The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install UI components by downloading registry configurations from an external third-party URL (ui.inference.sh) using the shadcn CLI. This pattern executes remote configurations to generate files and install dependencies within the user's project environment.
Evidence: npx shadcn@latest add "https://ui.inference.sh/r/agent.json" in SKILL.md and references/starter-snippets.md.
Evidence: npx shadcn@latest add "https://ui.inference.sh/r/widgets.json" in SKILL.md.
[EXTERNAL_DOWNLOADS]: Fetches component definitions, primitives, and widgets from https://ui.inference.sh/r/agent.json, https://ui.inference.sh/r/widgets.json, https://ui.inference.sh/r/actions.json, https://ui.inference.sh/r/citations.json, and https://ui.inference.sh/r/sources.json.
[COMMAND_EXECUTION]: Uses the npx shadcn command to modify the local project structure and automate the installation of components from remote URLs.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to render UI components (like widgets and tool states) based on structured data returned by AI models. This creates a vulnerability surface where malicious instructions or deceptive content embedded in model-generated tool outputs could manipulate the application's UI or trick the user.
Ingestion points: AI SDK message streams and tool results processed in ChatSurface and via uiTransform in SKILL.md.
Boundary markers: The skill does not provide specific instructions for isolating or sanitizing model-provided data before it is rendered in UI widgets.
Capability inventory: The skill utilizes components like <Response>, <Tool>, and various widgets to render model output, and supports client tools that can perform browser-side side effects.
Sanitization: While the skill suggests that invalid payloads should fail safely, it does not mandate specific sanitization for the content being rendered.

inference-agent-ui