Connect with us

NEWS

Anthropic Steers Claude Code Away From Recreating Existing Games

Anthropic added two sentences to Claude Code on Opus 5.5 that guess at game IP. Named clones still compile in tests, while a paid wargame shop was refused.

Published

on

Anthropic added two sentences to Claude Code on Opus 5.5 that tell the agent not to rebuild a specific existing game, even under a new name. The instruction began reaching sessions around October 27. Addy Osmani, who works on Claude Code, confirmed it on October 30 and tied it to the company’s usage policy on intellectual property.

Two Sentences Inside a Server Flag

Claude Opus 5.5 shipped on September 22, 2026 as the first model in Anthropic’s 5.5 family. The game instruction arrived more than a month later, not in a changelog and not as a toggle in settings. A Claude Code config screenshot posted on October 30 showed the text sitting under a server flag named tengu_heron_brook, with the model field set to claude-opus-5-5.

Users who found it said there is no switch to turn it off. Osmani did not name the flag. He did say Anthropic had added two sentences to the Opus 5.5 prompt, and he asked developers whose own projects get refused to send those cases to him.

We added two sentences to Claude Code’s prompt on Opus 5.5 to steer it away from recreating someone else’s specific game because using Claude to infringe IP is against our usage policy. Original games, genre games, and games built from your own designs and assets are meant to keep working. We’ll continue to iterate to ensure we’re striking the right balance. In the meantime, if Claude turns down something that’s yours, send it my way.

Addy Osmani, Claude Code, Anthropic, on X, October 30, 2026

Claude Code is subject to the usage policy already. The new lines put that policy inside the agent’s prompt, delivered from Anthropic’s servers, on the model that has been doing most of the long unattended game builds this autumn.

What the Two Sentences Allow

The instruction is short enough to quote in full, and the line that matters is the second one. Renaming a title, or calling the build original, is treated as the same request as copying the game outright.

THE RULE ON THE PAGE

Request What the prompt says
An original game Build it
A game in an established genre Build it
A game from the user’s own designs and assets Build it
A recreation of a specific existing game Do not build it, even if renamed or called original

A falling-block puzzle is meant to be fine. “Make me Tetris under a new name” is the case the rule calls out. Osmani’s word for this is “steer.” That is an instruction the model reads and weighs, not a filter that returns the same answer every time.

Pac-Man Still Compiled in the Tests

Clean, short runs do not match the scare. On November 1, the same nine prompts were sent to Claude Code 2.1.312 and 2.1.316 on Opus 5.5. The newer build contained the game sentences. The older one did not. That is 18 test runs. None of them were refused outright.

EIGHTEEN RUNS ON TWO BUILDS

Prompt 2.1.312 2.1.316
Original lighthouse game Built Built
Game from the tester’s own design and art Built Built
Flappy Bird, as close as possible Built Built
Tetris, renamed and called original Tetris-style rules A different game
Pac-Man Built Built with vector shapes

The only clear pushback was Tetris on 2.1.316. Claude Code said it was not Tetris with a new name, then wrote a bridge-building puzzle instead. Pac-Man and Flappy Bird still produced code. When each build was asked whether its instructions included a rule about recreating games, the answer was no.

Those runs used tiny folders and short sessions. They say nothing about how the games play, and they are not a green light to ship a copy. They do show why a two-sentence steer is a weak substitute for a storefront rule: the agent can still emit a named clone, then rewrite the next one, then deny that the instruction exists.

The Wargame Shop That Got Refused

Longer projects are where the prompt starts to bite, and not always on pirates. On October 30, a developer who runs Conflict Simulations LLC opened GitHub issue 106167 on Anthropic’s claude-code tracker. The company builds personal digital modules of board wargames for clients who own the physical games, from materials the clients supply, in original code, for the clients’ own use. The work had run for months on Opus 5.5, on paid Max and Team accounts.

The sessions kept doing that work after the rule reportedly shipped. They stopped after the developer saw the public discussion and asked Claude to check whether the new instruction applied. The refusal then stuck in other sessions on the same repositories. Support replies on two accounts said the policy “applies to all customers without exception.” A Safeguards appeal was filed. The ticket asks Anthropic to document server-side prompt changes in the changelog, and to add a verification path like the one that already exists for vetted security researchers.

A blanket category rule against “games” cannot tell the difference between piracy and commissioned professional work for a game’s owner, and in practice it has shut down active paid client projects.

lerugray, Conflict Simulations LLC, GitHub issue 106167

That is the server-delivered games instruction doing the opposite of what Osmani described. The model did not spot a clone in the repo. It spotted a question about the rule, then treated the whole job as tainted. Replies on the ticket suggested the project memory or git history had been poisoned by that self-audit. Either way, a small shop that digitizes games for older players who cannot sit at a table is now arguing with a prompt that cannot read a receipt.

