The Shift to AI-Native Development

Meng To has fundamentally changed his design and engineering process, moving away from traditional tools like Figma for initial design and manual coding for implementation. He now operates in an "AI-native" workflow centered on Codex, which he views as a significant evolution from standard LLM chat interfaces. By keeping his entire project environment local and integrated with AI agents, he can build, design, and plan within a single, unified workspace.

Building Personal Tools and Local Workflows

Meng emphasizes the importance of building custom tools to solve specific frustrations. Rather than relying on cloud-based "walled gardens" like Notion, he builds local, markdown-based applications that allow him to maintain control over his data. This local-first approach enables his AI agents to access his entire file system as context, allowing for more accurate and personalized outputs. He treats his computer as a living repository of his design history, fonts, and code, which he feeds into AI models to ensure the output reflects his unique taste rather than generic "AI slop."

The "Last 10%" and Human Taste

A core argument in the conversation is that while AI can handle the bulk of the work, the "last 10%" of polish is what separates high-quality products from generic AI output. Meng argues that designers must remain a step ahead of current AI trends—such as overused purple gradients or specific animation styles—by applying their own human memory and design history. He uses screenshots as his primary interface for communication with AI, finding it faster and more precise than typing or voice for describing visual bugs or layout adjustments.

Iterative Prompting and Animation

Meng reveals that he no longer writes code manually, relying entirely on AI to generate it. However, he stresses that this requires deep product knowledge and thousands of iterations. For complex tasks like interaction design and animation, he uses a clever "reverse-engineering" tactic: he provides the AI with a video of an animation he admires and asks the model to generate the corresponding prompt or code. This bridges the gap between visual intent and technical implementation, especially for complex subjects like shaders or motion design.

Key Takeaways

  • Stop chasing trends: Avoid "AI slop" by staying ahead of the baseline design trends and injecting your own unique design history into the model's context.
  • Use screenshots as your primary interface: When working with AI, a screenshot is the fastest way to communicate visual issues, padding, or layout problems without needing to write lengthy descriptions.
  • Build your own tools: If existing software doesn't fit your workflow, build a local version. Local files provide better context for AI agents than cloud-based SaaS tools.
  • The 10,000-prompt rule: High-quality software isn't built in one prompt. Expect to iterate thousands of times to refine features and edge cases.
  • Reverse-engineer with video: If you don't know the technical terminology for an animation or effect, record a video of it and ask the AI to describe the prompt or code required to replicate it.
  • Be an editor, not just a creator: Treat yourself as an editor of AI-generated work. Your value lies in your ability to spot bugs, refine details, and curate the final output.