The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted design data from the Stitch platform to automatically generate video annotations and callouts. Malicious instructions embedded within the design's HTML or metadata could potentially influence the agent's behavior during the extraction process.\n
Ingestion points: Fetches project metadata and HTML source code via the stitch:get_screen tool (SKILL.md).\n
Boundary markers: There are no explicit instructions or delimiters mentioned to isolate external design text from the agent's internal logic.\n
Capability inventory: The skill possesses the ability to execute shell commands via Bash and write files to the local file system (SKILL.md).\n
Sanitization: The instructions do not specify any validation or sanitization steps for the text extracted from the external HTML code before it is used in the video composition.\n- [COMMAND_EXECUTION]: The skill uses local shell scripts (scripts/download-stitch-asset.sh) and command-line tools like npm to manage the video project lifecycle, including environment setup and asset organization. These commands are standard for this type of automated workflow.\n- [EXTERNAL_DOWNLOADS]: Asset retrieval involves downloading screenshots from external storage (e.g., Google Cloud Storage) using curl. These operations target the vendor's expected infrastructure and are limited to downloading visual assets.\n- [DYNAMIC_EXECUTION]: The skill dynamically generates and writes React/TypeScript components (WalkthroughComposition.tsx, ScreenSlide.tsx) based on provided templates and screen manifests. These components are subsequently rendered into a video file by the Remotion framework.

remotion