The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands and the Python subprocess module to execute external binaries including ffmpeg, ffprobe, sips, sed, and mv for media processing and installation tasks.
The Python scripts scripts/process_video_bg.py, examples/Карусель Архетипы/render_animations.py, and record_video.py use subprocess.run to call ffmpeg and ffprobe for video cross-cropping, trimming, and encoding.
Instructions in SKILL.md and workflows/render_format_A.md command the agent to run sips -Z 2000 to resize images to avoid platform session crashes.
ONBOARDING.md contains a complex grep and sed pipeline to perform path replacements across the skill's files during installation.
workflows/handover.md includes shell commands using mv to organize final post assets and move drafting scripts to a subfolder.
[EXTERNAL_DOWNLOADS]: The skill interacts with external services to fetch media and process content from the web.
It uses curl to download AI-generated images from Higgsfield URLs and to upload user-provided media.
workflows/onboarding.md describes fetching data from YouTube, website articles, and URLs to provide the raw text for carousel generation.
[DYNAMIC_EXECUTION]: The skill generates executable HTML and Python content at runtime based on user data.
examples/Карусель Архетипы/generate_slides.py dynamically builds localized HTML files for Playwright rendering by interpolating post content into templates.
workflows/render_format_A.md instructs the agent to create a project-specific generate_slides.py script for each new carousel, which is then executed to produce the final assets.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data and possesses high-privilege capabilities, creating a significant attack surface.
Ingestion points: The skill is designed to ingest and summarize external data from YouTube transcripts, website articles, and user-provided text/images as specified in workflows/onboarding.md and SKILL.md.
Boundary markers: The instructions lack specific delimiters or "ignore embedded instructions" warnings for the ingested content, increasing the risk that the LLM may follow malicious commands embedded in external articles.
Capability inventory: The skill environment includes powerful tools such as shell access (ffmpeg, sips, sed), headless browser automation (playwright), network utilities (curl), and Python execution.
Sanitization: There is no evidence of sanitization or strict schema validation for external data before it is interpolated into HTML templates or passed to rendering scripts.

carousel