Two sentences set the rule: Claude Code will no longer recreate existing games

Two sentences, one new red line

On October 30, 2026, Anthropic added two sentences to Claude Code. Not a feature, not a new model — two lines of instruction in the system prompt: "Do build original games, games in a genre, and games from the user's own designs and assets. Don't build a recreation of a specific existing game, even if it is renamed or called original." Original games, genre games, and games built from your own designs and assets are fine. Recreating one specific existing game is not, even under a new name, even described as original.

This is worth a full article because it marks a shift in how governance works. For the past year, disputes over AI-cloned games mostly played out after the fact: takedowns, delistings, legal threats. Anthropic has moved the checkpoint to the moment of the request itself — not waiting until you have built the clone, but having the model say no at the moment you ask for it. For anyone building games with Claude Code, that is not distant policy news. It is what you will run into the next time you open a terminal.

First, the boundaries. The rule targets Claude Code running on Opus 5.5. It is not a blanket ban across all Anthropic products, and it is not a ban on AI-made games. In his public explanation on October 30, Anthropic technical staff member Addy Osmani stressed that original games, genre games, and games built from your own designs and assets are meant to keep working. The rule aims at one narrow class of request: copying one particular existing game wholesale.

Timeline: spotted first, confirmed second

The rule was not announced at a launch event. It was found by users first. Around October 27, X user chasm noticed a new server-delivered rule in Claude Code, filed under the internal identifier tengu_heron_brook. Reports later placed the instruction inside the code of Claude Code version 2.1.314, released just after midnight UTC on October 30, while version 2.1.312, released October 27, did not contain it. Notably, Anthropic's public changelog for the release lists dozens of other changes but not this one — it arrived not as a feature update, but as a quietly inserted governance instruction.

On October 30, Anthropic confirmed to Game Developer that the change was real, and Osmani posted his explanation on X: "We added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game because using Claude to infringe IP is against our usage policy." In other words, Anthropic's usage policy already prohibited using Claude to infringe intellectual property; the two sentences simply operationalize that policy inside every Claude Code session. Polygon's November 2 report confirmed the chain: user discovery, vendor confirmation, and a public staff explanation, all within days.

One detail captures the nature of the rule: according to its discoverer, it is delivered from Anthropic's servers, and there is no user-facing setting to switch it off. It is not a preference you can opt out of. It travels with the model, not with your machine — switch computers or reinstall the client and it is still there, because it sits in the instructions the model reads before every session.

What it blocks, and what it lets through

Read closely, the two sentences draw three lines. The first is the word "specific": the rule opposes recreating one particular game, not making a game in a genre. A platformer, a card battler, a farm management sim — none of those are targeted. Asking for a named, already-released game's levels, systems, and art direction packaged together is what crosses the line. The second line closes the most common workaround: renaming the clone or presenting it as original does not help, because the rule names that trick explicitly. The third line protects your own work: your own designs, your own assets, your own ideas remain in the allowed zone even if they resemble a genre.

The sharper the lines, the more visible the grey areas become. First, how does the model tell "recreation" from "homage"? A game with very similar mechanics but entirely new art and story — genre game or clone? The rule gives no test, and Anthropic has not published how the model is supposed to distinguish original material from copied material. Second, how does the model know whether you have the rights? If you own an old game or hold a license to remake it, the model has no way to verify your paperwork; it can only guess from the text of your request. Third, what counts as "your own designs and assets"? Does uploading a concept sketch you drew suffice, or does the model merely take your word for it? The rule answers none of these questions.

Real cases have already hit those grey areas. According to reports, on October 30 a developer representing Conflict Simulations LLC complained in Anthropic's official repository that Claude Code had begun refusing work on personalized digital board-wargame modules commissioned by customers — even after clearing the session and starting fresh, the model kept refusing requests involving the same repositories. That work is, in principle, exactly the "own designs and assets" side of the line, yet it was blocked anyway. Osmani's response was to invite people to report such cases to him — an admission that the false-positive rate has not yet been measured systematically.

Another widely reported set of tests shows the other side: Opus 5.5 reportedly refused to recreate Pokemon Diamond, citing its session instructions, yet agreed to help turn disassembled functions into readable C code and to rebuild a ROM from the user's own copy. The first is "recreating a specific game"; the second is "working with your own material," and the model did draw that distinction. But the same tests also show that enforcement depends on how the request is phrased. Reword the ask, and the same task may flip from refused to accepted. A prompt-level guardrail raises the bar; it does not lock the door.

Why now: a cloning wave forced the guardrail

Anthropic's timing is not accidental. Polygon's report sketches the backdrop: AI-generated games are flooding storefronts this year, Steam Next Fest was overwhelmed by AI titles that made it harder for players to find genuine indie gems, and the number of game pages on itch.io has doubled in two months, a surge the report links to vibe coding tools such as Claude. Worse than volume is cloning: other people's ideas rebuilt quickly with AI and shipped first.

