The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/package-loader.mjs script dynamically loads Node.js modules using the import() statement from paths computed at runtime, specifically targeting packages installed via npm. This allows the execution of code not strictly bundled with the skill.
[COMMAND_EXECUTION]: The skill executes shell commands via node:child_process (spawnSync) within scripts/package-loader.mjs to perform package installations (npm install). It also suggests the execution of various CLI tools for video processing and linting in SKILL.md.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to the npm registry to install dependencies like @hyperframes/producer and sharp. It also references external CDNs (jsdelivr, cdnjs) for loading libraries such as GSAP, Three.js, and Lottie in the generated HTML compositions.
[PROMPT_INJECTION]: The skill includes instructions that attempt to gate agent behavior, such as <HARD-GATE> blocks in SKILL.md which mandate specific design checks before proceeding with HTML generation. These are used as internal control mechanisms rather than adversarial bypasses.
[INDIRECT_PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection by processing external untrusted data.
Ingestion points: Data enters the context through transcript.json, frame.md, design.md, and audio data files.
Boundary markers: Absent. The instructions do not define delimiters to separate or neutralize instructions that may be contained within these data files.
Capability inventory: The skill can execute shell commands (spawnSync in package-loader.mjs), write to the file system (e.g., .hyperframes/expanded-prompt.md), and initiate network downloads.
Sanitization: Mostly absent. While layout-specific validation exists (e.g., fitTextFontSize), there is no sanitization to prevent the agent from interpreting ingested text as new instructions.

hyperframes