NewsDesk World

Press releases and business news from around the world

Made available through support from

Claude Code Told Not to Rebuild Existing Games

Artificial Intelligence ·

Anthropic added a two-sentence rule to Claude Code's prompt that bars recreating a specific existing game. Ars Technica doubts it will slow the wave of AI-made game copies.

A short addition to Claude Code's instructions now steers the model away from rebuilding a particular existing video game, which Ars Technica says could shield Anthropic from some liability. AI has made copying and porting games far more efficient than before.

Where the line is hard to draw

The rule tells Claude to build original games, games in a genre, and games from the user's own designs and assets. It bars recreating a specific existing game, even one renamed or called original. Deciding which side a project falls on is not simple.

One popular trend uses AI to merge two game concepts, such as Mario in Elden Ring or Spider-Man in Batman: Arkham Knight. Whether those count as recreations, or as transformative works, is unclear. The same goes for a direct Tetris copy, which the prompt apparently bars, versus a game in the Tetris genre, which it apparently allows. Courts can struggle with such questions even in the best circumstances.

A liability question for Anthropic

Anthropic software engineer Addy Osmani explained the reasoning on social media. Using Claude to infringe intellectual property already breaks the company's usage policy, he said, and the new lines enforce that for games. Ars Technica reads the move as a way to protect Anthropic itself, since AI companies have already faced legal trouble when their models produced images based on copyrighted works.

The legal picture for the projects is mixed. Decompilation that works by reverse engineering code generally passes legal muster when art and sound assets are left out. Nothing guarantees that every person using Claude for such a project would be that careful.

A community that gave up

Decomp.dev, a years-old community around the "brutally difficult, time-consuming task" of decompiling games, announced last month that it was shutting down. Its maintainer Luke Street wrote that the community side of decompilation had been broken, because newcomers race to port games using decompilations they did not help build. He said the site's progress counters, meant to keep people motivated, had turned into a numbers game.

AI coding tools are fuelling a stream of projects built on existing games. Some are quick copies of game demos, and that has made certain developers afraid to show early work in public. Others are a flood of decompilation and porting projects for old games, built with AI. Work that once demanded painstaking human effort is now far more efficient. Some veteran coders complain that the vibe-coded versions do not reach the quality of human-led ones.

Ars Technica doubts the rule will change much. Copiers can use other leading models that carry no such limits. Even if every major lab added them, cheaper open-weight models are not far behind in capability.

How the change came to light

A social media user who goes by "chasm" spotted it late last week. According to that report, the rule appeared in the Claude Code prompt on or around October 27, stored in a json file under the name "tengu_heron_brook".

XDA then checked independently. It found the rule in version 2.1.314 of the Claude Code binary, released on October 30. The same rule name had been used for other purposes in earlier versions, XDA reported. It also found that the Claude Opus 5.5 model seems to carry its own separate instructions limiting game recreations.

Osmani said Anthropic will keep adjusting the balance. He added that if Claude refuses something that belongs to the user, people can send the case to him. A handful of issue reports on Anthropic's GitHub suggest the prompt is already blocking some legitimate game development.

Our View

The rule hands a model a judgment that lawyers and judges find hard to make. Mashups may be the hardest test, since they sit between copying and new work. Anthropic's response to the early GitHub complaints, and whether rival AI companies add similar lines, would show whether this is a one-off or the start of a pattern.

Source: Ars Technica