0から10日でmacOSアプリをMVPリリースした話 — GPTとCursorをどう分業したかA solo developer built and shipped SnapSlim, a macOS image-compression app…

A solo developer has documented how they built and shipped SnapSlim, a macOS image-compression application, in just 10 days by systematically dividing development work between GPT and Cursor. The account, published on Zenn, is notable less for the app itself than for what it suggests about how far AI-assisted tooling can compress the timeline of individual software projects, from an empty repository to a minimum viable product.

SnapSlim is a straightforward utility. It supports JPEG, PNG, and HEIC formats, and its core interaction is drag-and-drop: users drop images into the app, and it reduces file size while trying to keep visible quality degradation to a minimum. That framing—smaller files with limited perceptual loss—places it in the familiar category of lossy and format-aware compression tools rather than anything experimental in its output. The app is written in Swift, the language Apple positions as the default for building native macOS and iOS software.

The developer frames the project explicitly as an experiment. The stated question was how much faster solo development could become if GPT and Cursor were incorporated into the workflow "seriously," rather than used casually for occasional snippets. To test that, the two tools were assigned distinct roles across the design, implementation, and testing phases. This division of labor is the central theme of the write-up, and it reflects a broader pattern emerging among indie developers who treat large language models and AI code editors as complementary rather than interchangeable.

In practice, the split appears to lean on each tool's strengths. GPT, as a conversational model, is well suited to higher-level tasks such as clarifying requirements, discussing architecture, weighing trade-offs, and reasoning through design decisions before code is written. Cursor, an AI-native code editor built on a fork of Visual Studio Code, is oriented toward the implementation surface: it keeps the project's files in context, can generate and edit code inline, and supports iterative refinement directly within the editor. Testing sits between the two, where a model can help draft test cases or interpret failures while the editor applies fixes. The developer's account suggests this separation reduced friction, though the exact boundaries they drew are specific to their own process and may not transfer cleanly to every project.

The 10-day figure is the headline claim, and it is worth reading with some caution. An MVP is by definition a pared-down release, and SnapSlim's scope—three image formats and a single primary interaction—is deliberately narrow. Building a focused utility is a very different undertaking from shipping a feature-rich application with extensive edge-case handling, localization, and long-term maintenance. The story is best understood as a case study of one developer's experience rather than a benchmark that guarantees similar results elsewhere.

The context matters because the tooling landscape has shifted quickly. Cursor has become one of the more visible AI coding environments, competing with GitHub Copilot, which integrates AI suggestions into existing editors, and with newer agentic assistants that can execute multi-step tasks. On the model side, general-purpose systems from OpenAI and rivals such as Anthropic and Google are increasingly capable at code generation and reasoning, which has encouraged workflows that pair a chat model for planning with an editor-integrated assistant for execution. Native Apple development adds its own considerations, including Xcode, the Swift toolchain, and macOS distribution requirements such as code signing and notarization, which typically fall outside what AI tools automate.

For readers, the takeaway is practical rather than promotional. The SnapSlim write-up offers a concrete example of how design, implementation, and testing responsibilities can be allocated between a conversational model and an AI editor, and it provides a data point on the speed such a setup can enable for a tightly scoped project. Whether that pace holds for larger or more complex software remains an open question, and the results are likely to depend heavily on the developer's own familiarity with Swift, macOS conventions, and the tools involved. As more solo builders publish similar accounts, a clearer picture should emerge of where AI-assisted workflows reliably help and where human judgment still sets the limits.