A Japanese developer built a working clone of the Windows 3.1 Program Manager using Claude Code in about an hour, then spent the following days polishing it into a shippable 12MB tool that now runs on both Windows 11 and Apple Silicon Macs. The project, called ReProgman, was first documented by Tom’s Hardware in a report published September 10, 2026, and it landed at a moment when “vibe coding” (writing software by describing what you want rather than typing it line by line) has moved from a niche hobbyist trick to a mainstream way of shipping real, if small, software products.
The developer, who goes by Mayuki, named the project after progman.exe, the file that powered the primary shell of Windows before Windows 95 replaced it with the taskbar and Start menu most people now take for granted. According to Tom’s Hardware’s Mark Tyson, Mayuki described the process directly: “I asked Claude to create a clone of the Windows 3.1 Program Manager,” and the AI assistant produced a usable first draft in roughly 60 minutes. That draft needed cleanup, but the shape of a decades-old interface was there almost immediately.
What ReProgman actually is, and why it exists
ReProgman is not an emulator and it is not a virtual machine. It is a native reimplementation of the look and feel of the Windows 3.1 Program Manager, the grid of program groups and icons that served as the entire desktop metaphor before Windows 95 introduced overlapping taskbars and a Start button. Reports describe the finished tool as a lightweight launcher, somewhere between 12MB and 16MB depending on the release build, that runs without an installer. Users download it, double-click, and the retro shell appears on top of a modern operating system.
That framing matters because it separates ReProgman from the long history of DOSBox-style emulation projects. Those tools recreate old hardware and run old binaries inside a sandbox. ReProgman instead recreates the interface from scratch, in modern code, so it behaves like a native app on whatever machine it’s launched from rather than like a window into an emulated PC from 1992. Tom’s Hardware’s coverage frames it plainly: if you are yearning for a quick-shot retro Windows experience on your modern Windows 11 or macOS system, this is a new vibe-coded clone of the Windows 3.1 Program Manager built for exactly that itch.
Mayuki has been candid about the tool’s appeal being nostalgia rather than utility. The developer’s own description of the experience, again per Tom’s Hardware’s reporting, sets expectations low and specific at once: “I think you’ll enjoy reminiscing with it for about 30 seconds.” That’s not a knock on the project. It’s an honest label for what a lot of retro-computing software actually delivers, a short jolt of recognition rather than a daily driver.
The one-hour build, and what “vibe coding” means in practice
“Vibe coding” is the term that stuck this year for a workflow where a developer describes a goal in plain language and lets an AI coding agent, in this case Anthropic’s Claude Code, generate the actual implementation. The developer then reviews, runs, and iterates rather than writing every line by hand. Tom’s Hardware’s account of ReProgman is one of the more concrete public examples of the workflow applied to a full, if small, application: a UI clone of a 30-plus-year-old shell, produced end-to-end by an AI coding agent inside a single working session.
According to that reporting, Claude Code worked pretty swiftly to churn out this clone of the Windows 3.1 shell, and the first version that was precipitated in about an hour was already functional enough to resemble the target interface, even if it needed further work before Mayuki considered it release-ready. The publication does not provide a public prompt transcript or session log, so the exact division of labor between what Claude generated unassisted and what required manual correction is not independently documented. What is documented is the outcome: a working shell clone, built fast, then refined over subsequent days into a distributable release.
That distinction is worth sitting with, because it’s the honest state of AI coding tools in September 2026. These tools can produce a strong first draft of a well-scoped, well-understood UI target almost immediately. They still benefit from, and in most reported cases require, a human pass to fix edge cases, tighten behavior, and prepare something for public release. ReProgman is a clean illustration of both halves of that equation rather than evidence that either half has been solved outright.
Platform support: Windows 11, and Apple Silicon macOS
The cross-platform angle is what elevates ReProgman from a Windows curiosity to a broader hardware story. The project explicitly targets Windows 11 on x64 PCs, Windows 11 on Arm devices, and macOS on Apple Silicon. That’s a meaningful spread. It means the exact same retro interface, coded once, deploys across Intel-class Windows PCs, Arm-based Windows machines like Snapdragon-powered laptops, and Apple’s M-series Macs, without three separate native codebases being hand-maintained by the developer.
Cross-platform reach like that has historically required either a heavyweight framework or a lot of platform-specific glue code, both of which take time most solo, nights-and-weekends developers don’t have. ReProgman’s multi-target build is possible in large part because Mayuki relied on .NET, a runtime designed from the ground up to compile and run the same managed code across Windows, macOS, and Linux hosts. Microsoft’s own .NET documentation describes exactly that goal for the platform, and it’s the toolchain choice, more than any single AI-generated line of code, that made a one-developer project able to ship for three architectures at once.
