The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external URLs, user-provided files, and text for media generation.
Ingestion points: transcribe, upload_asset, render_html_video, and various generation prompts in SKILL.md.
Boundary markers: Extensive use of tokens (e.g., @Image1, <IMAGE_REF_0>) and quoting for scripts to delimit data from instructions.
Capability inventory: File writing (saving assets), network fetching of media, and command execution (via the MCP server's rendering and generation tools).
Sanitization: The skill instructs the agent to treat example code as untrusted and to verify outputs.
[DYNAMIC_EXECUTION]: The skill supports rendering HTML, CSS, and JavaScript into images or video through specialized tools.
Evidence: references/tool-catalog.md and references/workflows/examples.md describe render_html_image and render_html_video tools.
Safety guidance: The skill explicitly warns that example code is "untrusted data" and that loading it does not authorize execution without review.

creativeclaw