The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and search markdown files from a local directory, which serves as an untrusted data source. Maliciously crafted notes could contain embedded instructions intended to influence the agent's behavior.
Ingestion points: The Read tool and search command snippets ingest prose content from notes into the agent's context.
Boundary markers: The skill encourages structured output formats (JSON/CSV) to separate metadata from content, though raw prose is processed during deep reads.
Capability inventory: The skill allows for SQL execution, link-graph traversal, and full-text search across the provided files.
Sanitization: The documentation explicitly instructs users to use parameterized queries with ? placeholders and specific functions like segment() to handle untrusted input safely.
[EXTERNAL_DOWNLOADS]: The skill documentation mentions that the underlying CLI tool will automatically fetch external dependencies and models depending on the configuration.
Automatic Installation: Switching to DuckDB or Turso storage engines triggers the automatic installation of @duckdb/node-api or @tursodatabase/database packages.
Model Fetching: The skill performs runtime downloads of embedding models required for vector similarity searches if they are missing from the local environment.
[DYNAMIC_EXECUTION]: The tool performs dynamic installation and loading of executable database engine packages at runtime.
Runtime Loading: The CLI tool manages the download and installation of native packages (such as the 110 MB DuckDB API) the first time the relevant engine is invoked, loading them into the active process.

sense