From Wet Lab to Web Dev: The Geneguessr AI Case Study
Coverage of lessw-blog
In a recent post, a LessWrong contributor details the creation of "Geneguessr," a web game built by a biologist with limited coding experience using advanced AI agent workflows.
In a recent post, lessw-blog discusses the development of "Geneguessr," a browser-based game inspired by Wordle and Geoguessr. While the game itself-which challenges players to identify human proteins based on similarity clues-is a novel educational tool, the primary signal for technical leadership lies in its construction. The creator is a wet lab biologist with minimal coding experience who successfully deployed a functional application using a suite of AI tools, specifically Claude, ChatGPT Codex, and the Model Context Protocol (MCP).
This narrative is critical for observers of the "AI-assisted development" space because it moves beyond simple code generation into the realm of agentic workflows. We frequently hear predictions about a coming flood of software generated by non-technical users. However, the reality often falls short due to the complexity of debugging and project management. This case study provides a concrete answer to the question: "What does the workflow of a successful non-coder actually look like right now?"
The author details a two-month development cycle that relied heavily on specific integrations. Rather than simply pasting code back and forth into a chat window, the developer utilized Linear MCP to manage the project's Kanban board and Playwright MCP to handle end-to-end testing. This usage of MCP (Model Context Protocol) allowed the AI agent (Claude) to interface directly with development tools, effectively bridging the gap between generating syntax and managing a software lifecycle.
The post also highlights the nuances of model capability. While Claude served as the primary engine for development and project management, the author notes that it struggled with complex, logic-heavy bugs. For these specific hurdles, the author turned to ChatGPT Codex, suggesting that a multi-model approach is still necessary for robust application development. The experience underscores that while AI can lower the barrier to entry, it currently shifts the user's burden from writing syntax to acting as a technical product manager and QA lead.
For developers and tech strategists, this post offers a grounded look at the current capabilities and limitations of AI coding agents. It demonstrates that while the tools are powerful enough to enable a biologist to build a React app, the process remains labor-intensive enough to prevent an immediate saturation of the app market.
We recommend reading the full post to understand the specific interactions between the AI agents and the testing frameworks used.
Read the full post on LessWrong
Key Takeaways
- The project serves as a proof-of-concept for non-technical users building complex web apps using AI agents.
- The workflow utilized Linear MCP for issue tracking and Playwright MCP for automated testing, demonstrating practical agentic integration.
- A multi-model strategy was required: Claude handled architecture and management, while ChatGPT Codex resolved complex bugs.
- The two-month timeline suggests that while AI enables creation, it does not yet eliminate the significant time investment required for debugging and refinement.