Anthropic bars Claude Code from cloning existing games – but it blocks too much

What's going on

Anthropic has changed Claude Code, its coding tool, so that it should no longer build clones of specific existing video games. According to Tom's Hardware, as reported by AI Times, developers using Opus 5.5 are also getting refusals on tasks that have nothing to do with copyright infringement. Anthropic has confirmed the rule but hasn't said anything specific about the individual false blocks.

The rule and how it was found

On October 27, Claude Code received a new rule, according to X user Chasm. It is delivered from Anthropic's servers through a configuration field called "tengu_heron_brook". There is reportedly no setting to turn it off. According to ShiftDelete, the post drew more than 1.8 million views within a day. The wording:

Build original games, games in a genre, and games from the user's own designs and assets. Do not build a clone of a specific existing game, even if it is renamed or described as original.

XDA Developers found the instruction in version 2.1.314 of the Claude Code program, along with a separate mechanism that can enable it for Opus 5.5 only.

Addy Osmani, who works on the technical side at Anthropic, wrote on X on October 30 that the company added two sentences to the prompt because using Claude to infringe intellectual property violates the usage policies. Original games, genre games and games built from your own designs should keep working. Anthropic will keep tuning the balance, he said, and if you get a refusal on your own material, you should send it to him.

False blocks on legitimate projects

Reports are piling up on GitHub, according to AI Times. Work stopped midway through development, or features were refused when they were modeled on how a specific game works. One case involved a program in the form of a tabletop game. Others reported refusals when building an anti-cheat system for their own game, or during security reviews, with the model citing security policies.

Reverse engineering was affected too. That means analyzing how a program is built and how it works, which is useful for compatibility, security and debugging. Reported cases include using a "Reverse Engineer Anything" tool, analyzing Windows crash dumps (error logs) and removing unwanted software from a Samsung TV. According to AI Times, it's unclear whether all of these cases stem from the same rule. The timing coincides with tighter cybersecurity restrictions: on the 6th, the cyber verification program was expanded, with three access tiers for advanced capabilities. Whether the two mechanisms are connected is unconfirmed.

XDA ran its own test. Opus 5.5 refused to build a clone of Pokémon Diamond, citing session instructions. But it helped with a decompilation (turning disassembled functions into readable C and rebuilding the ROM from your own copy) and with a static recompilation to run natively on Windows. That matters for fan ports of old console games.

Why now?

The rule arrives at a time of heavy concern among indie developers about fast AI clones. According to IGN, a flood of low-effort clones caused anger after Steam Next Fest. Alliance Arts (Drunken Goddess Reflux) and Who You canceled their Steam demos. According to tbreak, Digital Sun Games also scrapped the public playtest of ReVamp. At an event in Busan on October 27, Valve said it doesn't want to ban tools and announced no new rules. According to tbreak, Anthropic's move is the first by an AI provider itself.

Limits of the rule

• There is no definition of what counts as a "clone". Fan projects, homages and disguised clones fall into gray areas.
• It's unclear how Claude is supposed to check whether users have rights to the assets they upload.
• According to Game Developer (via tbreak), the rule may only deter the laziest copycats. Safeguards can often be bypassed with rhetorical tricks.
• Open-weight models aren't subject to any such restriction.
• Anthropic didn't say whether other models in Claude Code are affected.

Our take

The move shows a basic problem: a prompt rule has a hard time telling intent from method, because game analysis, anti-cheat and cloning use similar techniques. According to ShiftDelete, nothing changes for most developers. But if you work with reverse engineering or game mechanics, you risk false refusals. The fact that Anthropic takes feedback directly suggests it will keep adjusting. Whether the rule will really slow down clones on Steam is an open question. The sources think it's unlikely to do so on its own.