The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a framework architecture that aggregates content from multiple external sources, including Notion, Contentful, and remote GitHub repositories, to expose this data to AI agents via llms.txt and Model Context Protocol (MCP) servers. This creates a surface where an attacker controlling one of these external sources could inject instructions intended to influence the behavior of an agent reading the documentation.
Ingestion points: Content sources described in references/sources.md such as mdxRemote, notion, sanity, contentful, payload, and strapi.
Boundary markers: The framework includes a <Visibility> component for scoping content to agents, but the documentation lacks specific guidance on sanitizing or filtering instructions that might be embedded in data fetched from these external CMS platforms.
Capability inventory: The documented framework's CLI (blume) can execute build processes, perform automated translations using agents, and conduct network operations to fetch source content.
Sanitization: Documentation in references/sources.md warns that remote HTML and MDX are rendered as-is and advises users to only ingest content from trusted repositories.
[EXTERNAL_DOWNLOADS]: The skill documents legitimate framework features that fetch remote content from well-known services (e.g., GitHub via mdxRemote) during the site build process. These operations are neutral and consistent with the framework's primary purpose.

blume-docs