~/definitionOfDone.dev/blog/artcraft-adobe-and-the-premium-software-moat.md

Elmer Botha · · 11 min read

AI Rebuilt Adobe's Suite in Ten Days. Here's What It Couldn't Copy.

Last week, someone used AI to rebuild Adobe's creative suite and gave it away. The ArtCraft project released seven free, open-source applications, each aimed squarely at an Adobe…

TL;DR

  • ArtCraft released free, open-source rebuilds of seven Adobe apps (Photoshop, Illustrator, Premiere Pro, After Effects, Lightroom, Acrobat and InDesign), written in Rust largely by Claude Opus 5.5. The Photoshop rebuild went from first commit to over a thousand merged pull requests in about ten days. Within days the project had tens of thousands of GitHub stars.
  • That's real, and it matters. Copying a mature product's features used to take years and a funded team. Now it takes weeks and an agent budget. "We have more features" has stopped being a moat.
  • It isn't AI slop. I've tested it, and it works remarkably well. The repos back that up: PhotoCraft alone has more than 6,800 tests, real-world PSD files tested on every pull request, fuzzing, performance budgets enforced in CI, and unsafe code forbidden outright. That's an engineering discipline, not a demo.
  • The project's own progress documents are candid about what's still open. The menus are close to 100% complete. Each app's estimate of "ready for real work" sits between 45% and 62%. The gap is file compatibility, reliability, feel and ecosystem.
  • ArtCraft's business model gives the game away. The editors are free; the money is meant to come from selling AI generation credits. The value has moved from the feature list to the compute behind it, which is where Adobe has been heading too.
  • The companies most exposed aren't the giants with format gravity. They're premium products whose only defence is a feature list.
  • If you build or buy premium software, ask what's left once someone can regenerate the interface from screenshots and documentation. That's the actual product.

Title card reading "Adobe Got Cloned …and it's not slop", with an "Open source" label pointing to a row of seven ArtCraft app icons, each a sketched animal.

Last week, someone used AI to rebuild Adobe's creative suite and gave it away.

The ArtCraft project released seven free, open-source applications, each aimed squarely at an Adobe product: PhotoCraft for Photoshop, VectorCraft for Illustrator, FilmCraft for Premiere Pro, EffectCraft for After Effects, LightCraft for Lightroom, PrintCraft for Acrobat and DesignCraft for InDesign. They're native Rust, MIT or Apache licensed, with no account, no subscription and no telemetry. Most coverage credits Atlanta developer Brandon Thomas, who says he built them with Claude Opus 5.5; Futurism describes a four-person team and quotes co-founder Michael Caffrey calling Adobe's cancellation practices "scummy." PhotoCraft alone reportedly passed 20,000 GitHub stars within hours, and Linux Compatible counted about 43,000 by 11 October.

The speed is real, and it's in the commit history. PhotoCraft's first commit was on 30 September 2026. By 10 October it had 1,004 merged pull requests, with a median of 3.4 hours from opening a PR to merging it. Reviewers who tried the apps (Gizmodo and Tedium among them) found the same thing I did when I tested it: it looks and works enough like the real thing to be unsettling. The rough edges they hit are the kind a codebase moving this fast closes in days, not years.

Both halves of that verdict matter. The interesting question isn't whether ArtCraft replaces Adobe for every professional this year. Thomas himself has said the apps are still early, and the roadmaps say the same. The question is what this does to the economics of premium software, and the evidence for that is unusually good, because ArtCraft publishes its own scorecard.

What got copied

Start with what ArtCraft proves, because it's significant.

A credible Photoshop alternative has always been possible in principle. GIMP has existed for decades, and Affinity showed a well-funded team could build a serious competitor. What stopped most challengers was cost: years of engineering to reach the point where a professional opens the app and finds their tools where they expect them.

ArtCraft collapsed that cost. PhotoCraft implements all 628 of Photoshop's menu items. VectorCraft has 348 of Illustrator's 375 wired up. LightCraft's checklist of Lightroom features sits at 81%. Gizmodo's reviewer described PhotoCraft as feeling like "glitchy, but real, Photoshop," which is a remarkable sentence to write about a ten-day-old codebase.

