The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as sales call transcripts and web reviews, which is a vector for indirect prompt injection. \n
Ingestion points: Research data provided by the user or through connectors like ~~call recordings and ~~CRM (SKILL.md). \n
Boundary markers: The skill uses [A] and [NEEDS PROOF] tags to label data, but does not provide explicit boundary markers or instructions to the model to ignore potential instructions within that data. \n
Capability inventory: The skill is capable of reading and appending to local files including /foundation/brain.md and /context/skill-sessions.md (SKILL.md). \n
Sanitization: No specific sanitization or filtering of the external research content is defined.

alternatives-map