What changed in Claude Code on 27 October?
Anthropic has added a rule to Claude Code that stops the coding agent from rebuilding a specific existing video game. Claude Code is the company's agentic coding tool, used by developers to read, write and refactor code from a terminal or editor. The new instruction appeared in the tool's rules around 27 October 2026, according to users who inspect its internal instructions. It was first flagged publicly by a user posting as Chasm.
The rule sits under an internal flag named tengu_heron_brook. Game Developer, which reported the change, quotes the text in full. "Do build original games, games in a genre, and games from the user's own designs and assets," the rule begins. The second sentence closes the loophole: "Do not build a recreation of a specific existing game, even if it is renamed or called original."
Anthropic confirmed the change through Addy Osmani, a member of its technical staff. In a reply on X, he said the company added "two sentences to Claude Code's prompt on Opus 5.5" to steer the agent away from recreating someone else's game. He added that original games, genre games and games built from a user's own designs "are meant to keep working".
The rule appears to be delivered from Anthropic's servers rather than through a client update. Claude Code's public changelog, which lists fixes for each version, does not mention it. The developer who flagged the change said there was no way to disable it, according to Game Developer. The report was published on 30 October and updated once Anthropic confirmed the change.
That makes the restriction easy to miss and hard to audit. Nothing in the interface announces it, and Anthropic has not published a list of similar prompt rules. The instruction has no version history of its own, unlike a normal release note. A developer can only discover it by asking for a clone and reading the refusal.
Why is Anthropic doing this now?
The change follows weeks of complaints from game developers about AI-assisted clones on Steam. PC Gamer counted 8,011 demos listed for the October 2026 Steam Next Fest, roughly double the number in June 2026. In June 2022, the same festival had just over a thousand demos.
PC Gamer attributed much of the surge to Anthropic's own model. Its review of the week ending 18 October cited "massive efficiency gains for vibe coders" in Opus 5.5 and noted that the model had been used to decompile commercial games. The same piece found at least seven Roblox clones of a small game called Escape the Walls within days of its creator posting footage. When that creator shared work in progress, one user replied: "You have exactly 7 hours before someone makes this on Roblox and gets 20x your revenue." PC Gamer described the reply as prophetic.
The reporter was blunt about what the festival contained. He wrote that the vast majority of the demos would be "meritless slop", and that many of them probably did not exist as concepts before October. That is an assessment rather than a measured figure, but it captures the complaint that reached Anthropic.
Studios have responded by withholding builds. Digital Sun, the maker of Moonlighter, cancelled a public demo of its next game. Its marketing manager, Israel Mallén, told PC Gamer that "if you can't show your game to the world without fearing it'll be cloned, something's broken". PC Gamer also reported that the creators of Drunken Goddess Reflux pulled a planned Steam demo.
Anthropic's launch post for Opus 5.5 shows how central game building is to the model's pitch. It describes a tester who had several Claude models build a game from a single prompt, and says Opus 5.5 scored highest on graphics and polish. The model was released on 22 September 2026 and costs 40% less to run than Opus 5.
The capability Anthropic advertised is the same one developers say is being used against them. The rule tries to keep the first use case, building something new, while discouraging the second. Anthropic has not withdrawn the tool from game development, and its statement stresses that original work is meant to keep working.
What does the rule allow, and what does it block?
The rule separates inspiration from reconstruction. Original games, games within a genre, and games built from a user's own designs and assets are still permitted. A recreation of a specific title is not, even when it has a new name.
That covers most ordinary requests. Platformers, battle royales and farming simulators are genres, and the rule leaves them open. A project that reproduces a particular game is the target, including one that has been renamed to look original.
The renamed clause matters because renaming is the simplest workaround. A clone called something else is still a clone, and the rule closes that gap in the same sentence that bans the copy. It also puts the burden on the user to supply their own designs and assets rather than to lean on a known game.
The line is still fuzzy. Games borrow mechanics, pacing and art direction from one another constantly, and the rule does not define how much borrowing becomes a recreation. A developer can describe a project as genre work while aiming at a specific title, and the model has to judge which description is honest.
Osmani acknowledged the risk of false positives. He asked users to send him cases where Claude refused work that was genuinely theirs, adding that the team would keep iterating. That invitation is itself a signal that the boundary will move as examples arrive.
There is also a difference between imitating a design and lifting a codebase. PC Gamer draws that distinction, noting that simple games have spawned imitators for years while wholesale copying is a different act. The rule targets recreation, not the borrowing of ideas.
Cloning predates generative AI. Doom, released in 1993, produced so many imitators that "Doom clone" became a genre label. What changed is the cost and speed of producing one, not the appetite for copying a hit.
How enforceable is a prompt-level rule?
The rule is an instruction to a model, not a filter in the product. That makes it softer than a code-level restriction or a legal takedown, and much harder to measure.
