The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation in references/sau-adapter.md provides explicit instructions for the user to clone and install a third-party tool from a non-trusted GitHub repository (dreammis/social-auto-upload). The installation process involves downloading external browser binaries and multiple dependencies from unverified sources.
[COMMAND_EXECUTION]: The helper script scripts/publish.mjs executes the external sau tool using the node:child_process.spawn method to perform video uploads.
[DYNAMIC_EXECUTION]: The scripts/publish.mjs script allows the path to the sau binary to be defined via a command-line argument (--sau). This allows the agent or a user to point the script toward any executable on the system, which is then executed with arguments parsed from the project files.
[INDIRECT_PROMPT_INJECTION]: The skill parses video and thumbnail file paths directly from Markdown files in the user's project directory and passes them as arguments to the upload command.
Ingestion points: Markdown files for each platform (e.g., 抖音.md, youtube.md) located within the path provided by the --episode argument in scripts/publish.mjs.
Boundary markers: The SKILL.md file mandates a "List execution plan" step and "Get per-platform authorization," and the publish.mjs script requires manual confirmation ((y/N)) before processing each platform.
Capability inventory: The skill has the ability to read files (node:fs/promises.readFile), write to files (appendFile, writeFile), and execute external processes (node:child_process.spawn).
Sanitization: The script uses path.resolve to handle paths extracted from the markdown content (via fieldPath in scripts/publish.mjs) but does not validate that the paths point to legitimate media files, potentially allowing an attacker to specify sensitive system files for upload.

self-media-video-publisher