{"slug": "claude-code-game-rule-3-allowed-builds-1-banned-2026", "title": "Claude Code Game Rule: 3 Allowed Builds, 1 Banned [2026]", "summary": "Anthropic added two sentences to the Claude Code system prompt on Opus 5.5 to stop the coding tool from recreating a specific existing video game, a change reported on November 2, 2026 and updated November 3, with coverage from The Verge. The permission sentence reads \"Do build original games, games in a genre, and games from the user's own designs and assets,\" while the restriction sentence reads \"Don't build a recreation of a specific existing game, even if it is renamed or called original.\" Anthropic framed the edit as enforcement of its existing usage policy against intellectual property infringement rather than a new product restriction, leaving open how the prompt-level nudge will be applied in practice.", "body_md": "Anthropic added **two sentences** to the system prompt of Claude Code to steer the tool away from recreating another person’s specific video game. Reports published on November 2, 2026 and updated on November 3 (including coverage from The Verge) describe the change. The company’s stated reason is short: using Claude to infringe intellectual property violates its usage policy.\n\nThe edit applies to Claude Code running on Opus 5.5, and it landed in the middle of a busy week for AI and games. We covered the first wave of reaction in our piece on [why two sentences will not stop the AI game-cloning boom](https://shattered.io/claude-game-cloning-rule-two-sentences-wont-stop-ai-copying-boom-2026/). This article takes a different route. It treats the Claude Code game clone rule as a design document, reads the exact wording, separates what is confirmed from what is not, and asks what it means for developers, studios and rival tools.\n\nThe short version of the Claude Code game clone rule: three kinds of game building stay open, one kind gets discouraged, and a long list of practical questions has no public answer yet. That gap between a tiny rule and a large open question is where the real story sits.\n\nDon't miss new tech stories on Google\n\nAdd Shattered once in the Google app and our stories show up in your news suggestions.\n\n[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)\n\n## The Claude Code Game Clone Rule in Plain Terms\n\nClaude Code is Anthropic’s coding tool. Like most agentic tools, it works from a system prompt, a block of standing instructions the model reads before it reads your request. Anthropic changed that block for Opus 5.5. The change runs two sentences. One tells the model what it should build. The other tells it what to avoid.\n\nThat is a small edit by any measure. A system prompt can run to thousands of words, and two sentences barely change its length. Yet the edit matters because it targets a behavior that developers have noticed for months: ask an AI coding tool for a game, name an existing hit, and the tool will often try to deliver something close to it. Anthropic now tells Claude Code to decline that specific path.\n\nThe company did not frame this as a new product restriction. It framed it as enforcement of an existing one. Anthropic’s usage policy already treats IP infringement as off limits, and the prompt change gives the model a concrete cue for one common case. That distinction shapes everything that follows, because a prompt nudge is a softer and more flexible instrument than a hard technical block.\n\nReaders searching for this story mostly want three answers. What changed, who is affected, and can they still build games with Claude. The answers are: two sentences were added, people trying to recreate a specific existing game are affected, and yes, original and genre games are meant to keep working.\n\n## What the Rule Text Actually Says\n\nThe exact wording is available, and it is worth reading slowly. The permission sentence reads: “Do build original games, games in a genre, and games from the user’s own designs and assets.” The restriction sentence reads: “Don’t build a recreation of a specific existing game, even if it is renamed or called original.”\n\nThree details stand out. First, the permission comes before the restriction, so the model reads the open door before the closed one. Second, the word specific does heavy lifting. The rule does not bar games that resemble others in style or structure. It bars recreating one particular title. Third, the closing clause blocks the obvious workaround. Renaming a clone, or calling it original, does not change the answer.\n\n| Build type | Status under the rule | Basis in the wording | \n|---|---|---|\n| Original game | Meant to keep working | “Do build original games” | \n| Game in a genre | Meant to keep working | “games in a genre” | \n| Game from your own designs and assets | Meant to keep working | “games from the user’s own designs and assets” | \n| Recreation of a specific existing game | Discouraged | “Don’t build a recreation of a specific existing game” | \n| Renamed recreation | Still discouraged | “even if it is renamed” | \n| Recreation labeled original | Still discouraged | “or called original” | \n\nNote what the table leaves out. The rule does not mention fan games, tribute projects, mods, remakes of the user’s own old titles, or parody. Those sit in gray zones, and the two sentences do not resolve them. Developers should expect the model to use judgment there, and judgment varies from one request to the next.