The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external URLs via scripts/harvest.py, which is stored in harvest.json. This content is then read by the AI agent to generate a storyboard.json that controls the video rendering engine.
Ingestion points: scripts/harvest.py fetches article text, headings, and image captions from user-provided URLs.
Boundary markers: The instructions do not include explicit delimiters or warnings to the AI to treat the harvested content as potentially adversarial data.
Capability inventory: The skill possesses significant capabilities, including shell command execution (remotion, npm, curl, agent-browser, sips) and file system writes.
Sanitization: While the generated storyboard.json is validated against a Zod schema, this validation ensures structural integrity but does not sanitize the natural language content that influences the agent's logic.
[EXTERNAL_DOWNLOADS]: The skill downloads font assets from a well-known repository during the rendering phase.
Evidence: scripts/build_fonts.mjs uses curl to fetch Noto Sans and Serif CJK variable fonts from github.com/notofonts/noto-cjk.
Context: These are official Google-maintained font repositories, and the downloads are necessary for the skill's primary function of rendering Chinese typography.
[COMMAND_EXECUTION]: The skill executes several CLI tools to process media and perform rendering.
Evidence: scripts/harvest.py and scripts/render.mjs use subprocess.run and spawnSync to invoke bun, agent-browser, sips, remotion, esbuild, and npm.
Context: These executions are part of the intended workflow. The scripts pass arguments as lists rather than raw strings, which mitigates standard shell injection risks.

html2video