Anthropic has told its coding assistant Claude Code to stop rebuilding other people’s video games. A new standing instruction on the Opus 5.5 model refuses recreations of a specific existing game, even when the request renames the project or presents it as original. Original games, genre games, and projects built from a user’s own designs and assets are meant to keep working, and Anthropic says it will keep refining the balance.
What changed
The rule reached Claude Code around 27 October, delivered from Anthropic’s servers under the internal flag “tengu_heron_brook”. It is not a setting users can switch off, and it applies to Claude Code running on the Opus 5.5 model. Anthropic has not said whether other Claude models will follow.
The instruction itself is two sentences long. It tells Claude to build original games, games in a genre, and games from the user’s own designs and assets, but never to build a recreation of a specific existing game, even if it is renamed or called original.
The change surfaced when an X user posting as Chasm spotted the new entry in Claude Code’s rules and shared it publicly. Anthropic then confirmed it through a company spokesperson, who pointed Game Developer to a statement from Addy Osmani, a member of Anthropic’s technical staff.
Osmani said two sentences had been added to Claude Code’s prompt on Opus 5.5 to steer it away from recreating someone else’s specific game, because using Claude to infringe intellectual property goes against the company’s usage policy. He also invited developers whose own projects are wrongly refused to send those cases to him directly.
The timing matters because this is a prompt-level guardrail, not a new model release. It changes what one tool refuses, while the underlying model and its other uses stay the same. It also follows a rough patch for Claude Code security, after a fake Claude Code update planted a backdoor in Cursor.
Why developers pushed for it
The trigger was a wave of AI-generated clones that filled feeds during Valve’s Steam Next Fest in October. Players reported copies of existing games, including titles that had only just released their own demos, crowding out the originals they copied.
Some independent studios reacted by pulling public demos. Digital Sun Games was among the studios to cancel an upcoming public playtest, for fear someone might feed the demo into an AI tool and release their own version of its game.
Developers also pressed Valve to explain how it could stop clones from supplanting their games on Steam, prompting a rare explanation from the company to developers in South Korea. Anthropic’s rule does not solve that platform problem, but it removes one widely used tool from the cloning pipeline.
Our read
Two sentences cannot settle the hardest question here, which is what counts as a recreation. A game rebuilt level by level under a new title is clearly covered, but genre conventions, shared mechanics, and remakes of a developer’s own earlier work all sit in a grey zone.
Anthropic’s answer so far is iteration. Osmani has promised to keep adjusting the rule, and his invitation to report wrongly refused projects suggests the company expects false positives and wants to learn from them. That feedback loop matters more than the wording, because the edge cases will arrive faster than any policy update.
The rule also fits a broader tightening of Claude’s guardrails. Anthropic has spent recent weeks spelling out what Claude will and will not do, adding narrow, enforceable instructions where misuse has already shown up in the wild, then watching how developers react.
That pattern includes its rule against sustained, needless cruelty toward Claude, which takes effect on 12 November. In both cases, Anthropic is policing behaviour at the point of use, rather than announcing a sweeping new policy and leaving enforcement vague.
There is a limit to what this approach can achieve, and Anthropic’s critics have already named it. Safety limits on generative tools have repeatedly been sidestepped with rhetorical tricks, and anyone blocked by Claude Code can turn to open-weight models that carry no such restriction.
The rule will most likely deter the lowest-effort copycats, the ones who paste a demo into a tool and publish the result, rather than determined infringers. That is still a real effect, but it should not be mistaken for an end to game cloning.
Why It Matters
Coding agents have become a distribution channel for games, not just a way to write them. When a tool can turn a demo or a description into a playable copy in hours, the window between a studio showing its work and a clone appearing collapses.
Anthropic’s rule is the first clear sign that a major coding-assistant provider is willing to police that window directly. Other providers will now face the same question from their own developer communities, and their answers will shape how much protection indie studios can expect.
The open questions will decide whether the rule holds up. Developers need to know how Anthropic judges ownership of assets, whether mods and decompilations are affected, and what happens when a legitimate project is refused. How quickly and transparently those cases are resolved will shape whether developers treat the rule as protection or as friction.
Reports already describe developers saying Claude Code has turned down reverse-engineering and troubleshooting work. Anthropic has not confirmed a broader ban beyond game recreation, so those cases remain reports, not established effects of the rule. The distinction matters, and it deserves watching.
Background
Claude Code is Anthropic’s agentic coding tool, and Opus 5.5 is the model family at the centre of its recent releases. The game-clone rule arrived quietly, spotted by users before Anthropic confirmed it, and it remains a server-side instruction rather than a published policy document.
Source: Game Developer report and gHacks report
0 Comments