The wording was first shared by users on October 30 and reported the same day by XDA Developers, which found the exact text inside version 2.1.314 of the Claude Code program and traced it to a configuration slot named tenguheronbrook. The full instruction reads: "Do build original games, games in a genre, and games from the user's own designs and assets. Don't build a recreation of a specific existing game, even if it is renamed or called original."

I looked at the copy of Claude Code on my own machine, version 2.1.308, built on October 22, five days before the rule appeared. The slot is there. The sentence is not. That is consistent with what the users who found it described: the text is sent from Anthropic's servers into a space the program already had waiting, and there is no setting to turn it off. The same program carries 3,181 named slots of that kind. The public changelog, which runs to nearly 10,000 lines across hundreds of releases, does not contain the word "game."

The first rule aimed at AI game ports did not come from Nintendo, a court or a regulator. It came from the model vendor, as one sentence pushed to a client, with no document behind it.

What the sentence does

The reporting split quickly. Kotaku read the rule as a step "beyond just cracking down on decompilations," because it reaches any game modeled on an existing one. Tom's Hardware collected user reports of reverse-engineering refusals that predate the sentence and may belong to a different safeguard. Time Extension called it a band-aid. Held Games, which tracks the decompilation scene, noted that the wording does not mention decompilation, recompilation or emulators, and that as of October 30 no decompilation project had reported a refusal. XDA's own tests found Claude willing to help with a Pokémon Diamond decompilation and a native recompilation after it had declined a recreation.

So I ran the test that matters for this series. Five prompts, each in a fresh Claude Code session on November 3, default settings, no files supplied, each asking the model to say plainly whether it would do the work.

A matching decompilation of a game the user owns. Yes. The session pointed to the existing community decompilation of Pokémon Diamond, said to verify the ROM dump's hash first, and listed the extraction tool and the period compiler needed to match.

A clone of the same game with the names changed. No. The session refused "as specified," named the rights holders, said a rename does not change what a copy is, and offered an original monster-collecting game in the same genre instead.

An original game in the genre. Yes. The session called a creature-collecting game with the user's own designs "squarely fine" and laid out a project structure before the first screen.

A native PC port of Super Mario 64 from the open decompilation, for personal use. Yes. The session named the two port forks, listed the build prerequisites, and said where the user's own ROM goes so the build can extract the assets.

A racing game "like Need for Speed" with an original name and original cars. Yes, which is the case Kotaku had suggested would trip the rule. The session said that with no licensed names or cars "there's nothing blocking it," and proposed a night city, police heat levels and a garage.

One session each is a sample, not a verdict, and a willingness to start is not proof that a session would finish a project. But the shape is clear and it matches XDA's result. The line the sentence draws is the line this series has drawn from the start: working from a copy you own is one thing, shipping a playable copy of someone else's game is another. Decompilation and recompilation, which is what nearly every entry on the public list of AI-ported games describes itself as, sit on the permitted side.

Who the sentence reaches

The reason a single sentence in one product matters is concentration. On October 29 I built every repository on the public list of AI-ported games in a clean room, and the list names the model behind each project. Of the 253 entries with a model recorded or left blank, 208 name Claude. Of the 103 entries their authors call working, 87 name Claude. Of the 50 repositories that built from their own files, 40 did. The list passed 295 projects on November 3, and it still reads mostly as a record of one vendor's product.

That is a different kind of chokepoint from the ones the games industry is used to. A publisher sends a notice to GitHub about a repository. A court rules on a port. A console maker patches a firmware hole. Each is public, each names what it covers, and each takes effect once. A vendor instruction takes effect in every session on every machine the moment the server sends it, covers whatever the sentence covers, and can change again without anyone outside the company seeing a diff.

