The prompt for Claude Opus 5.5 was three sentences long:
I have a tool to work with Markdown on my computer in the dir markdown. I want to have the same as HTML version. make it really good, iterate and spend a lot of tokens until you are proud!
The tool in that directory was iA Writer. I had downloaded it from the developer's website and dropped it into a working folder, because I like an experiment to live in one place. The prompt never names it. The agent had to work out from the installed app what it was being asked to rebuild. An installed app turns out to be the most precise brief there is. In the terms of the Omakase principle, it fixes the outcome and the evidence of success in one bundle and leaves the method entirely to the agent.
About an hour later I had a browser version with most of the features working. Ten more minutes produced a native Mac app: 2.2 MB, ad-hoc signed, documents stored as plain .md files, a real menu bar, save on quit. It was not a perfect clone, as there were features missing: the native print and save dialogs, the Intel build. However, in the end, I got a working clone of an App that I really like. IA Writer has been on sale for fifteen years, its makers have built a whole design philosophy around leaving things out, and in 2025 it was an Apple Design Award finalist. People buy it because someone made a thousand small decisions on their behalf.
Thirteen giants and one small app
I ran the experiment because of ArtCraft. In early October its founder, Brandon Thomas, published Rust reimplementations of seven Adobe applications, written with Claude. The lineup has since grown to thirteen: stand-ins for Photoshop, Illustrator, Premiere, Lightroom, Acrobat, After Effects and InDesign, then Word, Excel, PowerPoint, Figma, Pro Tools and AutoCAD Thomas told commenters he had "one-shotted Photoshop". Ars Technica put his other line, "Software is over", in its headline.
The repositories, however, tell a quieter story. On ArtCraft's own page, three apps are marked early alpha and the other ten are still in development. PhotoCraft's README concedes that it is not yet a Photoshop replacement for daily professional work, and in mid-October its tracker listed about 1,500 open issues. Don't get me wrong: it's an impressive feat!
The law guards against selling, not building
Most of the coverage stops at the law. ArtCraft calls its apps clean-room reimplementations, and PhotoCraft's README says it was built from public specifications and observed behaviour only. The traditional clean room used two teams, one studying the original and one writing code without ever seeing it. Here there is one developer and a model trained on most of the public code ever written, and no one independent has verified the claim. According to tbreak's summary, Ars Technica's reading is that rebuilding functionality this way is generally lawful, while copying a recognisable interface invites trade dress claims. TweakTown points to Adobe's patent portfolio. Adobe has not commented in any coverage I could find.
ArtCraft also does not sell the clones. They are MIT or Apache licensed, and Thomas plans to fund the work by selling token-based access to ArtCraft's own image model. Selling a clone is where the legal exposure concentrates, and the people building them seem to know it. Copyright, trade dress and patents all guard against the same thing: someone else selling your work. They have little to say about a customer who builds their own. My clone of iA Writer will never be sold or leave my machine, so it will never see a courtroom, and nothing in the law stops it from replacing a purchase.
Who buys an app that is a prompt?
As someone who has built and sold apps, I find this both fascinating and threatening. When anyone can clone your app, who buys it? Small vendors of specialised software are the most exposed, and the reason is scope. A narrow tool with a few hundred well-chosen features is exactly the size an agent rebuilds in an afternoon. The niche that protected them from large competitors is what makes them cheap to copy.
Large vendors will sue, and they have the lawyers and the patents to do it. But litigation works on the supply side, against people who distribute copies. The threat sits on the demand side, in thousands of firms that will never ship a copy of anything. Klarna showed the pattern in 2024, when it moved off Salesforce and Workday onto systems it built itself. Adobe can take ArtCraft to court. It cannot take its customers' build folders to court. When the software you sell is, in effect, a prompt for Claude or ChatGPT, the price you can charge drifts towards the price of the tokens it takes to rebuild it.
Maintenance becomes a token bill
The standard objection is ownership. Building your own means you deploy it, and you fix it every time the operating system shifts underneath it. A vendor sells you that responsibility along with the code, and most buyers are glad to pay someone else to carry it. That objection assumes the builder walks away once the app runs. The agent that wrote the clone does not have to. It can stay on as a sentinel that reads the logs and patches what breaks.
The transcript of my experiment with Claude Code already shows the habit in miniature. When the preview server died between sessions, the agent restarted it. When its first test of save-on-quit failed because the editor ignored a simulated command, it switched to a synthetic paste event and ran the test again without being asked. The sentinel does not remove the cost of maintenance. It turns it into a token bill.
The labs pay cost price
Once maintenance is a token bill, the question changes. It is no longer whether you can build software, but how many tokens you can afford to spend building and keeping it. My clone cost an hour of a frontier model running flat out on a prompt that told it to spend lavishly. ArtCraft's thirteen apps cost far more. Everyone in that market pays retail for tokens, except the labs, which pay cost. Their compute is not unlimited and it is not free. But their marginal cost of cloning any application on the market is lower than anyone else's, and they already run the agents that do it.
If cloning is a token budget, the largest budgets belong to the people who sell the tokens. In December 2024, on the BG2 podcast, Satya Nadella described business applications as CRUD databases with a lot of business logic, and said that is where they would "collapse" in the agent era. The line was widely shortened to "SaaS is dead", which he never said. His claim was narrower: agents would absorb the logic layer, and the applications underneath would thin out. I asked at the time whether this was the end of SaaS as we know it.
Two years later the claim looks conservative. Agents did not stop at the logic. They reproduce the whole application, interface included, from an installed copy and a sentence of instruction. Nadella was talking about business apps such as Microsoft's own Dynamics. What arrived applies to every app, Word included.
Fifteen years in an hour
When iA Writer was named a finalist, its makers wrote that the only way to tell whether a design is good is whether it stands the test of time, and that being picked after fifteen years in the App Store was good news for design that lasts. They were right about the fifteen years. That is exactly what my clone copied, in an hour, for the price of tokens that never reached iA. My clone carries fifteen years of iA's decisions and cannot make the next one. I do not know who pays for the next fifteen years once the first fifteen are free, and I have not heard anyone explain it yet.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.