He One-Shotted Photoshop: ArtCraft, Rust, and the AI Reverse-Engineering Wave (2026)
Quick verdict
Brandon Thomas, a 15-year industry veteran, used Claude Opus 5.5 to build ArtCraft — clean-room Rust clones of seven Adobe apps (Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, Acrobat Pro), free, cross-platform, with browser versions via WebAssembly. The alpha builds are rough, the original 100%-in-a-month promise has slipped to months, and the clean-room defense covers code but not trade dress. The capability claim, though, is no longer hypothetical.
The Claim
Brandon Thomas — a software engineer with fifteen years of industry experience, including early work at Square and roughly a decade of large-scale Rust projects — announced a suite of seven open-source applications cloning Adobe's Creative Cloud flagships, built in Rust using Claude Opus 5.5 as the primary development instrument: PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere), LightCraft (Lightroom), EffectCraft (After Effects), DesignCraft (InDesign), and PrintCraft (Acrobat Pro).
His summary of the method, on Reddit and Hacker News: "I one-shotted Photoshop" — followed by the declaration that software is over.
Strip the rhetoric and the verifiable core is still remarkable: seven cross-platform applications, in weeks rather than years, released under MIT and Apache licenses for macOS, Windows and Linux — a platform Adobe has never officially supported — with WebAssembly builds that run in a browser, and an architecture designed to be agent-ready: a CLI, a JSON control channel, and an MCP server so other AI agents can operate the tools.
The Method: Clean-Room, With an LLM Doing the Reading
Thomas describes the approach as clean-room replacement: study the original's behavior, reimplement from scratch without copying code or assets. It is a legally established practice — and the part that actually needed proving was whether a frontier model could do the reading at the scale Adobe's suite demands.
The stack choices are deliberate and worth noting:
- Rust, no Electron. Native performance and memory safety, against the resource-heavy web-view frameworks common in modern desktop apps
- WebAssembly builds for browser execution
- Agent-ready interfaces (CLI, JSON channel, MCP server) — the tools are designed to be operated by other agents, not just humans
- Funding through token-based access to Thomas's pre-existing ArtCraft visual AI model and IDE — a revenue stream he says prevents the project from becoming an orphaned small-team effort
The motivation is equally concrete. Thomas cited "the dark pattern 'cancellation fee' one too many times" — and the launch lands weeks after Adobe settled a $150 million federal lawsuit over difficult-to-cancel subscription plans. The timing is the argument.
The Promise Curve
This is where the story gets honest, and where coverage has been chasing a moving target.
The original claim: 100% feature parity with Adobe within a month. After Hacker News pressure and visible gaps in the alpha builds, that has been revised: 99% parity over months, not one month. Five of the seven apps are still marked in development; PhotoCraft and the PDF tool are labeled early alpha.
Real testing matches that trajectory. PhotoCraft launches fast, opens images, and handles basic operations credibly — layers, masks, brushes, curves, text tools, and PSD open/save. Beyond the basics: PSD files with smart objects break; a professional document that opens runs slowly with frequent small errors; FilmCraft crashes when importing many video clips and misrenders some assets. Thomas is offering bug bounties and recruiting open-source contributors to stabilize what shipped.
None of this refutes the velocity claim. It prices it: weeks produce scaffolding that resembles a decade-old product; the last mile — calibration, color accuracy, plugin ecosystems, stability under professional load — is exactly the part AI generation does not shortcut. Whether ArtCraft becomes a viable professional competitor on the order of GIMP or remains a proof of concept is undetermined.
The Legal Line
Clean-room reverse engineering is legally established — study the behavior, reimplement without copying code. Thomas used it, and it mitigates the most obvious claim: copyright in the code.
It does not cover everything:
- Trade dress. Mimicking an application's interface layout invites claims that the clean-room process does not answer. The coverage on this story keeps circling the same point: the apps are designed to be instantly recognizable as Adobe's
- Patents. Adobe holds thousands covering specific features and workflow mechanics; a clean-room process does not clear patent exposure
- The HN consensus, worth stating plainly: reverse engineering functionality is generally acceptable; cloning the look and feel is where the legal trouble starts
The project's own hedge — "in the logo and name we made some adjustments" — is not a legal strategy. It is an acknowledgment of where the line sits.
The Wave, Not the Exception
ArtCraft did not appear alone. rea — a separate project packaging the same technique as a tool — sat at the top of GitHub Trending the same week at 18,168 stars, gaining 4,655 in a single day, described as "reverse engineer anything with agents, from app behavior down to native binaries." The AnyPS5 porting tool rode the same list. Two projects, one technique: agents doing the reading that reverse engineering always required, at a scale one person can now direct.
That is the transferable finding, and it rhymes with what we keep seeing across this series: the volume work that used to require a team — reading a large surface, mapping behavior, scaffolding an implementation — is now agent work. The judgment calls, the last-mile polish, and the legal exposure remain human.
What To Watch
- The parity trajectory — whether 99% over months survives contact with professional workloads, or becomes the new "one month"
- Adobe's response — the $150M settlement shows the company under pressure on pricing; its tolerance for interface clones is a separate question, and its patent portfolio is large
- Whether the agent-ready architecture matters more than the apps — MCP-controllable creative tools are arguably the more durable idea here
- The legal filing that has not happened yet — a trade dress claim would set the boundary for the entire wave
Summary
ArtCraft is the clearest demonstration yet that frontier models can carry the volume work of reverse-engineering and reimplementing large commercial software — seven apps, three platforms, weeks. It is also a demonstration of where that capability stops: the alpha builds are rough, the parity promise has already slipped, and the parts of Photoshop that took Adobe decades — calibration, stability, ecosystem — are precisely the parts the model did not generate.
The legal line is equally instructive: clean-room covers the code, and the interface is where the exposure starts. The next filing in this story will probably be Adobe's.
For the wider pattern, see the skills gold rush — the same week's evidence that agent tooling is becoming an economy — and the agent-security reckoning for what happens when the same autonomy points at systems nobody consented to.
Related Articles
Keep reading