The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external AI providers and user-supplied content, creating a surface for indirect prompt injection.
Ingestion points: The skill ingests data from external provider responses (e.g., rewritten prompts, generation receipts) and user-authored prompts or media files (SKILL.md, workflow.md).
Boundary markers: Instructions explicitly direct the agent to "Treat prompts, provider-native requests, receipts, and media as opaque creative artifacts" and require inspection in the "agent/user loop" (SKILL.md).
Capability inventory: The skill executes platform CLI commands (renku generation execute, renku media import) and internal Node.js scripts for materializing HTML visualizations (SKILL.md, scripts/).
Sanitization: The visualization materializer script performs character escaping (e.g., /[<>&  ]/g) on JSON payloads to mitigate injection risks in generated HTML components (scripts/materialize-generation-configuration-visualization.mjs).
[COMMAND_EXECUTION]: The skill relies on executing the renku CLI to perform its core functions, including generating context, validating requests, and importing media.
Evidence: Multiple files document the use of commands such as renku generation execute and renku media import to process and store generated artifacts (SKILL.md, workflow.md).
[DYNAMIC_EXECUTION]: The skill includes Node.js scripts that dynamically generate HTML fragments for the Codex harness's visualization capability.
Evidence: The script materialize-generation-configuration-visualization.mjs reads a template and a JSON payload, performs replacement, and writes a temporary HTML file for rendering (scripts/materialize-generation-configuration-visualization.mjs).

media-producer