\n\n## What Addy Osmani Put on the Record\n\nThe clearest public explanation came from Addy Osmani, a member of technical staff at Anthropic. His statements are the only on-the-record quotes in the reporting we reviewed, so they carry the weight of the whole story.\n\n### The statement on the change\n\nOsmani said: “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.” The verb steer is the key word. A prompt cannot guarantee behavior. It nudges a model toward a pattern, and the model can still miss, or be pushed off the path by a persistent user.\n\n### The statement on what stays allowed\n\nOsmani also addressed the worry that the change would hit legitimate work. He said: “Original games, genre games, and games built from your own designs and assets are meant to keep working.” The phrase meant to keep working is careful. It states intent, not a guarantee, and it leaves room for the false positives that any text-level rule produces. Our follow-up on [how the rule can snag legitimate developers](https://shattered.io/claude-game-clone-rule-legit-developers-troubleshooting-false-positives-2026/) walks through that risk in practical terms, and our earlier note on [Osmani’s two-sentence explanation](https://shattered.io/anthropic-osmani-claude-code-two-sentence-game-recreation-rule-2026/) covers the original statement.\n\n## Why Anthropic Points to Its Usage Policy\n\nAnthropic’s framing is legal and policy-driven rather than technical. The company says the usage policy already prohibits using Claude to infringe intellectual property, and the prompt edit makes the model act on that policy in one domain. That reasoning matters for two reasons.\n\nFirst, it signals that no new law or court order forced the move, at least as far as the reports show. The sources we reviewed do not name a studio, a game title, a legal demand or a lawsuit connected to the change. They also do not mention a licensing agreement. If pressure from a rights holder did drive the decision, nobody has said so on the record.\n\nSecond, it ties enforcement to the company’s own terms rather than to a promise about outcomes. A usage policy is a contract with users. Anthropic can point to it when it declines a request, and it can point to it if a rights holder ever asks what steps the company took. A visible prompt-level cue is a small but concrete step on that record.\n\nLarge AI vendors face growing pressure to show they take copyright seriously. The pressure comes from creators, publishers and regulators, and games are an easy place to act. A game is a defined product with a recognizable name, art style and rule set. A model that rebuilds a named game on request looks bad in a screenshot. A model that politely declines and offers an original alternative looks responsible. The two sentences buy that optics improvement at almost no engineering cost.\n\n## What Stays Allowed: Original, Genre and Own-Asset Games\n\nThe permission sentence matters as much as the restriction. It names three safe lanes, and each fits a different kind of developer.\n\n- **Original games.** A new concept, new characters, new mechanics. This is the cleanest lane and the one Anthropic most wants to protect.\n- **Games in a genre.** A platformer, a roguelike, a card battler, a farming sim. Genres are broad families of play, and no single company owns them.\n- **Games from your own designs and assets.** If you drew the art, wrote the design document and own the rights, Claude Code is meant to build it with you, even if your design resembles a well-known style.\n\nThe genre lane is the one most developers will test. Genre conventions are shared property in the games industry. Health bars, double jumps and loot tables belong to no one, and a rule that blocked them would break most game development. By naming genre games as allowed, Anthropic tells the model that building in the style of a category is fine. The line sits at the point where a request stops describing a category and starts describing one product.\n\nFor solo developers and small teams, this is the reassuring half of the story. AI-assisted game development has surged. Our report on [new game pages on Itch.io doubling to more than 6,000 a day](https://shattered.io/itch-io-new-game-pages-double-6000-per-day-ai-2026/) shows how much volume small creators now push through. Most of that output is original or genre work. The new rule is not aimed at it.\n\n## The Renamed-Clone Loophole the Rule Closes\n\nThe final clause of the restriction, even if it is renamed or called original, targets the most common workaround. Before the change, a user could ask for a famous game with the names swapped out and the sprites redrawn. The mechanics, level order and structure would match, but the title screen would not. A text filter that only looked for trademarked names would miss that request entirely.\n\nAnthropic’s wording shifts the test from names to substance. The question is whether the thing being built is a recreation of a specific existing game. A model can reason about that question using the structure of the request, not just a keyword list. That is a harder job, and it is also the only approach that holds up when users get creative.\n\nIt also creates a new class of edge case. Suppose a user describes a game in detail without naming it, listing every mechanic, level layout and enemy behavior of a well-known title. Is that a request for a recreation, or a request for a game in a genre? The two sentences do not say, and the reports do not describe a test case. Expect developers to probe this boundary and share results on forums within days.