The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents a pattern for compiling and executing generated code at runtime in references/ai-generated-video.md. It provides a React hook, useCompiledComposition, which uses new Function to execute code transpiled by @babel/standalone. While the documentation includes security warnings regarding sandboxing and CSP, the pattern itself allows for the execution of dynamically generated content.
[INDIRECT_PROMPT_INJECTION]: The skill documents workflows for ingesting untrusted data that could influence agent behavior.
Ingestion points: User prompts are passed to LLMs to generate code (references/ai-generated-video.md), and calculateMetadata functions frequently fetch external JSON data via fetch (references/compositions-props.md).
Boundary markers: The documentation suggests using structured output (Zod schemas) and providing specific system prompts for code generation, but does not enforce strict delimiters in the provided code examples.
Capability inventory: The skill uses bunx to execute CLI tools, supports network operations through fetch, and can write arbitrary files to the disk using the <Artifact> component.
Sanitization: The skill mentions stripping Markdown fences and trailing prose from AI outputs, but largely relies on the user to implement secure sandboxing.
[COMMAND_EXECUTION]: The skill heavily relies on and documents the execution of shell commands using bunx and remotion CLI for project scaffolding, rendering, and version management across various files (e.g., SKILL.md, references/rendering-ssr.md).

remotion