Building a speech-to-Markdown app with three coding agents (www.apptornado.com)

In a recent experiment, a developer tested three AI coding agents—Cursor CLI, Claude, and Gemini CLI—against a straightforward task of building a speech-to-Markdown app using the web speech recognition API and SvelteKit. The results highlighted a significant performance gap, with Cursor CLI excelling by delivering a production-ready application right away, while Claude produced a workable app after a minor fix, and Gemini struggled with basic setup, necessitating manual interventions. This experiment's findings are notable for the AI/ML community, as they challenge existing perceptions of coding agent capabilities. Cursor's Composer-1 model demonstrated superior efficiency and code quality, sparking a discussion about its underrepresentation in coding leaderboards. The results raise important questions regarding the training breadth of AI models, particularly around niche frameworks like Svelte, and the relevance of current coding benchmarks, which often focus on more commonly used technologies. This could signify a need for more comprehensive evaluations of coding agents, especially in diverse programming contexts.