Anthropic Tells Claude Code to Stop Rebuilding Other People's Games

Claude Code game clones get a hard line

Anthropic has drawn a line inside Claude Code. The AI coding assistant will no longer help you rebuild a specific existing game, even if you rename it or describe it as original. The change arrived as two sentences added to Claude Code's system prompt on the Opus 5.5 model, and it is not a setting you can switch off.

The rule, first spotted by users around October 27 and confirmed by Anthropic on October 30, reads in plain language: build original games, games in a genre, and games from your own designs and assets, but do not build a recreation of a specific existing game. Anthropic engineer Addy Osmani confirmed the change publicly, saying the company added the sentences because using Claude to infringe intellectual property already violates its usage policy.

This is a small change with a big signal attached. For years, AI coding tools have lived in a gray zone: they would happily help you clone a game if you phrased the request cleverly enough. Anthropic just decided to police that gray zone itself, at the prompt level, before regulators or courts force the issue.

What Claude Code still allows

The rule is narrower than a ban on AI game development, and that distinction matters. Asking Claude Code for a farming sim, a deck builder, or a roguelike is fine. Building a game from your own art, music, and design documents is fine. The refusal kicks in when the target is a particular commercial game, reproduced level by level or mechanic by mechanic under a fresh coat of paint.

That is a sensible boundary in theory. Genres are not owned by anyone. A specific game, with its levels, characters, and art direction, is. Anthropic is betting that most developers can tell the difference, and that its model can too, most of the time.

The early tests suggest the rule has teeth. In one reported test, Opus 5.5 refused to recreate a well-known handheld role-playing game and cited its session instructions. The same model agreed to help with a decompilation project based on the user's own copy, turning disassembled functions into readable code. That is exactly the split Anthropic says it wants: help with your own work, refuse to copy someone else's.

Why now: the Steam Next Fest clone wave

Timing explains a lot here. During Steam Next Fest in October, players complained that their feeds filled up with low-effort AI-generated clones of existing games, including copies of titles that had only just released their own demos. At least one studio reportedly pulled a public playtest after clone-like projects appeared. Independent developers, who live or die on a few weeks of visibility, were furious.

Valve, for its part, said at a South Korea event on October 27 that it does not intend to ban AI-assisted games from Steam outright. That leaves the enforcement job to the tool makers, and Anthropic just volunteered. It is also worth noting what Anthropic did not do: it did not announce a new detection system, a watermark, or a legal settlement. It added two sentences to a prompt. The modesty of the fix is part of the story.

The enforcement problem nobody has solved

A prompt-level rule is easy to write and hard to enforce consistently. The rule gives no definition of what counts as a recreation, and it is unclear how Claude can verify whether a user actually owns the assets they upload. One developer building licensed board-game adaptations reportedly had Claude refuse work they described as properly licensed. False positives like that are the cost of a blunt instrument.

Determined cloners can also route around it. Rephrase the request, split it into generic steps, use another tool, or switch to an open-weight model with no such restriction. Anthropic's rule raises the effort required to clone a game; it does not make cloning impossible. Anyone who tells you otherwise is selling something.

There is a second, quieter risk. When a coding agent starts refusing work based on hidden instructions, developers lose predictability. A refusal after a routine model update can look like a bug. Anthropic's answer, in effect, is that refusals are now part of the product's behavior, and developers should treat a sudden refusal as a signal to check what changed, not as a glitch to work around.

What this means for developers

If you build games with AI help, the practical advice is simple. Keep your own assets and design documents organized, because ownership is now the thing that separates allowed work from refused work. Describe mechanics in genre terms rather than naming a target game. And if you hold a real license for an adaptation, expect to spend time proving it to a support human, since the model cannot verify a license on its own.

For the wider AI industry, this is a template worth watching. Other labs face the same pressure from creators, and a two-sentence prompt change is far cheaper than a lawsuit or a content filter built from scratch. Expect similar rules to appear in other coding tools, applied unevenly, argued about constantly, and enforced just well enough to matter. The Gemini 4 Argon release showed how fast frontier models ship; the Claude Code rule shows how fast their guardrails are becoming part of the product. Developers who rely on VS Code as an agent control center should assume these rules will follow them into every editor they use.

The bottom line

Anthropic did not solve AI game cloning. It made a choice about where its tool stands, and it made that choice visible. Original games keep working. Genre games keep working. Rebuilding someone else's game, even under a new name, now gets a refusal. That is a defensible line, an imperfectly enforced one, and probably the start of a much longer argument about what AI coding tools are allowed to build.