The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the gbrain memory system, including voice profiles and person-subject pages. If these source materials are influenced by external or malicious content, they could contain instructions designed to manipulate the agent's behavior during the drafting process.
Ingestion points: Data is pulled from people/<slug>-voice and people/<slug> via gbrain get, as well as broader searches using gbrain search and gbrain timeline in SKILL.md.
Boundary markers: The skill relies on structural conventions (Markdown headers and YAML blocks) but lacks explicit delimiters or instructions to ignore potential commands embedded within the voice samples or context data.
Capability inventory: The skill has the ability to write new pages to the people/ directory using gbrain put and generates prose content intended for human use.
Sanitization: There is no evidence of text sanitization, escaping, or filtering for instructions before the data is interpolated into the agent's context for ghostwriting.

draft-in-voice