The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions direct the agent to execute npx @getvhs/vhscli@latest, which downloads the latest version of the tool from the npm registry at runtime. This package is managed by the skill's author (getvhs).
[COMMAND_EXECUTION]: The skill operates by invoking the vhscli tool through shell commands, passing user prompts and file paths as arguments.
[DATA_EXFILTRATION]: The tool is designed to upload local media files (images, videos, and PDFs) to a remote server for processing by AI models. This behavior is the primary intended function of the skill and is clearly documented.
[CREDENTIALS_UNSAFE]: The tool manages user authentication by storing session tokens in a local file at ~/.vhs/session.json following a login command. This is standard session persistence for a command-line utility.
[PROMPT_INJECTION]: The skill ingests and processes untrusted external data (images, PDFs, and videos), which constitutes a surface for indirect prompt injection attacks.
Ingestion points: The agent passes paths for user-provided images (-i), PDFs (-f), and videos (-v) to the chat and generate commands.
Boundary markers: The skill does not define specific boundary markers for the file content, although the prompt guides recommend requesting structured JSON output to improve reliability.
Capability inventory: The skill utilizes shell command execution, file system access, and network communication.
Sanitization: No explicit sanitization or content filtering for the processed media is mentioned in the instructions.

vhscli