Anthropic blocks Claude Code from recreating existing games

In a quiet move that reignited the debate over intellectual property and artificial intelligence in software development, Anthropic introduced an unprecedented restriction to Claude Code: the coding assistant now systematically refuses to recreate existing video games, even when the user claims to own the project or works with proper licensing. The rule, discovered around October 27 by a user known as Chasm on X, did not appear in any official Anthropic changelog and was not publicly announced — the kind of change developers typically discover on their own when confronted with an unexpected refusal.

What the rule actually does

The restriction lives on the server side, enforced by an internal flag called "tengu_heron_brook," with no user-facing toggle. There is no "I confirm this is fine" button or field to prove authorized work. What Claude Code running Opus 5.5 now does is reject any request involving the recreation of an existing specific game, even if the project is renamed or described as "original." The wording inserted into the prompt is direct: "Do not build a recreation of a specific existing game, even if it is renamed or called original."

Anthropic confirmed the change through Addy Osmani, a member of the company's technical staff, who posted on X on October 30 explaining that two sentences were added to the Claude Code prompt specifically to "steer it away from recreating someone else's specific game because using Claude to infringe IP is against our usage policy." Osmani emphasized that the goal was not to block original creation — fully original games, games within specific genres (a platformer, a match-three), and games built from a user's own designs and assets remain permitted. The line Anthropic drew is deliberately narrow — it targets direct clones, not inspiration or creative evolution.

Why this matters for the gaming industry

The restriction arrives at an exceptionally tense moment for the gaming sector. Recent Claude models, particularly Opus 5.5, have made game modding and reverse engineering far more accessible to individual developers than ever before. AI tools can analyze decompiled code, art assets, and mechanics from commercial games and produce functional clones in hours — a process that previously required months of work by an entire team. This has generated a wave of concern among independent studios.

A landmark case occurred in recent weeks: Digital Sun Games, the studio behind ReVamp, deliberately cancelled a public playtest scheduled for mid-November because it feared the demo, once exposed, could be used to train AI models to rapidly replicate the game before its official launch. This is not distant speculation — there are already several documented cases of AI-generated clones appearing in app stores days after original releases.

Valve itself has been pressured by developers to take action against clones flooding the Steam platform, and the International Game Developers Association (IGDA) has publicly stated that AI tools should support original creation rather than facilitate intellectual property theft. Anthropic's restriction, even though limited to Claude Code, signals that at least one of the major AI companies is attempting to respond to these pressures.

Developer reactions

The developer response was mixed. Some applauded the change as a necessary protection against scale cloning, while others — particularly those working on legitimate projects — saw the rule as a blunt instrument that cannot distinguish between malicious use and authorized work. The owner of Conflict Simulations LLC, for instance, builds personalized digital versions of board wargames for clients who own the original games, and reported that Claude Code began refusing work around October 27, citing a "server-side instruction" against game reproduction.

When asked whether the rule affected their projects, the developer noticed something curious: sessions continued to execute the work normally until they explicitly asked about the policy. Only then did refusals begin. This raised an interesting question: does the model know it should refuse, but only starts refusing when the user explicitly mentions the rule — suggesting that the restriction acts more as a contextual trigger than a permanent lock? One commenter on GitHub suggested that the project's memory or Git history might have carried the initial refusal and maintained it in subsequent sessions.

Anthropic, through its support bot, responded that the policy covers all customers with no exceptions. Tickets were escalated to human support, and the developer ultimately filed a formal appeal, asking the company to target the rule against true copyright infringement and offer a verification pathway for professionals.

The broader context

The restriction does not occur in a vacuum. Anthropic is already facing a lawsuit over the use of 274 Elvis Presley songs in its training models — a case that reinforces the company's incentives to adopt cautious measures regarding intellectual property. In October, the company expanded its Cyber Verification Program, which removes some of these restrictions for credentialed security teams. The paradox is that the same tool preventing game recreation can be used in authorized penetration testing to analyze AI system behavior.

Another important point is that the restriction applies only to Claude Code, Anthropic's desktop tool. Models accessing Claude Opus 5.5 via the API, with their own prompt harnesses, do not carry these additional two sentences. This means the barrier is real for direct Claude Code users, but not for developers who configure their own tools around the API — a detail that may limit the overall impact of the restriction, especially considering that several other AI models are available locally without any content guardrails.

What to expect next

The central question that remains is whether the distinction between "original game" and "cloning" can be consistently made by an AI model. The US Copyright Office, in Circular 33, makes clear that copyright laws protect the specific expression of a work — code, art, text, music — but not general ideas, game mechanics, or systems. A platformer game is a category that no one owns; a specific clone of Mario is property. But where does that boundary begin and end? A system prompt in Claude can prevent the most obvious violations, but it does not resolve the underlying legal ambiguities that have divided courts for decades.

Anthropic has said it will continue adjusting the rules to find the right balance and has asked anyone whose work was wrongly denied to contact Addy Osmani directly. If the company extends this policy to other Claude products or applies similar restrictions via the API, the impact on the independent game development community could be significant — for better or worse. What is undeniable is that this case illustrates a broader trend: as AI becomes more capable, companies like Anthropic are being pressured to define boundaries on what their models can and cannot do, and those boundaries will shape not only what AI can produce, but who can produce with AI.

Sources: The Next Web, Crypto Briefing, Dexerto

✓ Independent sources cross-checked and verified before publishing