The Agent Skills Directory

[COMMAND_EXECUTION]: The skill describes the use of ffmpeg and ffprobe for read-only diagnostics and media metadata extraction from local files. It also mentions using curl to download temporary rendered frames from the service's own signed URLs to a local temporary folder for inspection.
[DYNAMIC_EXECUTION]: Mentions the use of node_repl js to initialize a browser runtime if required by the host environment for browser automation tasks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data sources including video transcripts, project metadata, and asset names. 1. Ingestion points: Untrusted data is ingested via read_project, browse_assets, and manage_transcript tool results in SKILL.md. 2. Boundary markers: No explicit delimiters or boundary markers are defined for the processed data in the instructions. 3. Capability inventory: The skill utilizes ffmpeg, curl, and a suite of MCP tools (mcp__chatcut__*) for file and project manipulation. 4. Sanitization: The instructions recommend shell-quoting URIs for curl calls and explicitly require stripping sensitive editor-boot-token parameters from URLs before displaying them to the user.
[DATA_EXFILTRATION]: The skill contains a defensive instruction to prevent credential leakage by requiring the agent to remove sensitive tokens (editor-boot-token) from internal-browser URLs before presenting them as external links to the user.
[PROMPT_INJECTION]: The skill includes specific instructions on how to handle user communication for authentication failures and project handoffs. These instructions, which were flagged by static detectors as potential concealment, are actually transparency guidelines that ensure the user is informed of specific steps for authentication (e.g., using codex mcp login chatcut) rather than receiving vague error messages.

chatcut-plugin-basics