The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses and utilizes sensitive local authentication data to perform its tasks.
Evidence: The instructions define and check for ~/cookies.json (Step 2) to authenticate requests to the Xiaohongshu platform. The skill subsequently uses the Bash tool to make network requests using these credentials.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted content from external web pages, creating a potential surface for malicious instructions to influence the agent.
Ingestion points: Content is fetched from Xiaohongshu links provided in $ARGUMENTS. The agent is instructed to parse __INITIAL_STATE__ and transcribe video audio.
Boundary markers: No delimiters or specific instructions are provided to the agent to distinguish between the fetched data and its operational instructions.
Capability inventory: The skill has permissions for Bash (network and command execution) and Write (file system modification).
Sanitization: There is no requirement for sanitizing, filtering, or validating the external content before it is processed by the AI for organization and summarization.
[EXTERNAL_DOWNLOADS]: The skill references an external AI model for its transcription tasks.
Evidence: The skill defines the Whisper model mlx-community/whisper-large-v3-turbo, which typically involves downloading large model weights from Hugging Face at runtime.

xhs-batch