AI coding design competition
Vibe checking Cursor, Windsurf, Replit, Lovable and Firebase Studio
This week I recorded a Youtube video instead of writing my weekly newsletter.
I set up a mini design competition to evaluate the design skills of the top five AI coding tools of the moment. You can find the full video here:
The results of the design competition:
Cursor: 17/25 points
Windsurf: 19/25 points
Replit: 23/25 points
→ Lovable: 24/25 points (winner)
Firebase Studio: DNF
The winning design can be found here.
My main takeaways:
Today’s AI tools are less about emergent behaviour and more about system design. I suspect that Lovable won the design competition because it was engineered with certain design principles in mind—not because it had access to more powerful AI models or a better user interface.
Agentic software works best with heavy use of AI reasoning capabilities. Since these processes tend to be very time consuming, agentic systems should be set up with self-evaluating feedback loops to minimise the need for user feedback for error correction.
Intent engineering and context engineering are two sides of the same coin. The best agentic AI software tools today allow users to set the context in ways their intent matches system design. Intentionality in these systems emerges from user and system design—not from the LLMs.
Thanks again for reading The Circuit!
PS here is the master prompt for the design competition:
Create an interactive landing page titled "The Fourth Separation" using Next.js and React.
The page should visually show the progression of existence through key transitional stages of existence becoming gradually more complex (four "separations"). The four stages need to be presented as a scrolling experience.
These are the four stages that need to be represented independently:
First Separation: Represent the split between space and time (e.g., Big Bang visualization).
Second Separation: Represent the emergence of organic matter from inorganic matter (e.g., animating simple molecules forming complex structures).
Third Separation: Represent the development of language, culture, and technology as distinct from purely biological organisms (e.g., abstract visuals of networks, symbols, early tools).
Fourth Separation: Give your own interpretation and visualisation of this stage, focusing primarily on visuals and animation. Minimal text can be used here if needed, but only if they support the animations and visualisation.
Interactivity: Implement scroll-triggered animations to reveal or emphasise each separation. Add subtle background effects (e.g., parallax, particle animations) that enhance the theme. It is up to you to decide between a vertical scrolling or horizontal scrolling movements.
You can take inspiration from the series of four panels of Sandro Botticelli depicting events from the life of Saint Zenobius (Four Scenes From The Early Life of Saint Zenobius, Three Miracles of Saint Zenobius / National Gallery, Three Miracles of Saint Zenobius / Moma, Last Miracle and the Death of Saint Zenobius), but your interpretation needs to be strictly 2025 and futuristic.
Last week in AI
A study by the Stanford Human-Centered Artificial Intelligence (HAI) institute published last week audited AI's potential for automation and augmentation across the U.S. workforce, revealing a significant disconnect between current AI development and workers' desires. They highlight the need for a more human-centered approach in the design of AI agents to ensure they effectively complement the workforce. The authoer propose a benchmark (WORKBank) to help better match AI agents and humans to the right tasks.
Anthropic researchers evaluated the ability of AI models to perform and monitor for sabotage, finding that current models are not yet sophisticated enough to carry out complex deceptive tasks. While the most advanced models showed some success, the research highlighted that AI systems are also not very effective at detecting such sabotage in other AIs. The study suggests that improving AI monitoring capabilities, particularly by giving them insight into other AIs' reasoning processes, will be critical for safety as these systems become more autonomous.
Anthropic announced that its AI coding assistant, Claude Code, now supports remote Model Context Protocol (MCP) servers, a significant step in making the tool more customizable for developers. This update allows Claude to connect directly to third-party tools and data sources like Sentry and Linear, pulling in real-time context such as error reports or project status without requiring developers to manage local server infrastructure.
Cursor launched its Slack integration to shorten the distance between team discussion and development. Developers can now mention a
@Cursorbot in any Slack thread, which allows a remote AI agent to read the conversation's context, understand the task, and even create pull requests in GitHub automatically. This feature aims to create a seamless workflow, turning bug reports and feature requests directly into actionable coding work without leaving Slack.