The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestrator in SKILL.md (Step 3 and Step 5) executes Python code via PowerShell by interpolating user-supplied strings directly into the script body. For example, the use of Path(r"<VIDEO PATH>") and <STATUS> allows for potential Python injection if the user-provided path or metadata contains quote terminators or escape characters. The skill does not implement sanitization or validation of these inputs before interpolation.
[EXTERNAL_DOWNLOADS]: The setup-video-pipeline.ps1 script automates the download of several components from remote sources. This includes ffmpeg via winget, the Python 3.11 runtime via the Python Launcher, and large Machine Learning models (Whisper and pyannote) from HuggingFace. These downloads target well-known and reputable services.
[CREDENTIALS_UNSAFE]: The installation script requests a HuggingFace API token from the user and persists it in the Windows Credential Manager using the cmdkey utility. While this is a standard and relatively secure practice for local CLI tools, the skill involves the agent directly handling and managing third-party authentication secrets.
[PROMPT_INJECTION]: The skill processes untrusted external video files to generate transcripts and summaries. This creates an indirect prompt injection surface; if a video contains adversarial spoken content or metadata designed to override agent behavior, it could influence downstream qualitative specialists or analysts that consume the resulting manifest and artifacts.
[DATA_EXFILTRATION]: The skill implements a telemetry system that logs operational metrics and user feedback to a local JSONL file at ~/.claude/skills/video-content-analysis/telemetry.jsonl. While no remote exfiltration of this data was observed in the provided code, it maintains a persistent local audit trail of skill usage and participant-related metadata.

video-content-analysis