The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external files (PDF, Office, and images) via the mineru-open-mcp.parse_documents tool to extract content for slide generation. This creates a surface for indirect prompt injection where malicious content in these documents could attempt to override agent instructions.
Ingestion points: External document parsing via mineru-open-mcp.parse_documents and reference_intake.
Boundary markers: The workflow enforces a mandatory deck-protocol.json structured file and a deck-protocol.review.md artifact which requires explicit user confirmation before any image generation begins.
Capability inventory: File system access (reading references, writing protocol and images), subprocess execution (CLI tools via MCP), network operations (API-based document parsing), and subagent spawning.
Sanitization: The skill explicitly forbids rendering internal metadata or source labels in slide text and enforces visual QA gates to check for alignment and defects.
[COMMAND_EXECUTION]: The skill relies on MCP tools that execute local CLI commands (e.g., uvx, node plugins/ppt-composer/src/cli.mjs) for environment auditing (doctor), document parsing, and final PPTX assembly (assemble-image-ppt).
[DYNAMIC_EXECUTION]: For decks exceeding 7 pages, the skill is instructed to use spawn_agent to create parallel subagents for image generation tasks based on the imagegen-jobs.json plan.
[EXTERNAL_DOWNLOADS]: Setup instructions and tool checks involve uvx and npm package prewarming, which download external dependencies to the local environment.

image-first-ppt