The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill generates an interactive mockup by embedding diagram data from the conversation into an HTML template.
Ingestion points: Diagram elements such as labels and titles are gathered from the conversation context and written to /tmp/layout.json as described in SKILL.md.
Boundary markers: No explicit boundary markers or warnings are used to prevent the agent from including untrusted instructions in the data.
Capability inventory: The skill uses scripts/interactive_mockup.py to write files to the /tmp directory and execute system-level commands to open the browser.
Sanitization: The generate_html function in scripts/interactive_mockup.py performs a direct string replacement for TITLE_PLACEHOLDER using the diagram title. This lacks HTML entity encoding, providing a surface for HTML/JavaScript injection if the agent includes malicious user-provided text in the title.
[COMMAND_EXECUTION]: The interactive_mockup.py script uses subprocess.run(["open", args.output]) to launch the default browser. While used for its intended purpose of showing the generated diagram, this involves direct interaction with the host's command execution environment.
[DYNAMIC_EXECUTION]: The skill instructions in SKILL.md direct the agent to dynamically write and execute Python scripts using matplotlib and seaborn to produce PNG mockups. The generation and execution of code at runtime based on changing conversation context is a form of dynamic execution.

visual-mockup