Teaching Your AI Web Design Some Actual Taste

Maneshwar creates git-lrc, a free Micro AI code reviewer that runs on every Git commit, allowing developers to discover the project on GitHub. The reporter questions the output generated by Claude Code, which creates modern landing pages that look identical to others. Maneshwar has been fighting this trend and shares how he overcame it without relying on a magic prompt or a lengthy design document. The key lies in three steps and a few tools.

The first step is to cultivate taste by collecting visually appealing web designs. This involves browsing sites like Dribbble, Pinterest, and Twitter to save screenshots and links of designs that evoke positive emotions. The collected inspirations can be organized using Claude Code, which creates a personalized inspiration library web app. This step aims to provide a starting point for future projects instead of starting from scratch with an average result.

Step two involves equipping Claude Code with additional tools that are not available by default. This includes using Impeccable, an open-source skill for better visual judgment. Impeccable identifies common slop patterns in design and instructs Claude Code to avoid them. It offers commands like /critique, /bolder, /quieter, and /polish to refine the design. Additionally, the Higgsfield MCP allows Claude Code to generate images and videos, which can be integrated into the project seamlessly.

Finally, for smaller components like buttons, cards, and pagination, Maneshwar recommends browsing 21st.dev, an open component registry. Each component comes with a copy prompt button, enabling Claude Code to rebuild the component within the project's codebase. The reporter emphasizes the importance of developing one's taste rather than relying on generic AI-generated designs, which can result in predictable and uninspired outputs.