The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on shell commands like ffmpeg, say, and ffprobe to process media and generate text-to-speech. In SKILL.md, the command ffmpeg -i "<视频路径>" ... incorporates user-provided video paths. This pattern requires the agent to strictly sanitize paths to prevent shell injection vulnerabilities.\n- [DYNAMIC_EXECUTION]: The validation script scripts/check-script-contract.mjs determines execution paths for sub-validators using environment variables such as LAOHAN_SERIES_RESEARCH_VALIDATOR. It subsequently executes these paths using node via spawnSync. This design allows for the execution of arbitrary code if the environment variables are manipulated.\n- [EXTERNAL_DOWNLOADS]: The skill instructions specify the installation of external dependencies including notebooklm-mcp-cli via pip and system utilities like ffmpeg, poppler, and whisper.cpp via brew. Additionally, it utilizes the SiliconFlow API (api.siliconflow.cn) for audio transcription, which involves transmitting local audio data to a remote third-party service.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from URLs and video transcripts.\n
Ingestion points: External URLs scraped via Jina Reader/agent-reach and transcriptions of user-provided videos.\n
Boundary markers: The workflow uses Markdown structure and JSON "contracts" for data integrity, but lacks specific defensive instructions to ignore malicious prompts embedded within external content.\n
Capability inventory: File system access, network exfiltration/scraping, and shell command execution (ffmpeg, say).\n
Sanitization: Shell commands use quoting for paths, but external text content is summarized and structured without dedicated filters for LLM instructions.

laohan-chuangzuo