AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Developer behind the AI-Built Adobe Clones Says They Are “Not Anywhere Close to Ready"

Collected Oct 22, 2026

Brandon Thomas published AI-built, open-source replacements for seven Adobe applications and then walked the claims back on Hacker News. "I wasn't ready to post this to HN," he wrote, adding that the work is not anywhere close to ready; he said he would post again once the project clears beta. Days earlier, announcement posts across five subreddits had projected a beta and roughly 100% parity within a month. The repository's own roadmap lines up with the more cautious Hacker News message rather than the Reddit pitch.

Thomas built the applications with Claude Opus 5.5 in Rust, with WebAssembly builds for the browser, mirroring Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign and Acrobat Pro. Everything ships under the MIT and Apache licenses, so the code can be reused commercially with attribution.

The numbers in the repository are the most concrete part of the story. PhotoCraft's ROADMAP.md, updated October 11, puts feature breadth at roughly 76%, derived by counting coverage: 628 of 628 menus, 56 of 68 tools, 30 of 35 panels, and 13 plus 4 of 31 formats. Readiness for real work is a separate and much lower figure, about 45%. A mainstream practitioner gets 43%, while someone using only core features gets 61%. The roadmap estimates 1,100 to 1,800 hours to beta and 2,300 to 4,100 to full parity, counted as Opus 5.5 agent wall-clock hours for one agent working sequentially. The comparison target is named precisely: Photoshop 2026, version 27.11.0, measured from the installed application on October 10.

The roadmap is harsher on the project than its critics. Users of versions 0.5 and 0.6 have filed 579 open issues. Twenty-two of twenty-five performance budgets are missed. And the document states that PSDs the project writes are not yet reliably accepted by Photoshop. For software positioned as a Photoshop replacement, round-trip file compatibility is not a cosmetic gap; PSD interchange is how most workflows hand files between designers, retouchers and print vendors, so a clone that cannot reliably save a file Photoshop opens has a hard ceiling on professional use regardless of how many menus it covers.

The pace, at least, is striking. The first commit landed September 30 and the latest on October 13. In those fourteen days the repository recorded 2,152 commits, 1,004 merged pull requests and 539,017 lines of Rust across 1,306 files. Git shows 317 distinct contributor identities, so the solo-developer framing stopped being accurate almost immediately. Claude appears as a named commit author 277 times.

People who installed the apps split roughly along the roadmap's line. One Reddit commenter, Painted-Arcana, wrote that the Photoshop clone is bad, that alt-click eyedropper does not work and that layouts cannot be changed much beyond static presets, concluding after about two minutes that it is not a Photoshop replacement. Photographers ran into the RAW pipeline: one commenter reported noise and banding when raising shadows on a Sony a7c file, and another found no RAW support for several Nikon bodies. The simplest app fares best; on Hacker News, commenter Krasnol reported PDFCraft working without bugs. In practice, this pattern is typical of breadth-first builds: menu coverage is cheap to generate, while color science, codec support and interaction polish are where the effort concentrates.

The repository also documents substantial process, which much of the coverage overlooked. AGENTS.md carries a clean-room rule stating that the team studied Photoshop and other proprietary editors for behaviour and look only, never copying code, shaders, profiles or assets, and implemented from public specifications such as the Adobe PSD spec, ICC profiles, ISO 32000 blend modes and published papers, plus observation. Tests gate every change, a panic_hunt test fuzzes every command with adversarial parameters, and format work is checked against a corpus of Photoshop-authored PSDs. The apps expose a JSON control channel and an MCP server so the coding agent can drive the running application and screenshot it to verify its own interface work.

That clean-room framing is where opinion divides. Commenter kmeisthax was blunt: "There is no such thing as clean-room vibe coding." The doctrine conventionally assumes two separated teams, one writing specifications and another implementing without seeing the original, so a single agent iterating directly against the installed application is a loose fit for it. This is a legal and methodological question rather than a technical one, and the practical risk for anyone adopting the code is that the provenance argument is thinner than an established clean-room process would provide; that is my reading, not a claim in the project's documentation.

The most measured reaction came from someone running a similar experiment. Commenter az_liberal_geek is building Hew, an open-source SketchUp replacement, also with Claude, and described the first version as fast and impressive but very buggy and failing in unexpected ways. Adding a major feature takes a prompt plus several hours of Claude work, followed by many more hours of the developer's own testing and refinement. He put parity at a year rather than a month, which lands near the repository's own estimate.

Thomas has since adjusted his public figure, telling Hacker News that 99% parity will take a while but will be measured in months and not years. That still leaves the timeline in dispute, since the repository's own 2,300 to 4,100 agent-hours assume one agent working sequentially and say nothing about the human review time that the Hew developer describes as the dominant cost.

Why it matters: teams evaluating AI-generated replacements for commercial creative tools should read the readiness figure, not the feature count, and should test file interchange before committing. For developers, the more transferable lesson is the working setup: agent-drivable test harnesses, fuzzing and a reference corpus are what turn fast generation into something auditable. Expect more projects of this shape, and expect their public claims to be corrected by their own roadmaps.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

A developer used Claude Opus 5.5 to build open-source replacements for seven Adobe applications, promoted them across Reddit with a projection of roughly 100% parity within a month, then told Hacker News the project was not anywhere close to ready. Its own roadmap puts it at 45% ready for real work and 2,300 to 4,100 agent-hours from full parity. By Steef-Jan Wiggers