The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download a pre-built binary package (ZIP file) and clone a repository from an untrusted third-party GitHub user account (mikesheehan54).
Evidence: wget https://github.com/mikesheehan54/Claude-Code-Design-AI/releases/download/Software/ClaudeDesign.zip and git clone https://github.com/mikesheehan54/Claude-Code-Design-AI.git.
[REMOTE_CODE_EXECUTION]: The installation instructions guide users to download, extract, and then execute code from an unverified source via npm run dev. Running code from an unknown repository is a significant security risk as the contents could be malicious.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data that could contain malicious instructions.
Ingestion points: Untrusted data enters the agent context through the screenshotToReact function (processing user-supplied images) and the MarkdownRenderer component (rendering AI-generated or external content).
Boundary markers: No explicit boundary markers or "ignore embedded instructions" warnings are present in the provided code snippets to isolate untrusted data.
Capability inventory: The skill includes network capabilities via the Anthropic SDK and UI rendering components.
Sanitization: While the MarkdownRenderer component supports a sanitize prop, its effectiveness depends on implementation, and no sanitization is mentioned for the image-to-code conversion process.

claude-design-ui-framework