The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes large volumes of raw concept stubs, which are often derived from external sources like signals and web ingests. This creates a vulnerability to indirect prompt injection where instructions embedded in notes could influence the synthesis process.
Ingestion points: Raw concept stubs are fetched using gbrain query "type:concept" and gbrain get as described in SKILL.md.
Boundary markers: The skill relies on conventions/quality.md for quote fidelity and back-link enforcement but does not use explicit delimiters to isolate untrusted data during LLM processing.
Capability inventory: The skill possesses file-write (gbrain put) and file-delete (gbrain delete) capabilities within the concepts/ directory.
Sanitization: The skill includes a 'grounding labels' system (VERIFIED, OPINION, NEEDS_SOURCE, UNSAFE) and a 'reputational-risk gate' to identify and remove potentially harmful or false content in Phase 5.
[COMMAND_EXECUTION]: The skill utilizes the gbrain CLI for querying, searching, and modifying the concept database. These commands are restricted to the vendor's internal tooling environment.
Evidence: The invocation section in SKILL.md documents the usage of gbrain query, gbrain search, gbrain put, and gbrain delete to manage the lifecycle of concepts.

concept-synthesis