Under the hood: .NET 10, MIT licensing, and a no-install release
Mayuki has stated that building and running ReProgman requires .NET SDK 10.0 or newer, Microsoft’s current long-term-support-track runtime. That requirement places the project squarely on modern tooling rather than legacy .NET Framework, which is itself a small signal about how quickly AI-assisted side projects can adopt current-generation infrastructure: there’s no migration debt to drag along when the whole codebase is generated fresh.
The project’s source code has been released on GitHub under the MIT license, one of the most permissive open-source terms available, allowing essentially unrestricted reuse, modification, and redistribution as long as the original copyright notice is preserved. End-user-ready builds, meanwhile, are reported at between 12MB and 16MB and require no installation step: download, run, done. That packaging choice fits the tool’s purpose. A nostalgia app that asks for admin rights and a Setup Wizard would undercut its own joke.
Why this story is landing now
ReProgman surfaces at a point where Anthropic’s Claude Code has become one of the most widely discussed AI coding agents among professional developers, alongside GitHub Copilot, Cursor, and a growing field of competitors. What makes this particular story travel is its scale mismatch: a full, working, cross-platform desktop-shell clone, produced by one person’s evening project, using a tool most enterprises deploy for far more mundane refactoring and boilerplate work. The gap between “AI agent writes a unit test” and “AI agent recreates a historic operating-system shell in an hour” is exactly the gap that makes a story like this spread on tech forums and social feeds.
It also lands inside a broader pattern this year of small, AI-assisted retro-computing and nostalgia projects getting outsized attention relative to their user base. These projects rarely change anyone’s daily workflow. What they change is the public’s calibration of what a single developer, with an AI coding agent and a free weekend, can now credibly attempt. A Program Manager clone is a toy. The one-hour build time attached to it is the part people are actually reacting to.
Market impact: what this means for AI coding tools
No stock moved and no earnings call cited ReProgman, and it would be a stretch to claim otherwise. The market impact here is reputational and competitive rather than financial: stories like this function as informal marketing for whichever AI coding agent gets credited, and Anthropic’s Claude Code is the one collecting the mindshare this week. Coding-agent vendors compete partly on benchmark scores and partly on exactly this kind of viral, easy-to-understand demonstration, because “built a Windows 3.1 clone in an hour” communicates capability to a non-technical audience in a way that a SWE-bench percentage never will.
That dynamic puts pressure on rival labs to either produce their own headline-grabbing demos or to point at more rigorous, task-based benchmarks as the better signal. Expect competitors to keep leaning on the latter argument publicly while privately tracking exactly which viral builds get attributed to which tool, because attribution in stories like this one is a real, if hard to quantify, marketing asset.
Competitive comparison: AI coding agents building small desktop apps
ReProgman was built with Claude Code, but it sits inside a crowded and fast-moving field of AI coding assistants that all pitch some version of the same promise: describe a target, get working code back. The table below lays out how the major players position themselves for exactly the kind of task ReProgman represents, a scoped, well-understood UI-recreation project rather than a large enterprise codebase migration.
| AI coding tool | Maker | Primary interface | Best fit for a project like ReProgman |
|---|---|---|---|
| Claude Code | Anthropic | Terminal / CLI agent | Autonomous multi-file builds from a single natural-language prompt, as used in this project |
| GitHub Copilot | Microsoft / GitHub | IDE inline + chat | Line-by-line and function-level assistance inside an existing editor session |
| Cursor | Anysphere | Standalone AI-native IDE | Whole-project editing with an AI-first editor built around agent workflows |
| Codex (ChatGPT) | OpenAI | Chat / CLI agent | Conversational scaffolding and iterative refinement of generated code |
| Gemini Code Assist | IDE plugin + CLI | Integrated assistance across Google Cloud and Android tooling |
The specific detail that distinguishes Claude Code’s role in this story is the agentic, terminal-driven workflow: Mayuki described handing the agent a goal rather than a sequence of granular instructions, and getting back a runnable draft rather than a series of suggested edits to accept or reject one at a time. That’s the workflow Anthropic has pushed hardest for Claude Code since its release, positioning it less as an autocomplete layer and more as a delegate that owns an entire small task from prompt to working build.
Historical context: from Program Manager to the taskbar, and back again
The Program Manager that ReProgman recreates was Microsoft’s answer to a very different computing era. Windows 3.1, released in 1992, organized every installed application into a grid of icons inside “groups,” each its own child window, with no taskbar, no Start menu, and no system tray as later generations of Windows users would recognize them. Windows 95 replaced that entire model in 1995 with the desktop metaphor still recognizable in Windows 11 today: a taskbar, a Start button, and windows that minimize instead of iconifying into a group.
