The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes third-party document corpora which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The skill reads external data via gbrain search and gbrain get_page in Pipeline Step 1 and 2.
Boundary markers: The skill includes explicit instructions for the AI to treat document text as DATA, not instructions, referencing the untrusted-content.md convention.
Capability inventory: The skill has the capability to modify the database by creating edges using the gbrain link command.
Sanitization: The instructions require the model to use judgment based on the citation context and explicitly ignore any imperatives (commands) found within the documents.
[COMMAND_EXECUTION]: The skill executes native system commands to interact with its knowledge base.
Evidence: Uses gbrain link, gbrain graph-query, and gbrain search to perform its primary function. These commands are scoped to specific data sources and are used for legitimate graph-building and verification purposes.
[SAFE]: The skill implements several safety and integrity measures.
It mandates a "test-before-bulk" approach to verify results on a small sample.
It requires a verification walk (gbrain graph-query) to confirm that actions were performed correctly before reporting success.
It uses a mandatory provenance stamp (--link-source citation-graph) to ensure that all writes are auditable and can be bulk-removed without affecting other data.

citation-graph-ingest