The design choices are good ones, too. Every menu item, tool and dialog runs through one command registry that the interface, a command line and an AI-agent interface all share, so anything you can click, a script or an agent can also do. The apps compile to WebAssembly and run in a browser. Tedium's Ernie Smith noted they load in tens of megabytes rather than hundreds.

The uncomfortable lesson for every premium software company is that a feature list is now cheap to reproduce. If your product's value can be described by its menus, an agent can be pointed at the menus.

This is not AI slop

"Vibe-coded" has become shorthand for software that demos well and falls over on contact. ArtCraft gets lumped in with that, and it shouldn't be.

I've installed it and put it through real work, and it works pretty damn well. Tools are where you expect them, it opens files quickly, and it's fast in a way Creative Cloud hasn't been for years. The discrepancies I hit were small, and in a repository moving this fast they're the kind of thing that gets fixed within days. PhotoCraft logged 1,152 commits in its first ten and a half days, and every app in the suite was updated again within the last day.

What convinced me it isn't slop wasn't the demo, though. It was the engineering underneath, which is visible to anyone who opens the repos:

  • Tests as the gate. PhotoCraft has more than 6,800 test functions. Its contributor rules say plainly that every change comes with tests, and format code is covered by round-trip, synthetic-generator, oracle and fuzz tests.
  • Real files, every pull request. A CI job runs PhotoCraft against pinned, checksum-verified collections of real-world PSD files on every PR. A missing corpus fails the build rather than quietly skipping.
  • Compared against the real thing. PSD files normally contain Photoshop's own rendered image, and PhotoCraft diffs its output against it. EffectCraft goes further and drives a live copy of After Effects to compare keyframe behaviour; its first run found a real easing bug (errors up to 12.46 degrees), which was fixed and now matches to within 3e-11 degrees.
  • Performance as a contract. PhotoCraft keeps performance budgets that CI checks for regressions, with a written rule never to invent a number for a scenario that can't be measured yet.
  • Safety by construction. Unsafe Rust is forbidden at the workspace level, and warnings fail the build.
  • Honest scorecards. Every roadmap labels each number as measured or estimated and lists its own gaps in ranked order. Slop doesn't publish its weaknesses.

That's what separates this from a weekend prompt. The agents wrote the code, but someone designed the constraints the agents had to satisfy, and those constraints are the reason the output holds up.

What's still open

None of this means the work is done, and the repositories say so more clearly than most headlines. Each app publishes a roadmap with two kinds of number side by side: how much of the target's interface exists, and an estimate of how "ready for real work" it is.

AppInterface coverage"Ready for real work" (own estimate)
PhotoCraft (Photoshop)Menus 628 / 628~45%
EffectCraft (After Effects)Checklist ≈99%≈45%
DesignCraft (InDesign)99% of its checklist~47% (pre-alpha)
LightCraft (Lightroom)81% of a 517-row checklist~61%
VectorCraft (Illustrator)348 / 375 menu items~62%

The gap between those columns is where Adobe's business actually lives. Reading the gap lists across the seven repos, it falls into four buckets.

File gravity. Professionals don't just use Photoshop; they exchange PSD files with people who use Photoshop. FilmCraft can't yet open Premiere's .prproj project files, and its roadmap prices that at 60 to 120 agent hours. PhotoCraft has its strongest results here, because a PSD file normally stores Photoshop's own rendered image alongside the layers, which gives the team an answer to compare against. Even so, the README is careful to say a re-saved file renders the same but isn't byte-identical to the original. (One launch write-up reported "134 of 135 files round-tripping byte for byte," a figure that actually belongs to the standalone format library, not the app.) For a studio, "opens most of my files" and "opens the file my client sent at 5pm" are different products.

Reliability. Gizmodo found PhotoCraft's Free Transform sometimes undid itself. Linux Compatible reported "noticeable glitches, unpredictable tool behavior, and performance hiccups." None of that is surprising at ten days old, and at the pace these repos move, reported bugs are being closed quickly. But stability at Adobe's scale is earned through years of field reports across millions of machines, and that part takes time no matter who writes the code.

