AI-Rebuilt Adobe Suite Shows Promise but Falls Short of Hype

Brandon Thomas, a developer with fifteen years of experience including work on Square's payment systems, announced in October 2026 that he had rebuilt Adobe's entire Creative Suite using Claude Opus 5.5. The seven programs—PhotoCraft, FilmCraft, and five others still in development—launched as free, open-source tools written in Rust and runnable on local machines. The news spread rapidly across Reddit and developer forums, framed as a stunning achievement: decades of Adobe's engineering condensed into days by one person and one AI model.

Thomas had promised full feature parity within a month. PhotoCraft, the most polished of the seven, arrived with layers, masks, brushes, curves, text tools, and the ability to open and save PSD files. The interface was deliberately designed to feel familiar to Photoshop users. Early testers reported that the application launched quickly and handled basic image operations well. But the promise fractured almost immediately under real-world use. Users discovered that PSD files containing smart objects opened incorrectly. Professional designers found that moderately complex files would load but then crawl through operations, punctuated by small, frequent errors. FilmCraft, the video editor, crashed when importing large numbers of video clips and displayed materials abnormally. When pressed on Hacker News about the one-month timeline, Thomas revised his estimate: full parity would take months, not weeks. He explained that he had released the project publicly to gather users, collect bug reports, and invite community participation in development.

The gap between what looked finished and what actually worked revealed something fundamental about the difference between building software that appears complete and building software that endures. PhotoCraft's architecture was sound—separate Rust modules for PSD file reading and writing, image composition, filter calculation, and interface. Tests showed that smart object and filter information could now be saved and read back correctly. But when PhotoCraft attempted to recalculate Photoshop's filter effects, only 5 out of 30 test cases produced matching results. Photoshop's smart filters are a complete workflow: users convert layers to smart objects, attach effects like blur or sharpening without rewriting original pixels, then adjust parameters and filter order as needed, with Photoshop recalculating the image each time. PhotoCraft could save and read these operation records. It could not reliably reproduce the same visual output.

Premiere, Adobe's video editor, offers hardware-accelerated playback and export, plus support for a vast ecosystem of plugins that editors rely on to apply effects, process long footage, and move files between post-production teams. FilmCraft could perform multi-track editing, color grading, and export. It had separate modules for video codec, timeline editing, rendering, GPU processing, and export. But commonly used effects and Lumetri color grading still relied on the CPU, making processing glacially slow. It could not load third-party plugins like VST3 or Audio Units. Meanwhile, Adobe was already planning a new generation of image processing engines designed to compile filters and layer blending into GPU programs specific to each document, reducing the generation of temporary bitmaps and enabling real-time editing of complex files.

Thomas was not alone in this ambition. Photon Studio, another AI-coded alternative, had launched with claims of spending about two thousand dollars on model calls and using GPT-6 Astra to build a free photo editor; it attracted 170 active users on its first day, though the developer acknowledged initial bugs and instability. Concat, built with Claude Code, positioned itself as a partial alternative to CapCut, focusing on running automatic subtitles and text-to-speech locally to eliminate subscription fees. GenOffice aimed directly at Microsoft Office, integrating AI agents into the editing process for Word, Excel, PowerPoint, and PDF files.

AI coding had fundamentally altered how the industry measured completion. Building a window, toolbar, layers panel, and several filters used to consume months of grinding work. Now these visible components could materialize in days, responding to clicks, opening files, exporting results. The moment when software first looked like a finished product had been radically compressed. When bugs appeared, developers could describe them to an AI agent, which would fix them in minutes and pass tests. After dozens of iterations, it became easy to believe that remaining problems were simply a matter of running more cycles and consuming more tokens.

But mature software's hardest problems were not on any feature list or repository issue tracker. They were exceptions born from years of user feedback, sometimes from scenarios no one had anticipated yet. These required insight to discover and solve—not more tokens, not more iterations. Adobe's engineering director had recently written that rewriting Photoshop from scratch was almost impossible and unnecessary; the company was instead reshaping it from within, gradually adding and updating engines based on existing tools. When Apple released the iPad Pro in 2016, Adobe had ported the existing Photoshop codebase to mobile and web rather than rebuilding. The difference between a tool that looked like Photoshop and a tool that could reliably open a complex PSD file made ten years ago, edit it, save it, and have it still work in Photoshop was not a matter of degree. It was a chasm that no amount of AI-assisted coding had yet crossed.