The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes natural language input (user briefs) to generate design specifications and AI prompts. This creates a surface for indirect prompt injection where malicious instructions could be embedded in the intake data.
Ingestion points: references/site-flow.md and references/app-flow.md process user-provided briefs and intake batches.
Boundary markers: The workflow uses intermediate files like DESIGN-SPEC.md to structure the agent's tasks, though it lacks explicit delimiters for raw untrusted data.
Capability inventory: The agent has the ability to write files to the repository and execute shell commands for media processing.
Sanitization: The instructions recommend runtime validation using ArkType or zod for API request bodies.
[EXTERNAL_DOWNLOADS]: The skill is designed to download media assets from the fal.ai content delivery network.
Evidence: SKILL.md mandates downloading all generated assets (images, video, audio) from temporary fal.ai URLs into the local public/generated/ folder to ensure repository self-sufficiency.
[COMMAND_EXECUTION]: The skill employs local command-line utilities for video processing tasks.
Evidence: references/scroll-scrub.md provides ffmpeg command strings for re-encoding video for scroll-based animation and generating frame sequences.
[DATA_EXFILTRATION]: The skill facilitates communication with official fal.ai endpoints for media generation.
Evidence: The skill makes network requests to fal.run and queue.fal.run to generate content.
The instructions prioritize security by requiring the FAL_KEY to be handled only in server-side environment variables and never exposed in client-side bundles or public code.

fal-studio