The Agent Skills Directory

[PROMPT_INJECTION]: The skill aggregates and processes significant volumes of untrusted data from external sources (e.g., meeting transcripts, emails, and Slack messages), creating a vulnerability surface for indirect prompt injection. Maliciously crafted content within these ingested files could attempt to influence the agent's behavior during query or compilation tasks.
Ingestion points: Data enters the vault context through directories such as meetings/, raw/, customer-language/, recurring-questions/, and sales-objections/ via manual entry or automated synchronization from third-party services (e.g., Fathom, Gong, HubSpot).
Boundary markers: The skill uses metadata frontmatter (e.g., trust: unreviewed, sensitivity: internal) to categorize content, which provides context but does not serve as a definitive security boundary against adversarial content.
Capability inventory: The skill possesses capabilities for file system read/write operations, directory management, and integration with external tools like git and pandoc for version control and document rendering.
Sanitization: There is no explicit documentation of content sanitization or filtering logic applied to the body of ingested documents before they are parsed by the language model.
[COMMAND_EXECUTION]: The documentation in references/vault-config.md provides instructions for users to modify their shell environment configuration (~/.zshenv) to persist the COMPANY_BRAIN_VAULT environment variable. While common for tool configuration, this involves modifying persistent shell profile files.

company-brain