That three-decade gap is exactly what gives ReProgman its novelty. It’s not recreating a recent interface that current developers grew up on. It’s recreating one that predates the web browser as a mainstream consumer product, built by a developer using a tool, an AI coding agent, that didn’t exist in any form until the past few years. The nostalgia and the technology used to produce it are separated by roughly the same 30-plus years, which is part of why the story resonates beyond the usual retro-computing enthusiast crowd.
Hardware angle: why running natively on Apple Silicon matters
Running a Windows 3.1 interface clone natively on an Apple Silicon Mac, without Windows, without Boot Camp (which Apple discontinued years ago on M-series hardware), and without a third-party virtualization layer, is only possible because the app was written as a fresh, cross-platform managed-code project rather than ported from old Windows-specific source. Apple Silicon Macs use an Arm-based architecture entirely different from the x86 chips Windows 3.1 originally ran on, so any binary compatibility approach would require either emulation or, as here, a ground-up reimplementation targeting a runtime that already supports both architectures.
This is a small but real data point in a larger hardware trend: cross-platform managed runtimes like .NET have quietly made “write once, run on Windows x64, Windows Arm, and Apple Silicon” a realistic solo-developer capability rather than something that required a dedicated build and QA team. Combine that with an AI coding agent that can generate a working first draft of the UI logic in about an hour, and the barrier to shipping small, polished, multi-platform desktop software has dropped further than most casual observers of the AI coding space may realize.
What developers are saying
Mayuki on building with Claude Code
Mayuki’s own account of the build is the clearest first-hand data point available. As reported by Tom’s Hardware, the developer said: “I asked Claude to create a clone of the Windows 3.1 Program Manager.” That single sentence is doing a lot of work. It describes a goal-oriented prompt, not a line-by-line specification, which is the defining trait of the agentic coding workflow Claude Code is built around. Mayuki followed that up with a wry, low-stakes pitch for the finished tool: “I think you’ll enjoy reminiscing with it for about 30 seconds,” a fair summary of what a nostalgia toy like this is actually for.
Tom’s Hardware on the release
Tom’s Hardware’s own framing of the build speed reinforces the same point from the outlet’s side: “Claude Code worked pretty swiftly to churn out this clone of the Windows 3.1 shell,” and, on the specific first-draft timeline, “The first version that was precipitated in about an hour was…” functional enough to be recognizable as the target interface, according to the publication’s account. The outlet also captured the intended audience for the release directly: “If you are yearning for a quick-shot retro Windows experience on your modern Windows 11 or macOS system, then there’s a new vibe-coded clone of the Windows 3.1 Program Manager that may interest you.” (Tom’s Hardware)
ReProgman by the numbers
The table below pulls together the concrete, sourced specifications of the release as reported.
| Detail | Reported figure |
|---|---|
| Reported first-draft build time | About 1 hour, using Claude Code |
| Release package size | Between 12MB and 16MB |
| Installation requirement | None; runs directly after download |
| Runtime dependency | .NET SDK 10.0 or newer |
| Supported targets | Windows 11 x64, Windows 11 Arm, macOS on Apple Silicon |
| Source license | MIT (open source on GitHub) |
| Developer | Mayuki |
| Original interface recreated | Windows 3.1 Program Manager (progman.exe) |
Verifying and running ReProgman yourself
System requirements checklist
For readers who want to try the release rather than just read about it, the practical checklist is short. First, confirm your machine meets the stated runtime requirement: .NET SDK 10.0 or newer, available directly from Microsoft’s official .NET downloads page. Second, decide which build matches your hardware: Windows 11 on a standard x64 PC, Windows 11 on an Arm-based device such as a Snapdragon laptop, or macOS on an Apple Silicon Mac (M1 through the current M-series lineup). Third, because the release requires no installer, treat it like any other standalone executable from an independent open-source project: download it from the project’s official GitHub repository rather than a third-party mirror, and check the MIT license file included with the source if you intend to modify or redistribute it.
Because ReProgman is a UI clone and not an emulator, it won’t run actual 16-bit Windows 3.1 program binaries out of the box. It recreates the shell’s look and interaction model in modern code. That distinction matters for anyone expecting to relaunch decades-old .exe files inside it; that use case is what dedicated emulation projects and compatibility layers are built for, not a from-scratch UI clone like this one.
# Example: checking your installed .NET SDK version before building or running
# a .NET 10-targeted project like ReProgman
dotnet --list-sdks
# Expected output should include a line starting with 10.0.x
# If it doesn't, install the current SDK from Microsoft's official .NET site
# before attempting to build the project from source.The bigger pattern: AI agents and legacy interface recreation
ReProgman is a small project, but it’s an unusually clean test case for a question the software industry has been circling for a couple of years: how good are AI coding agents at recreating a well-documented, visually distinctive interface from scratch, without access to the original source code? A Windows 3.1 shell is an ideal target for that test. It’s simple by modern UI standards, extremely well-documented across decades of screenshots, tutorials, and nostalgia content, and visually unambiguous, which likely makes it easier for a model to reproduce accurately than a modern, animation-heavy interface with dozens of interdependent states.
