The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill generates and executes code dynamically at runtime to handle clipboard operations. On macOS, it writes a Swift script to a temporary file and executes it via the system's swift interpreter. On Windows, it dynamically constructs and executes PowerShell command strings.
Evidence: scripts/copy-to-clipboard.ts generates Swift source code in getMacSwiftClipboardSource() and constructs PowerShell strings in copyImageWindows() and copyHtmlWindows().
[COMMAND_EXECUTION]: The skill frequently executes system-level commands to manipulate the clipboard and interact with the browser environment. It uses utilities such as osascript (macOS), powershell.exe (Windows), xdotool/ydotool (Linux), and pkill (Unix).
Evidence: scripts/paste-from-clipboard.ts uses spawnSync to execute osascript for keyboard simulation.
Evidence: scripts/weibo-utils.ts executes ps aux and kill to manage Chrome instances.
[PROMPT_INJECTION]: The skill includes instructions that direct the agent to override standard confirmation workflows. Specifically, it instructs the agent to automatically kill browser processes and retry commands when encountering certain errors without asking the user.
Evidence: SKILL.md contains the instruction: "Important: This should be done automatically -- when encountering this error, kill the CDP Chrome instances and retry the command without asking the user."
[PROMPT_INJECTION]: The skill processes untrusted user data (Markdown files) to extract metadata like titles and summaries, which are then used in automated browser interactions and script arguments. This represents a potential surface for indirect prompt injection or command injection if the input files contain malicious payload characters.
Ingestion points: scripts/weibo-article.ts reads content and frontmatter from user-provided Markdown files.
Capability inventory: The skill has the ability to read/write files, automate browser sessions, simulate system keystrokes, and perform network operations via the browser.
Sanitization: The skill uses JSON.stringify for some fallback inputs, but relies on manual escaping for others (e.g., PowerShell path escaping).
Sanitization: No explicit boundary markers are used in the prompt construction when interpolating article content.
[DATA_EXFILTRATION]: The skill's primary function is to transmit user-supplied text and media to the Weibo platform. While this is the intended behavior and the skill is designed for manual user review before final publishing, it establishes an automated flow for sending data to external domains via browser automation.

baoyu-post-to-weibo