The damage has already changed developer behavior. Polygon notes that Alliance Arts and Who You pulled the demo of Drunken Goddess Reflux — a game with 25,000 Steam wishlists — from the October Steam Next Fest, fearing their work would be copied; Digital Sun likewise cancelled the public playtest of ReVamp in favor of a controlled test on Discord. When developers stop testing in public because they fear being cloned, what breaks is not just one game but the public feedback loop indie games depend on. Anthropic's rule is, in that context, a signal to indie developers that one of the biggest AI coding tools is starting to block at the request stage.

Contrast this with the other kind of event covered here in early October: Bethesda taking down a fan-made AI Fallout: New York, Activision removing posts about a Modern Warfare 2 decompilation. That is enforcement after the fact — expensive, slow, and case by case. Anthropic's approach is different: write the policy into the model's default instructions so every session carries its own first filter. Neither approach has published effectiveness numbers — nobody has disclosed what share of recreation requests are blocked or what the false-positive rate is — but the directional difference is real: from "deal with it after it ships" to "filter it when it is asked."

What this guardrail cannot stop

Keep expectations realistic. The rule has three hard limits. First, it only governs Claude Code on Opus 5.5. Users can switch to a tool without this instruction, or to a locally run open-weight model, and do the same thing. Anthropic controls its own product, not the ecosystem. Second, it is a prompt instruction, not a technical block. How firmly a model follows it depends on phrasing; indirect descriptions, split steps, and staged implementations are known ways to soften it. Third, it cannot verify ownership. A legitimate rights holder remaking their own game and someone cloning another studio's game can look identical in chat text, and the model has no external evidence to separate them.

Taken together, those limits point to an honest framing: the rule's main effect is not to "eliminate cloning" but to raise the cost of casual cloning and to state the vendor's position. For most ordinary users, it adds friction to a spur-of-the-moment attempt to clone a popular game. For determined bypassers, it will not hold. For wrongly blocked legitimate developers, the current remedy is an informal "report it to staff" channel, not a formal appeals process. None of that diminishes the rule; it prices a two-sentence rule correctly. It was never designed to solve the whole copyright problem.

There is also a longer-term variable: whether other vendors follow. If only Anthropic writes this sentence into its prompt, the effect is diluted by the rest of the ecosystem. If the major coding tools each add similar default instructions, the default cost of cloning rises across the board. But following comes with the same price: the same false-positive complaints and the same boundary disputes. Osmani said Anthropic will "continue to iterate to ensure we're striking the right balance" — which reads as an admission that the balance has not been found yet, and that this first version is only a starting point.

For game developers using Claude Code: four practical tips

First, put "this is mine" into the project, not just into the chat. Keep design documents, concept art, asset sources, and license files in the repository so the model can see evidence of ownership in context. If you are wrongly blocked, those materials are also your strongest attachment when you report it. Saying "this is my game" in chat is the weakest defense; materials in the repo are the strongest.

Second, describe mechanics in genre language, not by reference. "A turn-based tactics game on a hex grid, with morale and supply-line mechanics" is far safer than "a game like X," and it is also more likely to be classified as a genre game. The first describes systems; the second points at a specific work, and that is exactly what the rule targets.

Third, do not treat renaming as compliance. The rule explicitly covers recreations "even if renamed or called original," which shows Anthropic anticipated that workaround. Shipping a renamed clone not only fails at the model stage — if it ever reaches a store, the rename becomes the least helpful piece of evidence in a copyright complaint, because it shows you knew what the original was.

Fourth, if you are wrongly blocked, document before you complain. Save the model's refusal wording, the session time, the project background, and your ownership materials, and report through the channel Osmani opened. One false positive matters little on its own; the same class of false positives recurring is what actually pressures a vendor to adjust the boundary. The first version of any rule is imperfect, and how fast it improves depends on how well the failures are collected and seen.

Verdict: two sentences matter as much as the next sentence

Step back, and what Anthropic did looks very small: two sentences, no new product, no press release, not even a changelog entry. But that is precisely why it is worth watching — vendors are starting to write copyright policy directly into the default instructions of coding tools, rather than leaving it in a terms-of-service PDF. Policy is becoming runtime default behavior. More and more, the first answer to "can I build this?" may come from a quietly delivered prompt, not from a takedown or a lawsuit after the fact.

That turning point is not an endpoint. How much cloning two sentences block, how many legitimate projects they wrongly stop, whether rivals follow, and whether appeals graduate from "message a staff member" to a formal process — none of those answers are in the October 30 sentences. They will emerge over the coming months. For developers, the pragmatic stance is to treat the rule as a default that is already in force, not as a debate to be won: your own assets, clear genre descriptions, and documented ownership are the passports for making games under the new rule.

(Fact-check note: the rule text, the confirmation chain, and Osmani's quote follow Polygon's November 2, 2026 report; server-side identifiers and version details come from multiple contemporaneous reports, as Anthropic did not disclose them in its public changelog; the blocked-project complaint and test results are individual cases from public reports, not systematically verified, and are presented here as cases, not as general claims.)