The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/scrape-douyin.py downloads video files directly from Douyin's CDN (e.g., douyinvod.com) using urllib.request.urlopen and writes them to a local task-specific directory for analysis.
[COMMAND_EXECUTION]: The skill makes extensive use of subprocesses to run system tools. scripts/scrape-douyin.py executes ffmpeg to generate contact sheets from downloaded videos and mmx vision to generate descriptions. scripts/process-lark-douyin.mjs uses lark-cli to interact with Lark Base and python3 to invoke the scraper script.
[INDIRECT_PROMPT_INJECTION]: The skill represents an indirect prompt injection surface as it ingests untrusted text and visual content from external Douyin posts. This data is processed, summarized by a vision model, and then written into a structured format (Lark Base) and returned to the agent. A malicious actor could embed instructions in post captions or video frames to influence the agent's behavior during content processing.
Ingestion points: scripts/scrape-douyin.py (Douyin video pages and metadata).
Boundary markers: The skill uses a simple text header 【视频内容解析】 to separate vision analysis from the original caption, but lacks robust delimiters to prevent instruction interpretation.
Capability inventory: Subprocess execution (ffmpeg, mmx, lark-cli), file system writes (/tmp/douyin-scraper), and network access via Playwright and urllib.
Sanitization: Limited to basic regex-based cleaning of titles and engagement count normalization.
[CREDENTIALS_UNSAFE]: The skill accesses sensitive local files to maintain state and authentication, specifically ~/.config/video-download/douyin_cookies.json and browser user-data directories located at $HOME/Library/Application Support/Google/DouyinChrome. While intended for legitimate browser automation, this provides a mechanism for the agent to interact with authenticated sessions.

douyin-scraper