high-quality security instructions, requiring the agent to resolve absolute paths and confirm that target notes remain inside the vault after resolving symlinks. It explicitly states that "private notes are not uploaded to external services."
Capability inventory: The skill utilizes read_file, write_file, patch, search_files, and limited terminal access for path resolution.
Sanitization: The instructions mandate human clarification before actions that could overwrite work or expose private data, providing a manual verification layer.

obsidian