The promise of “vibe coding” – building applications simply by describing them in natural language to an AI – has rapidly moved from tech buzzword to demonstrable reality. While the results aren’t always polished gems, the sheer accessibility of application development is undergoing a significant shift. The effort required to create a web app through conversational AI is, as one developer recently discovered, surprisingly minimal.
The recent surge in interest stems from advancements in large language models (LLMs) and the emergence of platforms designed to leverage them for coding. The core idea is simple: instead of writing lines of code, users articulate their desired functionality and the AI translates that into a working application. This lowers the barrier to entry for aspiring developers and allows experienced programmers to rapidly prototype ideas.
One developer, frustrated with the limitations of existing e-reader devices – specifically the inability to simultaneously read and listen to audiobooks on a Kindle – decided to put vibe coding to the test. The goal: to create a functional e-reading web application with features currently lacking in popular commercial products, including real-time text highlighting synchronized with audio playback, dynamic background music tailored to the content, and even sound and visual effects triggered by specific words.
The project, dubbed “Tome Reader,” wasn’t undertaken with the intention of replacing existing e-readers, but rather as a proof of concept and a challenge to the capabilities of current AI coding tools. The developer chose to experiment with three leading LLMs: Gemini, Claude, and ChatGPT, rather than attempting to select a single “best” option. The approach involved a layered prompting strategy, designed to refine the instructions and leverage the strengths of each AI.
The process began with Gemini, used to establish a basic, functional version of the application. The developer then asked Gemini to generate a prompt – a detailed set of instructions – that could be used with other chatbots. This prompt was then refined by Claude, which identified potential improvements and addressed overlooked considerations. Claude, in turn, generated a further refined prompt, which was then tested with ChatGPT.
This iterative approach aimed to harness the collective intelligence of the three LLMs, creating a prompt that maximized the chances of success across different platforms. The final prompt was then used to generate the application code independently on each chatbot, allowing for a direct comparison of results.
Gemini proved particularly adept at establishing the core functionality of Tome Reader, resolving initial issues with text-to-speech (TTS) voice loading by implementing an initialization screen. Claude excelled at refining the application’s behavior, specifically enhancing the responsiveness of trigger words that activated sound and visual effects. However, Claude also introduced an unexpected constraint, limiting the activation of these effects to once per sentence to avoid overwhelming the user – a logical decision, but not one explicitly requested.
ChatGPT successfully recreated the project based on the final prompt, though it occasionally struggled with adding new features, such as a dedicated volume slider for the background music. The developer ultimately returned to Claude to address this issue and refine the prompt further.
The testing revealed inconsistencies in the user experience across the different platforms. Claude consistently offered a preview of the generated code and a direct download option, while ChatGPT sometimes required the developer to manually copy and save the HTML. Gemini consistently required the more cumbersome copy-and-paste method.
Despite these differences, the developer concluded that all three LLMs were capable of generating a functional version of Tome Reader, albeit with varying degrees of effort and refinement. The quality of the final product was heavily dependent on the quality of the prompt, highlighting the importance of clear and detailed instructions. The developer found that the performance between Gemini 3 Pro and the free versions of ChatGPT and Claude was surprisingly comparable, suggesting that access to premium models isn’t always necessary for successful vibe coding.
The experiment underscores a key takeaway: the real power of vibe coding lies not in the AI itself, but in the ability to effectively communicate your vision to it. A well-crafted prompt can unlock significant potential, regardless of the underlying model. While the technology is still evolving, the ease with which a complex application like Tome Reader could be created through conversational AI signals a fundamental shift in the landscape of software development.
The developer’s experience also highlights the potential for AI to address specific user frustrations with existing products. In this case, the limitations of Kindle devices served as the impetus for a creative solution, demonstrating how vibe coding can empower individuals to build tools tailored to their own needs.
