The Agent Skills Directory

[DYNAMIC_EXECUTION]: In scripts/fetch_note.py, the transcription model is initialized using funasr.AutoModel with the trust_remote_code=True parameter. This setting permits the library to execute arbitrary Python code contained within the downloaded model files from the remote repository.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted content from the web and passes it to an LLM without sufficient isolation.
Ingestion points: Raw text content is scraped from external Xiaohongshu notes in scripts/fetch_note.py.
Boundary markers: The LLM prompt in scripts/analyze_hook.py interpolates the scraped content directly into the user message without the use of delimiters or instructions to ignore embedded commands.
Capability inventory: The skill has permissions to write files to the local disk, execute system commands via ffmpeg, and perform network requests to the DeepSeek API.
Sanitization: No validation or sanitization is performed on the scraped content before it is processed by the AI agent.
[REMOTE_CODE_EXECUTION]: The use of trust_remote_code=True facilitates the download and execution of code from the ModelScope model registry, which represents a remote code execution vector if the source repository were to be compromised.
[COMMAND_EXECUTION]: The script scripts/fetch_note.py uses subprocess.run to call the ffmpeg binary. While used for processing video files, this pattern executes a system command on externally sourced data.
[EXTERNAL_DOWNLOADS]: The skill automatically downloads data, images, and video assets from Xiaohongshu domains (xiaohongshu.com, xhslink.com). It also downloads machine learning models from the ModelScope service.
[DATA_EXFILTRATION]: Scraped content and metadata are transmitted to the DeepSeek API (api.deepseek.com) for content analysis. This is part of the skill's core functionality but involves sending data to an external third-party service.

xiaohongshu-ingest