Anthropic Adds Two Sentences to Claude Code to Block Game Clones

According to a BeInCrypto report, Anthropic has made a very small change to Claude Code: two sentences added to steer the coding agent away from helping users replicate existing video games. The story frames the update not as a feature release but as a narrowing of what the tool will do.

Claude Code is Anthropic's command-line coding agent for developers. It can read and write across a codebase, run commands and carry out multi-step development work. Because it produces runnable code directly, a request to build something “just like” a known game can quickly drift toward reproducing protected art, mechanics or level design.

The change appears to sit at the prompt level rather than in model weights or a version bump. That limits what it can guarantee, since instructions can be worked around or overridden by more specific requests. Still, writing the constraint into the product itself signals that the risk has moved from conference panels into day-to-day operations.

The tension is familiar across generative coding tools. The more capable the agent, the more users expect a finished project in one pass, and a finished game is one of the most litigation-prone artifacts an agent can produce. Two sentences draw a default line between what the tool can do and what it should do.

Cost is the other consideration. Compared with training-time filtering or output review, prompt-level limits add almost no inference overhead and require no extra infrastructure. For an agent handling a high volume of calls, that is the lightest available control, at the price of relying on how faithfully the model follows instructions.

For developers, the practical effect is a shift in responsibility. A loose request can no longer simply be handed to the agent without thought; users may need to state originality requirements explicitly, or judge for themselves whether the output crosses a line. The lighter the restriction, the more of that judgment moves to the person using the tool.

The exact wording of the two sentences is not disclosed in the report, and it is unclear whether Anthropic pairs them with detection or appeals processes. What to watch is whether other coding-agent vendors adopt similar language, and whether users simply take the same request elsewhere. The constraint only holds if the wider ecosystem converges.