The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The 'Carousel Growth Engine' specialist (references/marketing-carousel-growth-engine.md) creates a vulnerability surface by ingesting untrusted data and exercising high-impact capabilities autonomously.
Ingestion points: The agent uses Playwright (analyze-web.js) to scrape arbitrary website URLs to extract content, features, and brand information.
Boundary markers: The instructions lack delimiters or warnings to ignore potential instructions embedded within the target websites.
Capability inventory: The skill allows the agent to generate images via Gemini, execute shell scripts (generate-slides.sh, publish-carousel.sh), and publish directly to TikTok and Instagram via an external API (api.upload-post.com).
Sanitization: There are no instructions for sanitizing or filtering content extracted from external websites before it is interpolated into prompts for image generation or captions.
Risk factor: The skill explicitly instructs the agent to operate with 'Zero Confirmation,' running the entire pipeline without user approval between steps and only notifying the user once publishing is complete. This lack of oversight makes the agent susceptible to following malicious instructions hidden on scraped websites.
[COMMAND_EXECUTION]: The skill instructs the agent to execute various local shell, Node.js, and Python scripts to perform its workflow.
Evidence: SKILL.md and references/marketing-carousel-growth-engine.md refer to executing analyze-web.js, generate-slides.sh, and publish-carousel.sh using tools like uv and playwright.
[EXTERNAL_DOWNLOADS]: The skill requires downloading external software and interacting with remote APIs.
Evidence: Instructions include running playwright install chromium to fetch the browser binary. It also performs network operations to api.upload-post.com for publishing content and fetching analytics.

marketing