The Agent Skills Directory

[PROMPT_INJECTION]: The skill's operation involves navigating to and reading content from external web applications, which creates a surface for indirect prompt injection where malicious instructions in the web UI could potentially influence the agent's behavior.
Ingestion points: The agent reads web application state using take_snapshot and browser_snapshot tools as specified in SKILL.md and references/overlay.md.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to distinguish between its primary instructions and content parsed from the target web applications.
Capability inventory: The agent possesses capabilities to execute browser scripts, click UI elements, write markdown and image files to the local repository, and manage git worktrees.
Sanitization: The skill does not define any sanitization or validation steps for the content ingested from the web applications before it is processed.
[COMMAND_EXECUTION]: The skill employs system commands and browser script execution to perform its tasks.
Git Worktrees: Utilizes git worktree add to create isolated environments, preventing file collisions during parallel documentation tasks.
Dynamic Script Injection: Uses evaluate_script (Chrome DevTools) and browser_evaluate (Playwright) to inject JavaScript templates into the browser's DOM for the purpose of rendering red-box highlights on elements.

guidewright-capture