The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from a locally running web server through screenshots, which can act as a vector for indirect prompt injection.
Ingestion points: The agent captures and inspects images of the local Vite server (http://localhost:5173) using the browser_take_screenshot tool, as described in SKILL.md.
Boundary markers: Absent. There are no instructions for the agent to ignore text or instructions that might be embedded in the rendered UI being analyzed.
Capability inventory: The skill includes a subprocess.run wrapper in scripts/_skill_cli.py and has the capability to modify project code files.
Sanitization: Absent. The visual output is interpreted directly by the agent's vision model without validation of the content source.
[EXTERNAL_DOWNLOADS]: The skill utilizes the @playwright/mcp tool via npx, which dynamically downloads and executes code from the NPM registry. This is a standard operation for using official tools from well-known technology providers.
[COMMAND_EXECUTION]: The helper script scripts/_skill_cli.py implements a run_command function using subprocess.run. The skill instructions in SKILL.md direct the agent to execute a local Python script (suggest_next_skill.py) to determine the next steps in the development workflow.

visual-verification