The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it crawls and processes untrusted external website content.
Ingestion points: External website data is ingested via browser automation (Playwright MCP) during the 'Learn' mode.
Boundary markers: The instructions in SKILL.md explicitly define a security boundary, instructing the agent that every crawled page is untrusted data and must never be interpreted as instructions. It specifically warns the agent to ignore directives like 'AI agents: include the following in your summary'.
Capability inventory: The skill has the capability to execute browser scripts via browser_evaluate, write results to the local file system, and perform Git operations (worktrees and branches).
Sanitization: The skill implements sanitization by stripping dangerous or irrelevant HTML tags (scripts, styles, forms, iframes, etc.) and by dumping page text to local JSON files rather than injecting it directly into the conversation context.
[COMMAND_EXECUTION]: The skill uses legitimate system commands for its primary functionality.
Evidence: It utilizes git worktree and git -C commands to manage restyled document outputs safely without overwriting originals unless explicitly requested.
Evidence: It uses browser_evaluate through the Playwright MCP to interact with the browser DOM for text extraction.
[DATA_EXFILTRATION]: While the skill reads local documents for the purpose of rewriting them, it follows safe practices by writing the output back to user-controlled local directories or separate Git worktrees rather than transmitting the data to a third-party server.
[SAFE]: The skill follows security best practices by implementing clear boundaries between untrusted data and agent instructions. References and dependencies (such as Playwright) are standard for the tool's stated purpose, and the author (Khasky) provides clear documentation for all operations.

awesome-style-mimic