AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Nov 2, 2026

Anthropic has added a rule to the system prompt for its Claude models aimed at preventing the re-creation of specific existing video games. The change was first spotted late last week by a social media user and AI decompilation enthusiast known as "chasm," who said a new rule appeared in the Claude Code prompt on or around October 27. Anthropic software engineer Addy Osmani confirmed on social media that the company had added two sentences to Claude Code's prompt on Opus 5.5, stating that using Claude to infringe intellectual property is against its usage policy. He said original games, genre games, and games built from a user's own designs and assets are meant to keep working, and that Anthropic will continue to iterate to strike the right balance.

The rule text instructs the model to build original games, games in a genre, and games from the user's own designs and assets, while not building a recreation of a specific existing game, even if it is renamed or called original. According to XDA, the rule is stored in a JSON file under the name "tengu_heron_brook" and appears in version 2.1.314 of the Claude Code binary, released on October 30. XDA also reported that the same "tengu_heron_brook" rule had served other purposes in earlier versions, and that the Opus 5.5 model appears to carry its own separate instructions restricting game recreations as part of its system prompt bundle.

The change lands amid a surge of game projects built with AI coding tools on top of existing work. That includes quick copies of game demos, which have made some developers reluctant to show early work publicly, and thousands of AI-assisted decompilation and porting projects for retro games. Decompilation, the practice of reverse-engineering a game's compiled code to rebuild readable source, traditionally took painstaking human effort; AI tools have made the process dramatically faster, though some veteran coders argue these ports and decompilation projects fall short of the quality of human-driven efforts.

The pressure has been visible in the community. Decomp.dev, a years-old community devoted to game decompilation, announced last month that it was shutting down. Maintainer Luke Street wrote that rapidly improving AI tools had made clear the community aspect of decompilation had been broken, with newcomers racing to port games, especially using decompilations they did not contribute to. He said the site had originally been designed to gamify decompilation, making incremental progress motivating when each fraction of a percent took real human effort, but that the same idea had warped into a hollow numbers game driven by people who just want to burn tokens.

The prompt is unlikely to meaningfully slow the broader copying wave. Game copiers can move to other frontier models that have not adopted similar restrictions, and even if the whole AI industry barred the use case, coders could turn to cheaper open-weight models that carry no such rules and trail the leaders only modestly in capability. This suggests the restriction functions less as a technical barrier than as a policy signal: it changes what Anthropic's own models will do by default, not what is technically achievable with AI assistance overall.

The more immediate benefit may be legal. AI companies have faced scrutiny when their models produce images closely based on copyrighted works, and a documented policy against generating clones gives a provider a clearer position that misuse violates its terms. Osmani's statement frames the change in exactly those terms. Decompilation based on reverse engineering has generally been treated as legally defensible when expressive elements such as art and sound assets are left out, but individual users of coding assistants have no guarantee of being that careful, guardrails or not.

The rule also raises a harder question of line-drawing. A popular pattern involves merging two game concepts, such as placing Mario inside Elden Ring or Spider-Man inside Batman: Arkham Knight. Whether such mashups count as recreations of a specific existing game or as distinct transformative works is unsettled, and the use of protected characters and imagery complicates it further. The same ambiguity applies to distinguishing a direct re-creation of Tetris, which the prompt ostensibly forbids, from an original game in the Tetris genre, which it ostensibly permits. Courts struggle with these distinctions; a two-sentence instruction asks a model to make the call reliably on every request.

There are early signs of friction. A handful of issue reports on Anthropic's GitHub suggest the new prompt is getting in the way of some game development use cases, and Osmani invited users to send him cases where Claude wrongly turns down work that is theirs. In practice, prompt-level restrictions tend to produce both false positives, where legitimate original work gets refused, and false negatives, where a determined user reframes a request until the model complies. That trade-off is likely to surface as complaints about blocked projects rather than as a measurable drop in cloning activity.

Why it matters: Developers using Claude Code for game work may find some requests refused, particularly ones that describe an existing title by name or lean on its characters and assets, and they should expect the boundary to be fuzzy and to shift as Anthropic iterates. For product teams, the change is a reminder that model behavior can be governed by system-prompt policy as much as by capability, and that such policies are provider-specific and easy to route around by switching models. The likely effect, by inference, is limited deterrence of deliberate copying alongside some added friction for legitimate and transformative projects.

Read at Ars Technica · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

The tsunami of imitators is way bigger than any one model maker can stem.