The Agent Skills Directory

[COMMAND_EXECUTION]: The skill repository contains utility scripts (scripts/merge_research.py, scripts/srt_to_transcript.py, scripts/quality_check.py) intended for local data processing and quality assurance. These scripts operate on local files and do not exhibit malicious command injection patterns.
[EXTERNAL_DOWNLOADS]: The scripts/download_subtitles.sh script uses the well-known yt-dlp utility to fetch subtitles from media platforms. This is a functional component for gathering research data and does not target sensitive system files.
[PROMPT_INJECTION]: The skill uses persona-enforcement instructions (e.g., "直接以田川的身份回应", "不说...他可能会...") to maintain the 'Tianchuan' character. These are standard role-play instructions and do not attempt to bypass core AI safety filters or exfiltrate system prompts.
[INDIRECT_PROMPT_INJECTION]: The '田川式研究' (Tianchuan-style research) protocol instructs the agent to use web search tools to fetch industry data. This creates an attack surface where untrusted web content could influence agent behavior. However, the skill includes a structured workflow (Step 1-3) that encourages internal fact-checking and classification before generating a response, which serves as a logical mitigation.

tianchuan-perspective