\n\nIntellectual property law has long separated ideas from expression. Game rules and general mechanics get thin protection, while specific art, characters, code and distinctive expression get more. A model has no easy way to apply that doctrine in real time. The prompt rule is a rough proxy for it, and rough proxies misfire in both directions.\n\n## What the Reports Do Not Confirm\n\nSeveral details that readers will want remain unconfirmed. We list them here so the line between fact and guesswork stays visible.\n\n| Detail | Status | Note | \n|---|---|---|\n| Two sentences added to Claude Code’s system prompt | Confirmed | Statement from Addy Osmani | \n| Change applies to Opus 5.5 | Confirmed | Named in Anthropic’s statement | \n| Reason is the usage policy on IP infringement | Confirmed | Stated by Anthropic | \n| Report date | Confirmed | November 2, 2026, with updates on November 3 | \n| Exact effective date | Unconfirmed | No date given in the reports | \n| Applies to every Claude product | Unconfirmed | Only Claude Code is named | \n| Rule can be disabled | Unconfirmed | One report attributes the claim to a user, not Anthropic | \n| Named games, companies or legal demands | Unconfirmed | None connected to the change | \n| Lawsuit or licensing deal | Unconfirmed | None reported | \n| Price or usage-limit change | Unconfirmed | None reported | \n\nThe disable question deserves a second look. One report relays a claim that the restriction can be turned off. That claim traces to a user, not to Anthropic, and the company has not confirmed it. Treat it as a rumor until Anthropic documents a setting. If it proves true, it would change the picture a lot, because a switch that lets users bypass the rule turns a policy into a speed bump.\n\nThe scope question matters just as much. Osmani’s statement names Claude Code on Opus 5.5. Nothing in the reports says the same two sentences appear in the Claude chat apps, in the API with a custom system prompt, or in other coding surfaces. Developers who build their own agents on the API control their own prompts, so a Claude Code edit does not automatically reach them.\n\n## Why a System Prompt Is the Tool Anthropic Chose\n\nAnthropic could have tried other levers. It could have retrained the model, added an output classifier, or blocked certain requests at the API layer. It picked a prompt edit instead, and the reasons are practical.\n\nA prompt edit ships fast. It needs no new training run and no new model release. A team can write two sentences, test them against a set of sample requests, and push the change the same day. For a fast-moving policy issue, that speed beats the slower and more thorough options.\n\nA prompt edit is also easy to adjust. If the rule blocks too many legitimate requests, engineers can loosen the wording. If it misses clones, they can tighten it. Osmani’s careful phrase about original and genre games being meant to keep working suggests the team expects to tune the rule as feedback arrives.\n\nThe tradeoff is reliability. Prompts work through persuasion, not enforcement. A long conversation, a clever framing, or a request split into small harmless steps can pull a model away from its standing instructions. Security researchers have shown for years that prompt-level rules leak under pressure. Anthropic’s own wording admits as much by saying the edit is meant to steer the model, not to bar it.\n\nThat is why the model version matters. The rule is tied to Opus 5.5. Anthropic tuned the sentences for how that model reads instructions. Our [comparison of Argon and Opus 5.5 on DeepSWE](https://shattered.io/gemini-4-argon-vs-claude-opus-5-5-coding-deepswe-77-9-vs-74-2-2026/), which put the two at 77.9% and 74.2%, shows why Opus 5.5 is a natural target. It is a leading coding model, so the behavior of its default prompt reaches a lot of real projects.\n\n## Historical Context: Game Clones Before AI\n\nGame cloning is older than any chatbot. The industry has fought it since the arcade era, and the legal record gives useful context for Anthropic’s wording.\n\nOn May 30, 2012, a federal court in New Jersey granted summary judgment to the owner of Tetris in its case against Xio Interactive, maker of the iPhone game Mino. The court found substantial similarity in the color, shapes, style and movement of the blocks, so the copying went beyond the idea of falling blocks. The case became a reference point because it showed that a clone can lose even when its name differs, as long as it copies the distinctive expression of the original.\n\nThat same year, Spry Fox sued 6waves Lolapps in January, accusing the company of cloning its puzzler Triple Town with an iOS game called Yeti Town. The companies settled in October 2012. Spry Fox took ownership of the Yeti Town IP, and the game came off sale, according to trade coverage including Game Developer. The dispute pushed many indie developers to take clone defense seriously. Both cases share the logic now visible in Anthropic’s second sentence: renaming does not cure copying.\n\nMobile storefronts spent the following decade flooded with knock-offs. Platform holders added takedown processes, but enforcement stayed slow and uneven. Cloning never needed AI. It needed time, an artist and a programmer. What AI changes is cost and speed. A solo person can now produce a near-copy in hours, which multiplies the volume that rights holders must police.