The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external document content which is then stored in a knowledge graph. The author explicitly documents this as an injection surface in the 'agent-access' documentation and provides mitigations.
Ingestion points: Document text extracted via the pipeline defined in references/pipeline.md.
Boundary markers: The skill recommends serving graph content as 'quoted evidence' and using tool descriptions to explicitly frame output as data rather than instructions in references/agent-access.md.
Capability inventory: The system defines a narrow tool surface (lookup_entity, neighbors, search_communities, add_facts) to interact with the graph, as detailed in references/agent-access.md.
Sanitization: Mitigation strategies include instructions for agents to treat graph content as findings to report rather than actions to take.

knowledge-graph