The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of user-provided Markdown and MDX content from local files and remote CMS sources (Notion, Sanity, Contentful, Payload, and Strapi). This content is interpolated into the agent's context and processed during project build, dev, and migration tasks, representing an indirect prompt injection surface.
Ingestion points: Local content root (default docs/) and remote content sources configured in blume.config.ts via adapters.
Boundary markers: Absent; the framework treats project content as authoritative input for generating documentation.
Capability inventory: The skill utilizes shell commands for project management (blume build, blume dev), performs network operations to fetch remote content, and generates executable code (Vite/Astro build process).
Sanitization: The documentation explicitly warns that enabling MDX for certain sources (such as Notion) allow for code execution at build time and advises enabling it only for trusted editors.
[COMMAND_EXECUTION]: The skill involves the execution of various development commands using the blume CLI (init, dev, build, doctor, validate, audit, etc.). These are necessary for the skill's primary purpose of managing a documentation project and are standard development workflows.
[EXTERNAL_DOWNLOADS]: The instructions guide the agent and user to install numerous peer dependencies from the npm registry, including official Astro integrations, AI SDKs, and CMS clients. These downloads target well-known and reputable organizations such as Vercel, Google, Anthropic, and the Astro community.

blume