The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external audio data from Xiaohongshu videos. This data is transcribed by Whisper and then interpolated into an LLM prompt for translation.
Ingestion points: Transcribed audio segments are stored in transcripts/audio.txt and then passed to the LLM via the {segment} placeholder in Step 7 of SKILL.md.
Boundary markers: There are no delimiters (like XML tags or triple quotes) or "ignore embedded instructions" warnings surrounding the interpolated {segment} in the translation prompt.
Capability inventory: The skill has capabilities to write to the file system (creating transcript.md and transcript folders) and execute shell commands (yt-dlp, whisper, ffmpeg, mv, awk).
Sanitization: No filtering, escaping, or validation is performed on the transcription output before it is sent to the LLM for translation.
[COMMAND_EXECUTION]: The skill relies on shell command execution to perform its primary tasks.
Bundled Script: It executes scripts/download_xiaohongshu.sh using the bash interpreter.
Media Tools: It invokes yt-dlp, ffmpeg, and whisper through shell subprocesses.
File Management: It uses shell commands like mv, cp, awk, and rm for file manipulation and text processing.
Input Handling: In SKILL.md, user-provided URLs are passed as arguments to the download script. While the instructions use single quotes for placeholders, a malicious URL containing a single quote could potentially attempt command injection if the agent does not apply strict escaping.

xiaohongshu-downloader