\n\nThat shift explains why a tool vendor, rather than a storefront, is stepping in. Catching a clone after release is expensive. Declining to write it in the first place costs far less. The Claude Code rule moves one small piece of enforcement upstream, to the moment of creation.\n\n## Market Impact: Indie Tools, Asset Stores and AI Pipelines\n\nThe reports give no price, usage-limit or revenue figure tied to the change, so any market impact is an inference. Still, the direction is clear enough to discuss.\n\nFor rights holders, the rule is a small positive. A major AI coding vendor now states, in the instructions its model reads, that recreating a specific game is out of bounds. Studios and publishers can cite that stance in conversations with other AI companies. Even an imperfect prompt-level rule resets expectations about what a responsible vendor does.\n\nFor indie developers, the effect is mixed. Those building original and genre games lose nothing, in principle. Those who relied on AI to prototype quick copies of a hit, for learning or for a tribute project, may find Claude Code less cooperative. Some will move to other tools. Others will rewrite requests to describe mechanics in generic terms, which is the behavior the rule tries to discourage.\n\nFor large studios experimenting with AI, the picture is different. Executives have made bold claims about speed. Our coverage of [an Xbox manager’s claim about building a three-hour game in two days](https://shattered.io/xbox-manager-stirrup-ai-game-development-three-hour-game-2026/) shows how fast the sector is moving. Big publishers building their own titles with AI face no conflict with the new rule, since they own their designs and assets. The friction lands on third parties who want to copy someone else’s work.\n\nFor competing AI vendors, the move sets a marker. If Anthropic gets credit for acting, rivals will face questions about their own defaults. A company whose coding tool still happily rebuilds named games on request will look behind. Expect journalists and developers to test competing tools with the same prompts within days of the news.\n\n## How Different Approaches to Clone Control Compare\n\nThe reports do not describe how rival tools handle game clones, so we do not claim to know their current policies. What we can do is compare the general enforcement approaches an AI vendor has available. The table below is our own framework, not an Anthropic document.\n\n| Approach | How it works | Strength | Weakness | \n|---|---|---|---|\n| System-prompt steering | Standing instructions tell the model what to avoid (the route Anthropic reports using) | Fast to ship and easy to tune | Can leak under persistent or creative prompting | \n| Usage-policy enforcement | Terms of service ban infringement, with account action after the fact | Creates a contractual basis | Slow and reactive | \n| Output classifiers | A separate system scans results for matches to known works | Harder to bypass with wording | Costly and prone to false positives | \n| Storefront takedowns | Platforms remove clones after rights holders complain | Targets the finished product | Clones ship first and get removed later | \n| Licensing deals | Vendors pay rights holders for permitted use | Clear legal footing | Hard to scale across thousands of titles | \n\nAnthropic chose the cheapest and fastest option on the list. That does not mean the company will stop there. A layered setup, with a prompt cue at the front and policy enforcement behind it, is the usual direction for mature trust-and-safety programs. The reports confirm only the first layer.\n\n## What Developers Should Do This Week\n\nPractical guidance beats speculation. Because the Claude Code game clone rule is about intent, not keywords, how you phrase a request matters. If you use Claude Code to build games, a few habits will keep you out of trouble and keep your projects moving.\n\n### If you build original or genre games\n\nDescribe your game on its own terms. Lead with your mechanics, your setting and your art direction, and skip comparisons to a named hit. Genre words are fine. A prompt that says you want a turn-based tactics game with a grid map and permanent character loss describes a genre. A prompt that asks for a copy of one named title describes a recreation. Anthropic says the first kind is meant to keep working.\n\n### If you build tributes or learning projects\n\nExpect friction. Learning exercises that rebuild a classic from scratch are a long-standing tradition in programming education, and the rule does not carve out an exception. Rework the exercise around your own designs and assets, or tell the model which mechanics you want to study and build them in an original wrapper. Keep records of your own art and design files, since the permission sentence names games from the user’s own designs and assets.\n\nTeams should also review their safety habits more broadly. Our guide to [five vibe coding safety practices](https://shattered.io/vibe-coding-safety-practices-engadget-five-2026/) covers the review steps that apply whenever an AI tool writes code for you. Provenance questions, meaning who made the assets and who owns the design, now belong on that checklist next to security reviews.