The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local utility scripts and media processing tools (ffmpeg, ffprobe) using secure subprocess patterns. It utilizes list-based arguments instead of shell execution, and all input paths are strictly validated using pathlib.Path.resolve() and symlink checks in scripts/extract_keyframes.py to prevent directory traversal or command injection.
[PROMPT_INJECTION]: The skill incorporates comprehensive defense-in-depth against indirect prompt injection. All data ingested from external platforms via the computer-use tool is explicitly marked as untrusted in references/evidence-schema.md. Boundary markers are implemented through deterministic validation scripts (validate_evidence.py), which ensure external content like transcripts and OCR results are handled as inert data and never executed as instructions.
[DATA_EXFILTRATION]: Data management is restricted to a user-defined vault root (<NOTES_VAULT>). The skill mandates path normalization and ensures that all writes occur within authorized platform-specific subdirectories (e.g., <NOTES_VAULT>/AI Wiki/raw/调研/爆款拆解/). It prohibits guessing paths or accessing sensitive system directories, and validates that output files remain within the run directory.
[REMOTE_CODE_EXECUTION]: The skill uses a segmented architecture where the main agent handles data collection and tool invocation, while a separate 'sub-agent' performs analysis in a restricted context. This prevents the analysis of untrusted content (titles, body text, comments) from leading to unauthorized tool usage or dynamic code execution.

viral-video-benchmark