> ## Content Index
> Fetch the complete content index at: https://www.regenraygun.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Anthropic dials back on Claude not copying games
- URL: https://www.regenraygun.com/anthropic-dials-back-on-claude-not-copying-games/
- Published: 2026-11-06T10:54:15.000Z
- Updated: 2026-11-06T10:54:15.000Z
- Description: Anthropic has narrowed the Claude Code rule that stopped it rebuilding existing games.
- Author: Jon Jordan
- Tags: Anthropic

The release notes for Claude Code 2.1.317 run to 87 items. They document OpenTelemetry exports hanging behind proxies that never answer, attachment chips in the VSCode panel staying small when the chat font size is raised, and a Windows volume check that now compares a folder's identity rather than its path.

They do not mention that Anthropic has changed what Claude Code will and will not build.

But that change is the reason the version number is being reported at all. On 4th November, five days after confirming the rule existed, Addy Osmani of Anthropic's technical staff [posted an update](https://x.com/addyosmani/status/2117848568535106001?ref=regenraygun.com): 

> Last week some of you ran into Claude Code pushing back on game projects it should have just built. The short rule added to the Opus 5.5 prompt to steer Claude away from recreating specific existing games, he wrote, was also kicking in when you referenced one just to describe the mechanics you wanted. It has been updated in 2.1.317\. You can now name a game to describe what you want and get an original game in that genre, and if you ask for a copy of the original, Claude offers to build something similar instead.

Read [the changelog](https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md?ref=regenraygun.com) for that version and none of it is there. The only occurrence of the word clone refers to git repositories. [Tom's Hardware covered the change](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-loosens-game-restrictions-claude-will-now-offer-to-build-original-game-in-same-genre-if-asked-to-copy?ref=regenraygun.com) the following day with no comment from Anthropic beyond the post.

So the fix arrived exactly as the rule did. [The original](https://x.com/addyosmani/status/2116213240363729013?ref=regenraygun.com) was two sentences pushed into the Opus 5.5 prompt under a feature flag, spotted by an X user reading the diff, and confirmed in a reply. Its replacement is a prompt edit announced on one engineer's personal account. Neither version appears in any document a developer could cite, link to or diff.

On the merits, the fix is correct and the diagnosis is honest. The failure mode Osmani describes is the exact one the complaints described: the rule could not distinguish naming a game as a reference from asking for a copy of it. 

Reference is the native vocabulary of game design. Every pitch, spec and task ticket in the industry describes what is being made by naming what it resembles, and a rule that treats "like *Vampire Survivors*" and "make me *Vampire Survivors*" as the same request was always going to refuse work it had no business refusing.

It is worth being precise that this narrows the trigger rather than the prohibition. Recreating a specific game is still off the table. What changes is what counts as asking. And whether being offered a genre-alike amounts to much of a concession is arguable — a developer who wanted a working port of something they own has not been given a smaller version of that, they have been given a different thing.

Two threads from last week are untouched.

The first is everything that was not a game. [Tom's Hardware documented](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-confirms-new-claude-code-limits-to-prevent-game-cloning-users-report-opus-5-5-is-also-actively-blocking-reverse-engineering-tasks?ref=regenraygun.com) Opus 5.5 refusing reverse-engineering work with no connection to games: software checks by its own former staffer, a product activation request, a Reddit user suspended and reinstated, plus reports of blocks on anti-cheat development and security reviews. A studio trading as Conflict Simulations said Claude Code had begun refusing [digital board-wargame modules commissioned by paying customers](https://wccftech.com/claude-code-game-restrictions-target-clones-but-developers-say-legitimate-projects-are-being-blocked/?ref=regenraygun.com). The update is scoped to game projects and says nothing about any of it. Those were the most damaging cases, and in public they remain where they were.

The second is that the rule is still porous in the way that matters. It lives in Claude Code on Opus 5.5, so calling the same model through another harness goes around it. A narrower rule catches less legitimate work, which is a real improvement, and it misses commercial cloners exactly as completely as the old one did, because anyone cloning a game for money has every reason to use a different client.

The part that genuinely worked is the feedback loop. Five days from confirmation to correction, with a named engineer collecting reports and acting on them, is faster than any platform holder has moved on anything this year. 

Which leaves the precedent sharper than before rather than softer. Five days is fast for a company and instantaneous for a governance regime. In nine days the set of games that could be built with the most widely used coding agent in the industry was narrowed and then partially widened again, by a company with no stake in games, and the entire documentary record of both moves is two posts on a social network.

Compare that to the other positions taken this autumn. Itch.io's disclosure requirement is a form you can read. Steam's AI disclosure field is published policy with a revision history. Pantaloon's no-generative-AI commitment is a stated editorial line you can hold it to. All three are weak in various ways but all three are legible – you can see the rule, cite the rule, and know when it changes.

Still, the most consequential rule in this space is the one that exists only as prompt text. That is not secrecy. Anthropic has been quicker to explain itself than most companies would be and invited developers to send back anything wrongly refused. 

But nobody has yet decided that changing what a model will build is the kind of change that belongs in a release note. Until somebody does, the governing document is a tweet, and the version number is the only thing you can actually look up.