Claude Code blocks game clones but guesses ownership from the prompt

Anthropic has confirmed a new instruction for Claude Code using Opus 5.5. The tool should assist with original games, genre projects and user owned designs, but should not recreate a specific existing game even after it is renamed.

Two system prompt sentences change work on game code

Anthropic technical staff member Addy Osmani said the company added two sentences because using Claude to infringe intellectual property violates its usage policy. The guidance applies to Claude Code with Opus 5.5. Anthropic says original games, genre games and work based on a user's own designs and assets should continue to work.

The instruction was also found inside the Claude Code 2.1.314 binary. Text embedded in the application does not prove identical deployment to every account and model. Tests reported by XDA nevertheless showed an active refusal in at least some Opus 5.5 sessions.

Protecting rights holders moves the decision into runtime

Anthropic is addressing a real problem: models can sharply reduce the cost of copying game design, code and assets. But this intervention happens during development rather than at the finished output. Claude Code decides from the request and available context before it can establish licensing or file provenance.

For professional teams, the precedent matters. A cloud coding agent can restrict a class of tasks through a remotely changed system prompt without any change to the user's repository. Vendor policy becomes part of the toolchain alongside the compiler and CI, but with a less predictable result.

The same request can describe a clone or a legitimate port

A model cannot reliably infer from a prompt whether the user owns the source, holds a licence or is preserving an older game. Static recompilation, a compatibility runtime and reverse engineering are also different from rebuilding a game from scratch. A broad refusal can catch legitimate work while a cleverly rephrased request may evade the restriction.

Anthropic says it will keep iterating and asks users to report cases where Claude rejects their own work. That is a useful emergency route, not a scalable proof of ownership for every session.

Error reports will reveal the cost of false refusals

The key signal is whether Anthropic publishes a clearer boundary between a clone, a port and a compatible implementation. Data on false refusals and a way to provide verifiable licence or ownership context would matter even more.

If the only remedy remains contacting a company employee, the policy will work mainly for a small number of visible disputes. A usable rule needs a predictable appeal inside the product rather than personal intervention on social media.

Lilith's verdict

Claude now stands at the game project's door and checks ownership from what the visitor says. That will deter some copyists, but an owner without the right wording may still be left outside.