The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to clone an external GitHub repository (https://github.com/GargantuaX/gemini-watermark-remover.git) and execute its scripts using Node.js to remove video watermarks. This involves downloading and running code from an unverified third-party source, which is a critical security risk.
Evidence: references/postprocess-sop.md contains the commands: git clone https://github.com/GargantuaX/gemini-watermark-remover.git, pnpm install, and node bin/gwr.mjs remove input.mp4.
[EXTERNAL_DOWNLOADS]: The skill utilizes pnpm dlx to install external agent skills from unverified developers and downloads several platform-specific tools.
Evidence: references/postprocess-sop.md suggests pnpm dlx skills add GargantuaX/gemini-watermark-remover.
Evidence: references/setup.md installs @larksuite/cli from the NPM registry.
[COMMAND_EXECUTION]: The skill requires extensive use of shell commands to control local and remote tools, including ffmpeg, ffprobe, git, and pnpm. These commands are used to process video files and manage dependencies.
Evidence: SKILL.md and references/setup.md specify checks and installation steps for these CLI tools.
[INDIRECT_PROMPT_INJECTION]: The workflow is susceptible to indirect prompt injection because it ingests data from Feishu (Lark) tables—such as product descriptions and rework comments—and interpolates them directly into prompts for large language models (LLMs) without boundary markers or sanitization.
Ingestion points: Data enters the context via lark-cli reading fields like 验收意见 (acceptance opinion) and 备注 (remarks) from the Feishu Base.
Boundary markers: The prompt templates in references/prompt-templates.md lack explicit delimiters or instructions to ignore potentially malicious embedded content within the ingested fields.
Capability inventory: The skill has significant capabilities including browser automation (chrome:control-chrome), shell execution, and the ability to upload/download files.
Sanitization: There is no evidence of filtering, escaping, or validating the external content before it is processed by the AI models.

tiktok-gemini-video-workflow