Anthropic confirmed that it added two sentences to the Claude Code instructions when running Opus 5.5 to prevent rebuilding of existing games, while the programmer reported that even unrelated decompilations of the game were rejected.

Anthropic confirmed that it added two sentences to the Claude Code prompt when the tool runs the Claude Opus 5.5 model, to prevent AI from recreating someone else's existing game. Claude Code is a programming tool that runs in Anthropic's command line window. Addy Osmani, a member of the company's technical team, said on X on October 30 that using Claude to infringe on intellectual property violates the usage policy, and that original games, genre games or games made from the user's own design and resources must still run normally. Anthropic has not made any separate announcement about this change.

The rule was discovered by user X named Chasm and posted on October 30, the post quickly exceeded 2 million views. According to the photo this person posted, the rule appeared around October 27 and has the following content: make original games, genre games and games from the user's own design and resources, but do not recreate a specific existing game, even if changing the name or calling it an original game. Anthropic has not published the full text. According to Chasm, this instruction is pushed down by the server so users cannot turn it off themselves.

The context is a wave of games being remade by AI. Some independent game companies have canceled public demos for fear of someone putting the demo into the AI ​​engine and releasing a copy first. Digital Sun Games, the company that makes ReVamp, is one of them. Game makers say the new rules may only deter sketchy copycats, as users can still switch to a similar open-ended model.

Side effects appear quickly. On Claude Code's GitHub page, there are many reports of being rejected midway through a game project, when simulating a gameplay mechanism or patching a board game. On October 30, Ray Weiss, CEO of Conflict Simulations, reported that Claude Code had stopped working on paid projects that it had supported for many months. His company makes digital copies of war board games for customers who own physical copies, using original code and documents provided by customers. He suggested that Anthropic differentiate between copyright infringement and custom development, and noted major changes in the Claude Code guidelines. Some other game developers reported that writing anti-cheat code and security checks for their own games were also blocked.

The scope is wider than the game. Decompilation, which means analyzing a software to understand how it operates, was once handled well by Claude Code, but now users report being rejected for many non-game related tasks: analyzing Windows crash dump files, removing pre-installed software on Samsung TVs, or using Reverse Engineer Anything (REA), an open source toolkit that allows AI to call a decompiler. Jarred Walton, technical writer for Phison Electronics USA, said his work still worked normally until October 31 and he could continue working after switching to Opus 4.8. A Reddit user said his account was temporarily locked and then restored after using Opus 5.5 with REA. These reports all come from users and have not been confirmed by Anthropic.

It is unclear whether the decompilation rejections are due to the two new sentences or due to specific cybersecurity limitations. Some reports on Reddit about reverse translation appeared from around October 21 to 24, before the game rules appeared. When launching Opus 5.5 on September 22, Anthropic said it would apply cybersecurity measures similar to Fable 5.1 and move most of these tasks to Opus 4.8. On October 6, the company expanded its Cyber ​​Security Verification Program so that vetted researchers face fewer restrictions. Decompilation also has legitimate uses beyond cybersecurity: in October, a Red Hat engineering manager used Claude Code to help build a scanner driver for Linux.

Osmani says Anthropic will continue to adjust to find the right balance, and invites users to send him their own instances of Claude refusing to belong. The broader problem lies in the nature of cloud tools: the provider can change the AI's behavior without the user changing the code, tools or prompts, so what worked yesterday may be rejected today.

As of November 2, Anthropic has not responded to a request for comment on the reasons for the decompilation rejections. Game-related blocking is currently only recorded in the Claude Code combination with Opus 5.5.