Google's Stitch Skills library turns coding agents into design tools

Google has published an open-source library of Agent Skills for Stitch, its collaborative design platform. The repository, hosted at google-labs-code/stitch-skills, contains plugins that let coding agents like Claude Code, Cursor, and Gemini CLI perform design tasks directly from the command line.

The library follows the Agent Skills open standard, a specification for describing agent capabilities. Each skill is a structured directory with a SKILL.md file, executable scripts, and resource directories. The design means any agent that understands the standard can invoke the skills without custom integration.

Three plugin groups ship in the repository. stitch-design handles core design workflows: converting frontend code to Stitch designs via HTML extraction, generating new screens from text or images, and managing design systems by uploading DESIGN.md files. stitch-build covers code generation and framework integration: converting Stitch screens to React components, producing React Native code with platform-specific StyleSheet output, and generating walkthrough videos using Remotion. stitch-utilities provides supporting tools: analyzing Stitch projects to produce DESIGN.md files, transforming vague prompts into polished specifications, and enforcing “premium, anti-generic UI standards” through a skill called taste-design.

The most significant shift here is not the tooling itself. It is the boundary crossing. Design tools and coding agents have largely lived in separate worlds. Agents edit code. Design tools handle screens. The Stitch Skills library builds a bridge between them. A single prompt like “Upload the frontend code at /path/to/dashboard into a Stitch project named ‘Dashboard-Migration-2026’” triggers stitch::code-to-design, which extracts HTML, infers a design system, and uploads the result. The reverse flow is equally direct: “Convert all screens in Stitch project projects/123 to React components” calls stitch::react-components and produces framework code.

Google is not the first to attempt this kind of design-to-code pipeline. Figma has plugins that generate code from designs. Anima and Zeplin offer similar bridges. What is different here is the agent-first architecture. The skills are not buttons in a UI. They are executable modules that any compliant agent can call. The agent decides when to invoke them, how to chain them, and what context to pass between them.

The timing matters. Coding agents are proliferating. Anthropic’s Claude Code, OpenAI’s Codex CLI, Cursor, and Google’s own Gemini CLI all compete in the same space. Each has its own plugin model. The Agent Skills standard offers a common interface. If it gains adoption, it could reduce the fragmentation that currently forces plugin developers to write separate integrations for each agent. The Stitch Skills library is a test case for that idea.

The repository also signals something about Google’s strategy for Stitch. The platform launched as a design tool with real-time collaboration. Now it is positioning itself as a hub that connects design, code, and agent workflows. The stitch-loop skill, which generates complete multi-page websites from a single prompt, suggests an ambition that goes beyond traditional design tools. Google is building toward a world where a designer describes a site in natural language, an agent generates the design in Stitch, another agent converts it to code, and a third deploys it. The skills library is the scaffolding for that pipeline.

There are limits to what the library can do. The skills depend on the Stitch MCP server being configured and running in the agent’s environment. That means environment variables, credentials, and server setup. The repository is clear about this in its prerequisites section. It is not a plug-and-play solution. It is a set of building blocks for teams that already use Stitch and want to integrate agents into their workflow.

The repository also carries a notable disclaimer. It is “not an officially supported Google product” and is “not eligible for the Google Open Source Software Vulnerability Rewards Program.” That language is standard for Google Labs projects, but it matters. Teams that build workflows around these skills should expect breakage. There is no guarantee of backward compatibility or maintenance.

The most interesting skill in the library may be taste-design. It generates DESIGN.md files that enforce “premium, anti-generic UI standards” with “strict typography and calibrated colors.” The existence of this skill suggests Google is aware of a problem that emerges when agents generate designs at scale: they produce generic, same-looking interfaces. The taste-design skill is an attempt to codate taste, to make the agent produce output that does not look like agent output. Whether that works in practice is an open question.

What the Stitch Skills library makes clear is that the boundary between design tools and coding agents is dissolving. Agents are no longer limited to editing text files. They can edit designs, generate components, enforce design systems, and produce videos. The skills library is a concrete example of what that looks like. It is not revolutionary. It is practical. And it points toward a future where the distinction between designing and coding becomes a workflow choice, not a tool boundary.