The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains a 'MANDATORY' instruction block requiring the agent to execute a specific curl command to a local service (localhost:8888) before any other action. This attempt to override agent behavior with fixed side-effects is a form of behavioral injection.
[COMMAND_EXECUTION]: The Tools/Render.ts file utilizes Bun's shell execution utility to run CLI commands like npx remotion render and npx create-video. This allows the agent to execute shell commands within the project environment.
[REMOTE_CODE_EXECUTION]: The Workflows/ContentToAnimation.md workflow dynamically generates a complete React project (including package.json and source files) in a temporary directory and subsequently executes npm install and npx remotion render. Executing code generated at runtime from potentially untrusted inputs poses a significant risk.
[REMOTE_CODE_EXECUTION]: Documentation files frequently instruct the agent to run npx remotion add <package>, which downloads and executes code from the npm registry at runtime during the skill's operation.
[PROMPT_INJECTION]: Indirect injection surface detected in Workflows/ContentToAnimation.md. 1. Ingestion points: YouTube transcripts, Web URLs, PDF files, and Tweets via Parser skill. 2. Boundary markers: Absent. 3. Capability inventory: Shell execution (npx), file writes (/tmp/), and network requests (curl). 4. Sanitization: Absent; content is directly interpolated into generated React source files.

Remotion