Anthropic Confirms Claude Code Rule Against Recreating Existing Games, as Reverse-Engineering Complaints Pile Up
A.I. / news
Anthropic Confirms Claude Code Rule Against Recreating Existing Games, as Reverse-Engineering Complaints Pile Up
A two-sentence prompt change on Opus 5.5 is blamed for refusals on developers' own projects. Anthropic says original games should keep working.

Anthropic has confirmed it added a rule to Claude Code that tells the model not to recreate a specific existing game, and users now say the same refusals are hitting reverse-engineering work that has nothing to do with games. Game Developer reported the confirmation on Oct. 30. Tom's Hardware followed on Oct. 31 with user reports of wider blocking.
What the Claude Code rule says
According to Game Developer, an X user called Chasm first spotted the rule in a Claude Code update pushed around Oct. 27. It sits under the flag "tengu_heron_brook" and reads: "Do not build a recreation of a specific existing game, even if it is renamed or called original." Chasm said it cannot be switched off.
An Anthropic spokesperson did not answer questions directly. The spokesperson pointed Game Developer to a post on X by Addy Osmani, an Anthropic member of technical staff, according to the outlet.

What Anthropic says it meant
Game Developer reports that Osmani said Anthropic "added two sentences" to Claude Code's prompt on Opus 5.5. He said the aim is to steer the model away from recreating someone else's specific game, because using Claude to infringe intellectual property violates Anthropic's usage policy. The wording here is as Game Developer relayed it, and the original post was not available to read.
Osmani also said original games, genre games and games built from a developer's own designs and assets should keep working, and that Anthropic will keep iterating. He asked developers whose legitimate work is refused to send it to him. The Terminal earlier reported on another Claude Code episode in which a fake 2.1.313 release planted SSH keys.
Osmani's move to Anthropic was reported on Sept. 8 by RuntimeWire, which said he had not disclosed a formal title.
Reports of wider blocking
Tom's Hardware wrote on Oct. 31 that, according to multiple user claims, Claude Code or Opus 5.5 is also declining harmless reverse-engineering requests, including ones that use the Reverse Engineer Anything tool. It listed alleged refusals to analyse a Windows crash dump and to clear junkware off a Samsung TV. Tom's Hardware said some reports may come from users with bad intentions, and that at least one came from a developer with a real track record.
These are user claims. Anthropic has not said the rule covers reverse engineering, and Tom's Hardware described the link as possible.
| Claim | Source | Status |
|---|---|---|
| Rule added around Oct. 27 | Chasm via Game Developer | Anthropic confirmed an addition |
| Two sentences on Opus 5.5 | Osmani via Game Developer | Anthropic-attributed |
| Crash-dump and TV refusals | User reports via Tom's Hardware | Unconfirmed |
Why studios are watching
Game Developer says studios fear fast AI-made clones. Digital Sun Games is one of several that cancelled a public playtest of its game ReVamp for fear someone would feed the demo into an AI tool. The outlet also notes the rule may stop only low-effort copycats, and asks how software would tell whether a user owns a game idea or its assets.
Both outlets say commenters pointed users to open-weight or local models, which carry no such rule. The Terminal's coverage of open-weight releases, such as Nanbeige4.3-33B, tracks what those models permit.
Anthropic has not said whether it will publish the full text of the rule or narrow it.
Sources
More in A.I.
- 01Nanbeige4.3-33B Ships Under Apache-2.0 With a 524,288-Token Window and Self-Reported WinsNanbeige Team's 33B model is built as a teacher for smaller agents. Its own table has Qwen3.8-27B ahead on SWE-Bench Pro and GDPval.
- 02GPT-5.6 Sol Launched a Factorio Rocket on Peaceful Mode, With Hauling Done by HandBlogger Ryan Madden's single run took 13.5 in-game hours. The agent carried supplies between areas because it never linked some production lines.
- 03Bootgpt Fits a Character-Level Language Model Into a 512-Byte Boot Sector Plus 23 Sectors of WeightsThe MIT-licensed Show HN project writes fake Linux boot logs on a 386-class machine with no operating system. A rival project runs a Llama 2 model in 1,189 bytes.
- 04Mistral Large 4 Weights Are on Hugging Face, and the Card Disagrees With the Launch PostThe 1-trillion-parameter model is ungated under a custom licence, but its active-parameter count and two benchmark scores differ from Mistral's October 6 announcement.