The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute several shell utilities, including ffmpeg, whisper-cli, nlm, and pdftoppm. These commands are invoked with arguments derived from user input, such as video file paths and project titles (e.g., ffmpeg -i "<视频路径>" and nlm notebook create "<标题>PPT"). This patterns presents a risk of command injection if the input strings are not properly sanitized before execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple external sources to generate its final output.
Ingestion points: Content fetched from external URLs (via Jina Reader or Agent Reach), audio transcriptions from user-provided video files, and raw text inputs.
Boundary markers: The skill does not define specific boundary markers or use instructions to help the agent distinguish between its core logic and potentially malicious instructions embedded in the ingested data.
Capability inventory: The skill possesses significant capabilities, including shell command execution, network access via curl and nlm, and file system write access for saving scripts and logs.
Sanitization: There is no mechanism described for sanitizing or filtering the ingested external content before it is processed by the language model.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to interact with the SiliconFlow API (api.siliconflow.cn) for speech-to-text and Google's NotebookLM via the nlm CLI. These are well-known services used appropriately for the skill's stated purpose.
[DYNAMIC_EXECUTION]: The skill involves the runtime invocation of external binaries and local Python scripts (such as clean_vtt.py) to process data and format files during its execution phases.

laohan-chuangzuo