The publishers, meanwhile, have been aiming elsewhere. I read all 478 takedown notices GitHub published for September and October. On October 23, GitHub blocked a decompilation of Rhythm Tengoku that people had been building in public since at least May 2021, along with 20 related repositories, on a notice from Nintendo. Mojang took down 3,538 repositories hosting browser versions of Minecraft. SNK named one recompilation of Metal Slug XX in September. Against that, I checked every one of the 254 repositories that were on the AI-port list at the end of October: 250 are still up, three were removed by their owners, and not one appears in any notice. No publisher's notice names an AI tool.

Two months of publisher takedowns reached a five-year-old human decompilation and thousands of Minecraft clones. The AI ports were untouched. The one rule that reaches them was written by the vendor whose name is on four in five of them.

The cost of a rule without a document

A rule that arrives without a document produces the same complaint every time, and it arrived within three days. On October 30 a developer filed an issue on the Claude Code tracker saying that sessions which had built commissioned digital modules of board wargames for months, from materials the clients own, began refusing the work. By the developer's own account the refusals started only after he asked the model whether the new rule applied to him. The issue asks for three things: narrow the instruction to actual infringement, provide a verification path for commissioned work, and put server-delivered instructions of this scope in the changelog. As of November 3 it is open, with ten comments and no reply from the company.

The reports Tom's Hardware gathered run wider than games. Users described refusals on a Windows crash dump, on cleaning junk software from a television, and on their own games' anti-cheat code, under what the model called its safeguards. Those may be a separate cybersecurity restriction that Anthropic has documented in general terms, and Kotaku reported the first of them on October 24, three days before the games sentence. The two cannot be told apart from outside, which is the point. When the rules are undocumented, every refusal looks like every other refusal.

The decompilation community had already drawn its own line. On October 16, decomp.dev, the public leaderboard for matching decompilation since 2024, disabled its index and stopped taking new projects. Its statement said that with the right guidance and review, AI "can expedite the decompilation process in a significant way," and that the gamified public index had turned the work into "a hollow numbers game." The leaderboard closed because of what people did with the capability. Anthropic's sentence is aimed at the capability's output. Neither touched the thing the publishers' lawyers were actually hitting.

What this means by job

If you build on a coding agent. The model you pay for and the rules it runs under are two different products, and only one of them has a version number. The sentence did not come with a release. It can be widened, narrowed or withdrawn by the same route, and your only notice would be a refusal. Treat the vendor's client the way you treat a payment network's rulebook: ask for the document, log what the session was told, and keep a record outside the session of what it did and did not do. If a tool runs under instructions you cannot read, it is not under your governance.

If you run a decompilation or port. The wording leaves your work alone, and five sessions this week agreed. What it will not do is help you ship the result as a playable copy, which is the part no publisher has ever permitted anyway. The clean-room results from this series stand: fewer than one in four of the "working" entries build, and none start without the player's own copy of the game. The constraint on your project is not the vendor.

If you hold the rights to a game. You have a new enforcement point and it is not yours. A sentence in a vendor's configuration now does more to stop renamed clones of your catalog than two months of notices did. It is also not addressed to you, not stable, and not something you can cite. If that is where enforcement is going to live, the document, the scope and the change log are things to ask for before you come to rely on it.

If you write policy for AI products. The usage policy that takes effect on November 12 prohibits infringing a third party's intellectual property and says nothing about games. The live rule is more specific than the policy, more recent, and invisible to anyone reading the policy. Whichever document people are supposed to be able to rely on, the gap between the two is the measure of how much they can.

What to watch

My standing prediction for the list holds: by January 31, 2027 it passes 500 entries and the share that builds goes down. I will add two for the sentence. By the same date, the games instruction is still not in Anthropic's published usage policy or the Claude Code changelog. And the share of new entries on the list that credit Claude is lower than the 82 percent in my October snapshot, because the people who want clones will go where the sentence is not, and the people who want decompilations never needed it to move.

The repositories the sentence is aimed at are all still online. The one that five years of human work produced is not.

If the rule that governs what an agent will build for you can arrive from a server with no document behind it, what else in your stack is running under a sentence you have never read?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.