That’s a meaningful caveat on how far this result generalizes. A one-hour build of a 1992 shell clone doesn’t tell you how an agent handles a modern, deeply stateful production app with authentication, databases, and real-time sync. But it does tell you something concrete about the current ceiling for narrow, well-scoped, visually well-documented UI recreation tasks, and that ceiling has clearly moved up in the past year.
Predictions: where this trend goes next
Based on the trajectory visible in this story and the broader AI coding tool market, a few predictions seem reasonable heading into the rest of 2026 and beyond.
- Expect more retro-interface recreation projects (classic Mac System 7, early GNOME or KDE builds, old game launchers) built the same way: a single developer, an agentic coding tool, and a short build window, published as free, open-source curiosities rather than commercial products.
- AI coding vendors will increasingly highlight small, viral, easy-to-understand demo projects like ReProgman alongside formal benchmark scores, because the former communicates capability to a broader audience than a leaderboard percentage ever could.
- Cross-platform managed runtimes like .NET will keep benefiting indirectly from this trend, since agentic coding tools appear to lean on well-documented, widely-trained-on frameworks to hit multiple hardware targets from a single codebase.
- Expect more scrutiny, not less, of exactly how much of these “built in an hour” projects is unassisted AI output versus developer cleanup, since public prompt transcripts and session logs remain rare even in widely covered examples like this one.
- Small nostalgia and utility apps built primarily through agentic coding will keep landing on GitHub under permissive licenses like MIT, continuing a pattern where AI-assisted side projects default to open, remixable code rather than closed distribution.
What this doesn’t prove, and what to watch for
It’s worth being precise about the limits of this story. There is no independently verified, minute-by-minute account of the build process, no published prompt log, and no third-party reproduction of the one-hour claim. The account rests on the developer’s own description as relayed through Tom’s Hardware’s reporting. That doesn’t make the claim false, but it does mean readers should treat “an hour” as the developer’s reported estimate of the first working draft, not a benchmarked, independently timed result.
What’s solidly documented, by contrast, is the shipped artifact: a 12MB-to-16MB, installer-free release that runs on Windows 11 x64, Windows 11 Arm, and Apple Silicon macOS, built on .NET 10, released under the MIT license. That combination of small size, no-install convenience, and genuine multi-architecture support is the part of the story that holds up regardless of exactly how many minutes the first draft took to generate.
Frequently asked questions
What is ReProgman?
ReProgman is an open-source clone of the Windows 3.1 Program Manager interface, built by a developer known as Mayuki using Anthropic’s Claude Code AI coding agent. It recreates the look and feel of the classic 1992 Windows shell as a lightweight, standalone app rather than as an emulator.
Can I run ReProgman on a Mac?
Yes. Reports confirm ReProgman targets macOS on Apple Silicon (M-series chips) in addition to Windows 11 on both x64 and Arm devices, since it is built on the cross-platform .NET runtime rather than Windows-specific code.
Does ReProgman let me run actual Windows 3.1 programs?
No. It is a from-scratch recreation of the shell’s interface and interaction model, not an emulator, so it cannot execute original 16-bit Windows 3.1 application binaries. Emulation projects and dedicated compatibility layers serve that separate use case.
What software is required to build or run ReProgman?
The developer states that building and running the project requires .NET SDK 10.0 or newer, available from Microsoft’s official .NET downloads page. Pre-built end-user releases reportedly do not require any installation step.
Is ReProgman free and open source?
Yes. The project’s source code has been released on GitHub under the MIT license, one of the most permissive open-source licenses, which allows reuse, modification, and redistribution with attribution.
How long did it take to build ReProgman?
According to the developer’s own account, as reported by Tom’s Hardware, the first working version was produced in about an hour using Claude Code, though additional work followed to refine that draft into a public release. This timeline is self-reported and has not been independently verified minute-by-minute.
What is “vibe coding”?
Vibe coding describes a workflow where a developer describes a desired outcome in natural language and an AI coding agent, such as Claude Code, generates the implementation, with the developer reviewing and refining the output rather than writing most of the code by hand.
Why does a Windows 3.1 clone matter for the AI coding industry?
It functions as a widely shared, easy-to-understand demonstration of what current agentic coding tools can produce quickly for a well-scoped, well-documented UI target, giving both enthusiasts and industry watchers a concrete, low-stakes example of the technology’s present capability.