\n\nFinally, do not count on workarounds. The reports include a claim, attributed to a user, that the restriction can be disabled. Anthropic has not confirmed it. Even if a bypass exists, shipping a clone built that way carries the same legal risk as any other clone, and the usage policy still applies to you.\n\n## Five Predictions and the Takeaway on AI Game Cloning\n\nThese forecasts are our analysis, not reported facts. We label them as judgment calls, and any of them could prove wrong.\n\n1. **Developers will publish bypass tests within days.** Forums and social feeds will fill with screenshots of prompts that slip past the rule and prompts that block legitimate work. Anthropic will likely adjust the wording in response.\n2. **Rival tools will face the same test.** Reporters will run the same game-recreation prompts on competing coding assistants. Vendors that fail will feel pressure to publish their own rules.\n3. **Anthropic will document the scope.** The open questions on effective date, product coverage and disabling are easy to answer in a help page. A clarification is the cheapest way to end speculation.\n4. **The fight will move from prompts to classifiers.** If clone requests keep slipping through, vendors will add output-side checks that look at what the model produced, not just what the user asked.\n5. **Rights holders will treat AI vendors as a new enforcement target.** Studios that spent a decade pursuing app-store clones will start asking AI companies for written policies, and some will push for licensing terms.\n\nThe common thread is pressure. One company’s two-sentence edit raises the baseline for everyone else. Once a leading vendor states a position, silence from the rest of the field reads as a position too.\n\nThe takeaway: with the Claude Code game clone rule, Anthropic made a small, fast, adjustable change, and it described it in modest terms. Claude Code on Opus 5.5 is now told to build original games, games in a genre, and games from the user’s own designs and assets, and to avoid recreating a specific existing game even when renamed. That is the entire confirmed substance.\n\nThe unconfirmed list is longer than the confirmed one. Nobody has said when the rule took effect, whether it covers products beyond Claude Code, or whether users can switch it off. No studio, title, legal demand, lawsuit or licensing deal appears in the reports. Readers should hold those gaps in mind before they repeat claims about what the rule does or why it exists.\n\nFor anyone building games with AI help, the practical message is calm. Original work and genre work should continue as before. Copying a named game should not. And for the wider industry, the episode shows how AI vendors may handle IP in the near term: through quick, visible, adjustable rules written into the instructions the model reads, with tougher tools held in reserve. For more coverage, visit our [gaming section](https://shattered.io/gaming/).\n\n## Frequently Asked Questions\n\n**What is the Claude Code game clone rule?**\n\nAnthropic added two sentences to Claude Code’s system prompt on Opus 5.5. They steer the model away from recreating another person’s specific video game, and they tell it to build original games, games in a genre, and games from the user’s own designs and assets.\n\n**Can I still build games with Claude Code?**\n\nYes. Addy Osmani, an Anthropic member of technical staff, said original games, genre games, and games built from your own designs and assets are meant to keep working.\n\n**Why did Anthropic add the rule?**\n\nThe company says using Claude to infringe intellectual property is against its usage policy. The prompt change gives the model a concrete cue to avoid one common case, recreating a specific existing game.\n\n**Does the rule apply to every Claude product?**\n\nUnconfirmed. Anthropic’s statement names Claude Code on Opus 5.5, and the reports do not say whether other products carry the same two sentences.\n\n**Can the restriction be turned off?**\n\nUnconfirmed. One report attributes a claim that it can be disabled to a user, not to Anthropic. The company has not documented any such setting.\n\n**Did a game company or lawsuit trigger the change?**\n\nThe reports do not say so. They name no studio, game title, legal demand, lawsuit or licensing agreement connected to the change.\n\n**When did the rule take effect?**\n\nThe exact effective date is unconfirmed. Coverage appeared on November 2, 2026, and some reports were updated on November 3.\n\n**Does renaming a clone make it acceptable?**\n\nNo. The restriction text says not to build a recreation of a specific existing game, even if it is renamed or called original.", "url": "https://wpnews.pro/news/claude-code-game-rule-3-allowed-builds-1-banned-2026", "canonical_source": "https://shattered.io/claude-code-game-clone-rule-three-allowed-builds-one-banned-2026/", "published_at": "2026-11-03 05:58:33+00:00", "updated_at": "2026-11-03 06:20:45.627167+00:00", "lang": "en", "topics": ["ai-policy", "ai-tools", "generative-ai", "ai-agents"], "entities": ["Anthropic", "Claude Code", "Opus 5.5", "The Verge"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/claude-code-game-rule-3-allowed-builds-1-banned-2026", "markdown": "https://wpnews.pro/news/claude-code-game-rule-3-allowed-builds-1-banned-2026.md", "text": "https://wpnews.pro/news/claude-code-game-rule-3-allowed-builds-1-banned-2026.txt", "jsonld": "https://wpnews.pro/news/claude-code-game-rule-3-allowed-builds-1-banned-2026.jsonld"}}