Game Developer lists several limits. Model guardrails can be sidestepped with the right prompting. The software cannot easily check who owns an idea or a set of assets. Anyone who finds Claude too restrictive can switch to an open-weight model that carries no such instruction.
Anthropic has not published refusal rates, an appeals process or a formal policy document for game recreation. Its public statement says only that the company will "continue to iterate" to strike the right balance. The change was described as a prompt on Opus 5.5, leaving open how other Claude models respond.
The rule is also separate from the company's written policy. Anthropic's revised Usage Policy takes effect on 12 November 2026 and bars users from infringing third-party rights. It does not mention games, and it does not set out a specific rule for game recreation.
A prompt rule is strongest against low-effort copying, which is the kind that floods a storefront. It is weaker against a determined developer using a different tool. Game Developer puts the ceiling bluntly: the rule "may only deter the most low-effort copycats". Enforcement therefore depends on friction, not prohibition.
That is not nothing. Most storefront slop is produced by people looking for the shortest path, and a refusal at the first prompt adds a step. But the rule cannot be the only control if the goal is to protect specific games from being rebuilt.
What does Steam's clone problem add?
The rule also moves part of the burden from the storefront to the tool vendor. Valve has taken a disclosure approach rather than a ban. Since January 2024, Steam has asked developers to describe AI use in its content survey.
Valve separates pre-generated content, made during development, from live-generated content, produced while a game runs. Developers "promise Valve that your game will not include illegal or infringing content". Valve says it reviews pre-generated AI content the same way it evaluates everything else, including that promise.
Live-generated content needs an extra step. Developers must describe the guardrails that stop the model producing illegal material, and players can report bad output from inside a game. Valve also does not release Adult Only sexual content made with live-generated AI. Much of the disclosure appears on the store page.
Valve has resisted banning AI tools outright. When the disclosure policy launched, Valve said the changes would let it "release the vast majority of games that use it". At Steam's Indie Connect event, Valve's Tom Giardino said the company does not want "to be gatekeepers telling developers which tools they should or shouldn't use", according to PC Gamer.
Giardino did not offer a specific plan for stopping clones, PC Gamer reported. His advice was to make games harder to copy by adding "features, personality and creativity" that come from the developer's own perspective. He also said that making Steam "a smarter and more dynamic platform" was the starting point of Valve's philosophy.
Anthropic's rule now handles one narrow piece of that problem before a build reaches Valve. The division of labour is becoming clearer. The model provider writes prompt-level restrictions, the platform requires disclosure, and the developer carries the contractual responsibility for what ships.
What does this mean for studios and Claude Code users?
For studios, the practical question is whether to show work in progress. Demos and playtests are how small teams find players, and cancelling them has a cost that is easy to miss. A game that is never seen is also harder to wishlist, and a closed demo cannot build the audience that makes a launch viable.
For Claude Code users, the change is a reminder that behaviour can shift between releases. A prompt rule can arrive without a version number, a changelog entry or a setting. Teams that depend on a specific output should test it rather than assume it, especially before building a pipeline around it.
The commercial logic is straightforward. Anthropic sells Claude Code to professional developers, and a coding tool associated with storefront slop is harder to defend inside a studio. A narrow refusal protects the brand at low cost, and it does so without removing the game-building feature from the product.
The risk is over-blocking. A developer prototyping an original game with familiar mechanics may now have to argue with the model. Osmani's request for examples suggests Anthropic expects some friction and wants to tune the rule rather than leave it fixed.
There is a wider signal too. Model providers are becoming rule-makers for creative work, and those rules are often enforced through prompts rather than published policy. Users see the effect, not the mechanism, which makes the rules difficult to contest or plan around. Teams buying coding agents should treat a vendor's prompt rules as part of the product, not an implementation detail.
What to watch next
The first test is whether the rule becomes formal policy. A prompt is easy to change, and Anthropic has not said whether the restriction will appear in its usage documentation or apply across its models. The revised Usage Policy lands on 12 November 2026, which gives the company a natural place to say more.
The second is scope. Anthropic framed the change around intellectual property, not games alone. Its usage policy already bars users from infringing third-party rights. Whether the same prompt-level approach reaches code, art or music is unclear, and no announcement has suggested it will.
The third is evidence from Steam. Valve publishes AI disclosures on store pages, and those pages will show whether clone-style releases fall. Watch also for Valve's response to developer complaints about copying, which has so far been encouragement rather than enforcement.
The fourth is competition. If Claude Code refuses work that rival tools accept, developers will route around it. Anthropic's balancing act is to look responsible without becoming the slowest option for the people who pay for the tool.
The larger question is whether prompt-level rules can substitute for clearer rules about training data, outputs and liability. Anthropic's two sentences are a start on output behaviour. They do not settle the underlying dispute about who owns a game idea, or who pays when a clone appears. A useful test would compare refusal rates across models and prompts, but Anthropic has not published that data for game recreation.
For now, the rule is narrow. It protects one category of work, at one layer of the stack, for one model in one product. That is worth watching precisely because it may not stay that way.