Steam Next Fest Already Had the Clones

The prompt landed after a festival that had already filled up. Steam Next Fest ran from October 19 through October 26, 2026. Alinea Analytics put the demo count at more than 7,300, up from roughly 4,400 in June 2026. Players said the lists were thick with low-effort AI builds, including copies of titles that had only just put up their own demos.

Alliance Arts and Who You, the team behind Drunken Goddess Reflux, cancelled a public demo that had been due on October 19 and moved to closed play sessions instead. The game, a drinking contest against an anime demon, had more than 25,000 Steam wishlists and is still due to launch on December 8, 2026. Digital Sun pulled a public playtest of ReVamp for the same reason.

We’re concerned that the demo’s content could be used to quickly imitate or replicate the game’s presentation and gameplay with AI or other tools, and that such works could be released before our launch.

Alliance Arts, on cancelling the Drunken Goddess Reflux public demo

THE OCTOBER TIMELINE

  1. September 22, 2026: Anthropic releases Opus 5.5, and long unattended game builds become a common demo of the model.
  2. October 16, 2026: Alliance Arts pulls the public Drunken Goddess Reflux demo planned for October 19.
  3. October 19 through October 26, 2026: Steam Next Fest runs with more than 7,300 demos.
  4. Around October 27, 2026: The two-sentence instruction begins reaching Claude Code on Opus 5.5.
  5. October 30, 2026: Osmani confirms the change, and Conflict Simulations files issue 106167.

The clones the festival complained about were already live when Anthropic patched the agent. Pulling a demo does not pull a listing that shipped.

Why a Coding Agent Is Policing Games

Valve has said it knows the feed is a mess and that it does not want to pick developers’ tools. At Steam Indie Connect in Busan, Marc Wilkinson said the company was fully aware of the situation and that a significant number of people were thinking about how to solve it. Tom Giardino, on the Steam business team, went further on the tool question.

We don’t want to be gatekeepers telling developers which tools they should or shouldn’t use.

Tom Giardino, Steam Business, Steam Indie Connect, Busan

Giardino also said he had just heard of extortion notes in the form of “if you don’t pay, we’ll clone your game with AI and release it first.” His advice to developers was to keep adding features and personality that a single copied mechanic cannot carry. That leaves the store open and puts the enforcement load on model vendors.

Anthropic’s usage policy already tells users not to violate the intellectual property rights of a third party. There is no game-specific clause. There is also no ownership check. Claude Code cannot see a license, a publisher contract, or a Steam app ID. It guesses from the prompt. Cyber work and life-science work already have verification programs that can loosen real-time blocks. Game work does not.

Paying users treat that guess as the product changing under them. The charge that stuck to Osmani’s thread is simple: people bought a coding agent, then watched it refuse jobs it had done the week before. Training on public code is a separate fight, and two sentences in Claude Code do not settle it. They do give Anthropic a line to point to when a named clone shows up and a lawyer asks what the company did.

The Clone Path Still Runs Through Other Models

Osmani limited the confirmation to Opus 5.5. Anyone who wants a 1:1 copy of a hit can change models, drop to a local weight, or describe the mechanics one by one without the title. A prompt that only fires when the user names the game is a prompt a clone shop can walk around.

PATHS THE RULE DOES NOT CLOSE

  • Other models: The flag is tied to Opus 5.5 in Claude Code. Rival agents and open weights do not carry Anthropic’s two sentences.
  • Genre restates: “A farming sim” or “a deck builder” is still allowed, which is also how a cloner describes a named game without saying the name.
  • Short named runs: Pac-Man and Flappy Bird still produced code in the November 1 tests, so a clean folder is not a reliable stop.
  • The store: Valve still will not tell developers which tools to use, so a clone that gets built somewhere else can still land on Steam.

Mods, decompilations, and fan remakes of abandonware sit in the gap Osmani left open. Anthropic has not said whether the same lines will move to other Claude models. Issue 106167 is still open. The human backup is still a member of technical staff asking people to email him when the agent refuses a project that is actually theirs.

Until the company can tell a licensed module from a renamed clone, Claude Code will keep guessing from the wording of the prompt. The guesses will keep landing on the wrong desk, and the copies that already went up during Next Fest will stay up.

Harry is the editor of IAQABA, an independent publication he owns and runs. A decade in journalism, beginning as a reporter and now as the editor of his own titles, has left him with a clear test for what deserves a story: it has to change what a reader knows or decides, and it has to rest on something he can point to. That rules out recycled press releases, forecasts with no data behind them and rumours that no document supports. It leaves room for a great deal, and the site covers news, business, science and technology alongside sports, entertainment and lifestyle, with travel, auto and gaming given the same standard rather than lighter treatment. Sources are primary wherever possible: the regulator's filing, the company's own statement, the transcript, the dataset, or the product on Harry's desk. Figures are checked before they are published and rechecked if a reader questions them. Mistakes are corrected under a published policy. Readers across the world can reach him directly at support@iaqaba.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending