Code Was Never the Product
Every few weeks, one social network or another announces the death of an entire corner of the software industry. The script barely changes: a developer, an AI model, a one-line instruction (“I want a copy of such-and-such software”), and then a triumphant thread. The latest example, shared by Pirat Nation, presents seven free, open-source applications, written in Rust with help from Claude Opus 5.5, as “clean-room” alternatives to Adobe’s flagship tools.
I’m using this suite as an example, not a target. It simply happens to be the most polished of these announcements, which makes it the one that best reveals what all of them leave out. I won’t pass judgment on the quality of its code: I have neither the time nor the pretension, and that isn’t my point. My point lies elsewhere. Code is a very important part of software, but a very important part is not the whole, and it is even less the meeting point between a product and a customer.
What was released
Behind the project is Brandon Thomas, described as a fifteen-year industry veteran, who runs ArtCraft. The method: eight agents running in parallel, which produced seven programs imitating Photoshop, Illustrator, Premiere Pro, Lightroom, After Effects, InDesign and Acrobat Pro, released under the MIT or Apache licenses.
The numbers are impressive. In eight days, the seven repositories racked up more than 42,000 stars on GitHub, more than 23,000 of them for PhotoCraft alone. The same article counts 24 crates and over 1,700 tests for PhotoCraft, and 35 crates and over 1,600 tests for FilmCraft. Nothing to quibble with on volume: it is a considerable body of work, turned out in a span of time that would have been unthinkable three years ago.
The author, to his credit, hides nothing about where things stand: the badges read “early alpha” or “in development,” and he is counting on bug reports and community contributions to stabilize the whole thing. The problem is not the project’s honesty. It is the story people tell about it.
A prompt is not a specification
“Build me a Photoshop” is a remarkably precise instruction about everything you can see, and a completely empty one about everything you can’t. A model can reproduce menus, panels, layers, masks and shortcuts, because all of that is documented, demonstrated in thousands of tutorials, photographed from every angle. What it cannot guess is everything that was never written down anywhere: why a given behavior exists, which customer it was built for, which corrupted file from 2009 it still manages to open.
That is the whole difference between the vibe coder who has mastered a prompt and the engineer who has mastered the ins and outs. The first reinvents the wheel by watching it spin; the second knows what road it runs on, what loads it carries, and what breaks when you swap it out. A few months ago I wrote that nobody reads the code they deploy anymore. These announcements go one step further: nobody reads the software they claim to replace, either.
The number that gives the game away
One detail in ArtCraft’s messaging says it all. According to the team, it would take another 140 to 220 agent-hours, or 25 to 40 hours of real time, to reach 95% feature parity with Adobe’s tools. Forty hours to catch up with thirty-five years. It’s an appealing figure, and above all an optimistic one, judging by the project’s own documents.
Those documents, to give them their due, are unusually candid. For PhotoCraft, they report 625 out of 625 menu items wired up and overall feature parity of roughly 86%, yet only 115 of 170 PSD files rendered pixel-perfect, and readiness for everyday professional work estimated at 25 to 35%. It’s all right there: every single menu responds, and barely a third of real work is possible. The gap between “menus wired up” and “ready for real work” is not an outsider’s criticism; the team measures it themselves.
Feature parity measures a list of checked boxes. It does not measure software. What’s missing isn’t a handful of forgotten features: it’s real-world files, edge cases, odd behaviors at the margins, in other words exactly what makes a professional willing to trust a tool with a full day’s work. The bug tracker already shows it: among the 447 open issues on PhotoCraft are a crash when opening a 32-layer PSD and PSD files that simply refuse to open.
I’ve written before, on an entirely different subject, about how a rising metric can come out of a broken measurement setup and look exactly like genuine progress. A parity percentage is precisely that: a number that goes up and tells you nothing about what matters.
A file format is a territory, not a specification
PhotoCraft’s README states that its PSD support is based on the specification Adobe has published. That’s rigorous, and it isn’t enough. A specification describes an idealized format. It does not describe thirty-five years of files produced by dozens of successive versions, third-party plugins, sloppy exports and bugs that became de facto standards. The real PSD format is the world’s stock of PSD files, not the document that defines it.
The same goes for page layout. A print shop doesn’t expect “a PDF”: it expects a compliant PDF/X, with properly managed color profiles, correct bleeds and embedded fonts. Designers don’t choose InDesign for its menus; they choose it because their printer, their agency and their clients can all read what comes out of it.
Reality has its own way of answering perfect redesigns. I described it in the case of IPv6, the Esperanto of networking: a technically superior solution, born in 1998, that the world stubbornly refuses to adopt. In software, you don’t beat the incumbent by being right; you beat it by being compatible with it.
Codecs, or the law before the code
FilmCraft made a bold choice: to do without FFmpeg and ship its own H.264, HEVC, ProRes and AAC codecs. On paper, that’s Rust purity all the way down. In practice, it means taking on two mountains at once.
The first is legal: these formats are covered by patents, some of them bundled into licensing pools, and the question isn’t whether you know how to write them but whether you have the right to distribute them. The second is technical: FFmpeg and hardware encoders represent two decades of optimization, edge-case fixes and compatibility with millions of malformed files. Anyone who has spent time tuning a constant-quality encode knows that a codec isn’t judged by what it does with a demo clip, but by what it does with the worst source you can throw at it.
GIMP, Inkscape and Scribus already gave us the answer
Here is the argument that, to my mind, settles the debate. If code were the wall protecting Adobe, it would have come down long ago. GIMP has been around since the mid-1990s, Inkscape and Scribus since 2003. They are free, open-source programs, built by real teams over decades, by contributors who knew their subject intimately. They are respectable, often excellent in their own niche, and they remain marginal in professional studios.
It isn’t for lack of code. It’s that Adobe’s customers aren’t buying code. The only alternatives that have genuinely gained ground, such as Affinity or DaVinci Resolve, did so with companies behind them, years of investment and a patient compatibility strategy. I’ve detailed elsewhere, regarding Adobe’s pricing policy, how the giant itself keeps pushing customers toward the exit. The door has been open for years: a prompt isn’t what will get anyone through it.
Habit is the real asset
ArtCraft makes a telling choice: the interface deliberately copies Adobe’s menus, panels and shortcuts, so that a Photoshop user feels instantly at home. That’s an admission. If copying the look is the main selling point, then Adobe’s true asset is not its code but the habits of tens of millions of people.
Those habits were built by schools that teach Photoshop, job listings that require it, thousands of presets, brushes, scripts and plugins, tutorials in every language, and forums where every problem already has an answer. It’s a network effect, and I know from experience how much it weighs. I ran FreeBSD servers in production before finally facing facts: the Linux ecosystem crushed the elegance of the BSDs, not through technical superiority but through sheer numbers. You don’t copy a network effect by copying menus.
The knowledge you can’t prompt
LightCraft promises a photo library and a RAW developer. It is probably the most telling case. A credible RAW developer depends on thousands of camera and lens profiles, measured one by one and updated with every new camera release, and on a color science that owes more to the lab than to the algorithm.
This isn’t a matter of lines of code but of accumulated data, physical measurements and deliberate aesthetic choices. I devoted an entire post to Sony’s color science and the Minolta legacy: even a giant that makes the sensors doesn’t always manage to render skin tones correctly. That kind of knowledge isn’t in any training corpus, because it was never written down in a form a model could copy.
Who pays for maintenance five years from now?
Professional software isn’t a deliverable; it’s a long-term commitment: fixing bugs, keeping up with new operating systems, processors, formats and cameras. On that front, ArtCraft’s model deserves a careful read. The team plans to fund development by selling access to its own AI model, offered inside the applications alongside third-party models. And this isn’t a foundation or a volunteer collective, but a company specializing in AI-generated imagery.
There’s nothing illegitimate about that. But it changes the nature of the thing: the free suite becomes a distribution channel for a paid service. We know how this works, as I described regarding AI loops and the dependency that comes bundled with them: the tool is open, the brain lives elsewhere. The day the revenue doesn’t follow, who will maintain seven applications, homegrown codecs and a PSD engine? No professional bets their livelihood on a question with no answer.
What these projects genuinely bring that is new
Let’s be fair, because not everything here is a mirage. These applications are native, with no Electron or web views, and most of them also run in a browser. Above all, PhotoCraft exposes more than 500 command-line commands, also available through an MCP server, the protocol that lets an AI agent drive an application.
That may well be the real story, and it’s a fascinating one: the target customer is no longer just the designer, but the agent that will work on their behalf. Adobe never designed its tools to be driven by a machine, and that market is genuinely new. It echoes the shift I noted with EmDash, the CMS Cloudflare built for agents first. On that terrain, starting from scratch is an advantage. For the designer who has to deliver a printed catalog by Friday, it’s a handicap.
The objections I expect
Some will object that Linux eventually beat the proprietary Unixes, and that open source always wins in the end. True, but Linux won through thirty years of community, maintainers, companies that invested, and painstakingly built compatibility. Not through an eight-day sprint.
Others might hold my own writing against me: I’ve argued that Odoo’s complexity is commercial rather than technical. I stand by that, and it isn’t a contradiction. I was talking about deployment, which is made artificially intimidating in order to sell integration services. Here, I’m talking about what the product is worth once it’s deployed. The first kind of complexity is a barrier someone puts up; the second is value that accumulates.
Scarcity has moved
Brandon Thomas foresees a world where any developer will be able to recreate paid software as open source in order to “take back the internet and own tech forever.” He’s right about one essential thing: code is becoming abundant. I wrote as early as January that the marginal cost of programming is collapsing, and these seven repositories are a spectacular demonstration of it.
But when a resource becomes abundant, scarcity doesn’t vanish; it moves. That’s the thesis I recently defended regarding the age of abundance Musk keeps promising, and it applies here point for point. Scarcity now lives in compatibility with the existing stock, in data that is measured rather than generated, in the right to distribute, in the confidence that a tool will still be around and still maintained five years from now, in the habits of millions of users.
None of that can be prompted. You can ask an AI for a copy of Photoshop. You can’t ask it for a copy of thirty-five years of customers.