Anthropic adds two-sentence rule to Opus 5.5 system prompt to stop Claude cloning existing games

Anthropic has added two sentences to the system prompt of Claude Opus 5.5 that steer Claude away from recreating specific games that belong to someone else, and the company has confirmed the change. The source is an article in Chinese from QbitAI (量子位), published 2026-11-01.

What changed

The change surfaced when an X user going by chasm noticed that Claude Code appeared to be quietly blocking game cloning. Anthropic technical staff then confirmed the addition to the Opus 5.5 system prompt, according to QbitAI.

The rule for Claude Code, as QbitAI reports it, reads in translation:

You may create original games, games of a specific genre, and games built with the user's own designs and assets. Do not recreate specific existing games, even if the name is changed or the game is described as original.

QbitAI says Opus 5.5 was released on September 22. The outlet attributes a surge in game reverse engineering, decompilation, porting and cloning to the model's coding ability. Examples it cites include browser ports of GTA 5 and Call of Duty, and hackers running Xbox-exclusive Halo 2 and Nintendo's The Legend of Zelda: Breath of the Wild on a jailbroken PS5 through emulation. These are QbitAI's reports, not Anthropic's statements.

Reported side effects

According to QbitAI, the Claude Code GitHub issues page was flooded within days of the rule appearing. Developers building fully original games say Claude stopped mid-project, citing suspected recreation of an existing game. Others say they were refused when they wanted to reproduce a common game mechanic or patch their own tabletop game.

Industry context

QbitAI reports that indie studio Alliance Arts cancelled the public demo of its upcoming game 《醉酒女神反流》 (title as given in the source). The studio feared that someone would feed the demo to an AI and release a clone before its own launch.

The outlet also describes a user experiment with the free Claude Opus 4.6 model. A single prompt asking for a Pokémon clone as close to the original as possible produced a fully playable game in under two hours, per QbitAI. The outlet says some commentators call the situation the game industry's "Napster moment."

QbitAI's criticism

The article is openly skeptical of the rule. QbitAI argues that it:

• governs only Claude, while other models, including ChatGPT and open-source models, remain available for cloning;
• can be bypassed by stripping names and labels and splitting a clone into context-free technical tasks;
• produces false positives on legitimate projects.

QbitAI also relays online criticism that frontier AI labs trained their models on publicly available information, art and code, and now worry about copyright infringement by others. The outlet characterizes the rule as more posture than practical enforcement. That is the outlet's opinion, not a confirmed finding.

What this means

The change is a system-prompt instruction, not a new model or a new Claude Code feature. It works by steering the model's behavior, so what it blocks depends on how Claude reads a request. The GitHub reports suggest the model sometimes judges original projects and common mechanics to be clones. Developers using Claude Code or Opus 5.5 for game work should expect that some requests may be refused, and that the allowed categories in the quoted rule (original games, genre-based games, games built from the user's own designs and assets) mark the boundary Anthropic has set. QbitAI's view that the rule does not stop cloning is its own assessment, and the article gives no data on how much cloning activity has changed since the rule appeared.