Feel. Tedium found the color pickers "rough, not elegant," and hex color entry that worked in PhotoCraft but not in DesignCraft. A checklist can confirm a dialog exists. It can't confirm a designer's hands will trust it after twenty years of muscle memory.

Ecosystem. Plug-ins, fonts, stock libraries, cloud libraries, team licensing, support contracts, and the confidence that the vendor will still exist in five years. Several ArtCraft roadmaps list plug-in hosting and ecosystem work as needing "owner decisions," which is another way of saying it can't simply be generated.

None of these four is about code volume, which is exactly why AI hasn't collapsed them yet.

Follow the money

The detail I find most revealing is how ArtCraft plans to pay for itself. The editors are free. According to Linux Compatible and The Open Letter, revenue is meant to come from the ArtCraft AI studio, a credit-based service for generating images, video and 3D through third-party models, priced from about $10 to $60 a month. The site also says you can bring your own models and compute and skip the subscription entirely.

So the free editor works as a funnel, and the paid product is AI generation. That's not far from where Adobe has been taking Creative Cloud itself, with generative credits bundled into its plans. Both companies have reached the same conclusion: the editing features are becoming the cheap part, and the compute behind generative AI is the part people will keep paying for.

That's the real threat to premium software, and it's subtler than "free clone kills incumbent." It's pricing power. Adobe's subscription has always been justified, implicitly, by the depth of the tools. When a free, credible-looking alternative exists for the essentials, that justification gets weaker for the casual and mid-range users at the edges of the customer base, even if no professional switches. It doesn't help that Adobe settled with the US government earlier this year for $150 million over how hard its subscriptions were to cancel. ArtCraft's pitch, "stop renting," lands because the frustration was already there.

The risks on ArtCraft's side

This cuts both ways, and the project carries real risk.

Legal. ArtCraft describes its apps as clean-room reimplementations built from public specifications and observation, and the repos forbid copying Adobe's code, assets or profiles. But Adobe holds patents on workflows, not just code, and Linux Compatible flags trade-dress concerns for apps designed to look this familiar. Tedium compared the project's posture to Napster: not legal, but it still changed everything. I'm not a lawyer, and nobody yet knows how this will play out.

Sustainability. Seven apps chasing seven of the deepest products in software is an enormous maintenance surface. The repos' own estimates of the remaining work run to thousands of agent hours. Agents make that cheaper, but not free, and a credit-based AI studio is an unproven way to fund it.

Perception. Several outlets advised caution about running largely AI-generated software, and Smith noted that ordinary users won't audit the code. The engineering evidence above is a strong answer, but most buyers will never read a CI file. ArtCraft's real challenge here is reputation, not quality.

And the "software is over" framing has already softened. Thomas initially suggested full parity within a month, then revised that on Hacker News to "99% parity taking months rather than years." Months is still remarkable, but it isn't over.

What it means for premium products

The companies most exposed by ArtCraft aren't Adobe. Adobe has file formats the industry is built around, a huge ecosystem, enterprise contracts and its own AI models. It will feel pressure on price at the edges, not an existential threat this year.

More exposed are the premium tools one tier down: products whose value is a well-built feature set, without a proprietary format everyone exchanges, without an ecosystem, and without data or services behind them. For those, ArtCraft is a preview. The question their buyers will start asking is, "What am I paying for that an agent can't rebuild?"

For anyone building or buying software, the useful exercise is to run that question against your own product:

  1. What survives a feature clone? List what's left once someone regenerates your interface from screenshots and docs. Formats others depend on, data, integrations, reliability history, support and compliance usually make the list. Menus don't.
  2. Where is your switching cost, and is it earned or imposed? File gravity and ecosystem are earned. Cancellation fees are imposed, and ArtCraft exists partly because users resented them.
  3. Where is the value moving? If it's moving toward compute, models and data, as ArtCraft's own business model suggests, then that's where the product strategy needs to be.

AI didn't make Photoshop free. It made the parts of Photoshop that fit in a screenshot free. Everything that doesn't fit in a screenshot (the files, the trust, the ecosystem, the years of edge cases) is what premium software has left to sell. The vendors that know exactly what that is will be fine. The ones that were really selling a feature list just found out.

newer REA - Reverse Engineer Anything