The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an indirect prompt injection surface by scraping external listing URLs via Firecrawl and interpolating the results directly into script generation prompts for Gemini AI.
Ingestion points: The listingUrl provided in the workflow request is processed by the scrapeProperty function in rules/url-to-video.md.
Boundary markers: The prompt templates SHOWCASE_PROMPT and SOCIAL_PROMPT lack explicit boundary markers or delimiters (like XML tags or 'ignore instructions' warnings) when injecting the JSON-stringified property data.
Capability inventory: The skill possesses significant capabilities including file writing (Write), shell access (Bash), and network operations via webhooks and external SDKs.
Sanitization: While the code references validatePropertyData, the logic is not provided, and the AI output is directly parsed as JSON, which can be manipulated if the source listing contains malicious instructions.
[DATA_EXFILTRATION]: The workflow includes a webhook notification system that sends POST requests to a webhookUrl. If this URL is user-controlled and not restricted, it could be used for Server-Side Request Forgery (SSRF) to interact with internal services or exfiltrate metadata.
[EXTERNAL_DOWNLOADS]: The implementation relies on standard SDKs from well-known technology providers for its core pipeline, including @google/generative-ai for scripting, openai for text-to-speech, and @remotion/lambda for video rendering. These are documented as trusted integrations within the skill's context.

real-estate-workflows