One-Shot Photoshop

Over the weekend a developer named Brandon Thomas announced that he had used Claude Opus 5.5 to rebuild Adobe's creative suite from scratch: seven applications in Rust, clean-room, open source, including a Photoshop reimplementation he says the model produced in one shot. "Software is over," he wrote. The repositories are real, and eight days after the first one appeared they say something more interesting than the slogan.

The claim is auditable in a way most launch claims are not. The storytold organization on GitHub holds PhotoCraft as a Rust workspace of twenty-four crates with more than 1,700 tests and over 640 commits in its first eight days, and the project publishes two generated documents most software companies would never ship. The first is a parity file, produced by a build task that walks Photoshop's entire menu tree against the live command registry. It reads 627 of 627 menu items live, one hundred percent. The second is a scorecard, and it reports that 56 of 139 settings currently do nothing, that 22 of 47 performance scenarios run over budget, and that PSD compatibility passes its oracle check on 229 of 309 files.

Both documents are true at once, and the distance between them is the actual product update. A menu tree is a specification, and reproducing specifications is what language models are good at; the project's own contributor guide states the goal as 1:1 Photoshop parity, same menus, shortcuts, behavior and file fidelity, with every feature drivable by an agent. The gap between every menu item existing and every menu item working is where the next months of this project live, and the marketing site is honest about it in its own way, labeling PhotoCraft early alpha and most of its siblings in development.

The Promise, Revised

Thomas's parity promise has already been walked back once in public. A pledge of full parity within a month, made when the suite launched, became a Hacker News revision days later: 99 percent parity, measured in months, not years. Against the scorecard the revision looks conservative. What has not been revised is the mechanism, and the mechanism is the part with a future: every app in the suite exposes a command-line interface, a JSON control channel, and an MCP server, so an agent can drive the whole thing headless. The Photoshop interface is the announcement. The automation surface is the product.

Beyond Adobe

The stranger fact is that the suite has already outgrown its target. Alongside the seven Adobe counterparts, the same organization publishes clean-room rebuilds of Microsoft Word, Excel and PowerPoint, of AutoCAD, and of Pro Tools. PhotoCraft passed 25,000 stars in its first week, and the last thirty commits across the workspace carry twenty-two distinct author names, one of them credited to Claude itself. Whatever this began as, it is not one developer anymore.

Adobe has said nothing publicly, and the legal theory writes itself: familiar from day one is the trade-dress claim and the confession in the same sentence, and the courts have seen clone-and-succeed cases before. But a suit, if it comes, will be about the surface. Underneath it sits the first large-scale demonstration that a frontier model plus a patient harness can produce a credible industrial codebase in days, which is either the end of software scarcity or the start of a very long bug list. The scorecard, to its credit, is honest about which one it is today.

The Scorecard Is the Product

Most launches hide their gaps; this one generates a document of them. 627 of 627 menu items live, and 56 of 139 settings that do nothing, in the same file, from the same build task. The parity file is the marketing. The scorecard is the engineering. A project that ships both in the same commit is either the most confident or the most disciplined launch of the year, and the interesting bet is that it is the second one.