Anthropic confirms a Claude Code prompt change for specific game recreations

What happened?

Anthropic technical staff member Addy Osmani said on October 30 that the company added two sentences to Claude Code's Opus 5.5 prompt. The stated purpose is to steer the system away from using Claude to infringe intellectual property by recreating someone else's specific game. The reported wording draws a line between building original games, games in a genre and projects based on the user's own designs or assets, versus recreating a particular existing game even when it is renamed or described as original.

Osmani also asked users to report cases where Claude rejects their own work, which indicates that Anthropic expects false positives at the boundary and is still collecting product feedback. The public confirmation does not say how an appeal would be handled or whether a refusal can be overridden by supplying a license. It is therefore a statement of intended behavior, not a complete rights-clearance system.

Source: Addy Osmani / Anthropic ↗ · Source: Anthropic ↗ · Source: Anthropic ↗

The scope is narrower than a game-development ban

Anthropic's public explanation is not a blanket prohibition on game work. It expressly says original games and projects using the user's own material are meant to continue working. That leaves difficult categories unresolved: a licensed remake, a mod, a compatibility layer, a classroom reconstruction, a reverse-engineering exercise and a project that mixes user-owned and third-party assets may not present the same facts. The public statement does not publish a complete decision tree or a guarantee that every account sees identical behavior.

The separate legal/compliance documentation is useful context but not a test result. It describes restrictions on third-party applications routing Claude.ai credentials on behalf of users and distinguishes those applications from ordinary use of an unmodified Claude Code binary. It does not establish how the new prompt behaves for every model, interface, rollout cohort or edge case.

The confirmation also matters because the change is described as a prompt instruction, not as a public compiler rule or a separate account setting. A user can therefore encounter the boundary inside a normal coding session without seeing a new product control. The source does not say whether the instruction is applied to every Opus 5.5 request, only Claude Code, or a staged portion of the service. It also does not publish an appeal process, an audit log or a way to distinguish an IP refusal from a generic failure. Those missing details make the safe description a confirmed steering change with incomplete rollout documentation.

Source: Addy Osmani / Anthropic ↗ · Source: Anthropic ↗ · Source: Anthropic ↗ · Source: Anthropic ↗

Why does it matter?

This is a product-policy change delivered through the model prompt rather than a new visible setting. That matters for developers because the same command can now produce a different refusal depending on how a request is framed and which assets it references. A team building a legitimate genre game should keep its original designs, licenses and asset provenance clear, because “not a copy” is not itself a technical guarantee that a request will be accepted. The public evidence supports a stated intent, not reliable prediction of every boundary.

It also changes how to evaluate coding-agent demonstrations. A refusal or successful generation is evidence about one prompt, account and rollout, not proof that Claude Code either bans or permits game cloning in general. Workflows should separate code-generation capability from permission to reproduce protected expression, preserve a human review path for assets, and avoid treating an internal prompt label or a single social-media test as the full policy.

A useful editorial distinction is between refusing to reproduce expressive content and refusing to implement compatible mechanics. A game can share genre conventions, physics ideas or interface patterns without copying protected assets, but a model may not reliably infer that boundary from a short prompt. Developers should give the agent original specifications and licensed inputs, then inspect generated code and assets before distribution. That workflow protects the user even if the model's refusal boundary changes again.

Source: Addy Osmani / Anthropic ↗ · Source: Anthropic ↗ · Source: Anthropic ↗ · Source: Anthropic ↗