The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill generates HTML and SVG content based on user-provided descriptions, which presents a potential surface for indirect prompt injection.
Ingestion points: Technical descriptions and diagram labels provided in the user prompt (e.g., "Diagram X", "Sketch the architecture").
Boundary markers: No specific delimiters are mandated for separating user input from the diagram structure, although output is restricted to a structured SVG format.
Capability inventory: The skill uses the agent's capability to write .html files to the local disk and interacts with a local platform listener tool.
Sanitization: Instructions explicitly require using safe DOM construction via textContent and createElement, and strictly forbid using innerHTML with variable data to mitigate XSS risks.
[DYNAMIC_EXECUTION]: The skill generates interactive artifacts that include inline JavaScript and CSS to facilitate communication with the agent platform.
It implements a local handshake mechanism (window.__CLAUDE_SUBMIT_URL__) for data interaction within the sandbox.
It provides guidelines for using localStorage safely by requiring unique prefixes to prevent data leakage between different artifacts.
[EXTERNAL_DOWNLOADS]: The skill allows the use of external assets from a well-known and trusted service.
It references CSS stylesheets from Google Fonts to provide high-quality typography for the generated diagrams.

html-svg-diagrams