The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/install-bridge.sh script clones a third-party repository from GitHub (https://github.com/samuelgursky/davinci-resolve-mcp) and installs its requirements via pip. This source is not recognized as a trusted organization or vendor.
[DYNAMIC_EXECUTION]: The bridge/claude_bridge.lua script runs inside DaVinci Resolve and uses dofile() to execute Lua code blocks generated at runtime by the Python client (bridge/resolve_lua_bridge.py). This architecture allows the execution of arbitrary logic within the Resolve environment.
[DYNAMIC_EXECUTION]: Multiple scripts (scripts/add-media.py and scripts/build.py) dynamically modify the Python sys.path to include directories for importing the bridge modules based on environment-specific paths.
[EXTERNAL_DOWNLOADS]: The scripts/add-media.py utility utilizes yt-dlp to download content from external URLs (such as YouTube), which introduces potential risks from untrusted web content.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by processing external data:
Ingestion points: Media metadata from URLs (add-media.py), transcripts (ref-cuts.py), and local project specification files (plan-spec.py).
Boundary markers: The skill lacks explicit delimiters or warnings to prevent the agent from interpreting instructions embedded within these data sources.
Capability inventory: The agent can perform shell command execution (ffmpeg, yt-dlp) and interface with a professional video editing suite via a custom Lua bridge.
Sanitization: While the bridge escapes data for Lua literals, there is no validation to prevent malicious instructional content from influencing the agent's behavior.

davinci-resolve