Claude Game-Clone Rule Snags Legit Devs: 2 Lines [2026]

Anthropic's Claude Code version 2.1.314 carries a new instruction, internally labeled `tengu_heron_brook`, that tells the model not to build a recreation of a specific existing video game even when the user renames it, a change first reported around October 27, 2026 and confirmed by an Anthropic staffer. Tom's Hardware reported the rule blocks game cloning and reverse engineering, while users and GitHub issue reporters say plain reverse-engineering and troubleshooting requests are being refused as collateral damage; PC Gamer tied the change to weeks of AI-generated clutter on Steam. The rule reaches users through Anthropic's servers rather than a local install, and the complaints about overreach remain user claims rather than findings from Anthropic or an independent audit.

Claude now refuses to rebuild specific existing video games, and the fallout is landing on people who never planned to clone anything. Tom’s Hardware reported that Anthropic’s model is blocking game recreation and cloning, while users say plain reverse-engineering and troubleshooting requests are getting caught in the same net. PC Gamer, for its part, tied the change to weeks of AI-generated clutter on Steam. This story is about that second half: the collateral damage. We first flagged the Claude Code rule https://shattered.io/claude-code-blocks-video-game-recreation-tengu-heron-brook-2026/ when the setting surfaced inside the tool’s binary. Since then, more reporting has arrived, an Anthropic staffer has confirmed the change, and developers have started posting concrete examples of work that stopped halfway. This analysis sorts what is confirmed from what remains a user claim, explains how the mechanism works, and lays out what it means for studios, hobbyists and the AI coding market. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F Claude Blocks Game Cloning: What Happened This Week The first signs appeared around October 27, 2026. Reports describe a new instruction inside Claude Code version 2.1.314 that tells the model not to build a recreation of a specific existing video game, even when the user renames it or calls it original. The instruction sits under an internal label, tengu heron brook , that earlier reporting links to remote changes in Claude Code’s system instructions. By October 30 the story had spread across gaming and developer press. Tom’s Hardware led with the broad framing that Claude blocks game cloning and reverse engineering. PC Gamer focused on the Steam angle. XDA Developers described the move as a test of a possible crackdown on vibe-coded ports of existing games. Smaller outlets, including Techtroduce and tBreak, ran headlines on the same restriction. Three points anchor everything else in this piece. First, the rule targets recreations of specific existing games. Second, the rule reaches the model through Anthropic’s servers, so users did not have to install anything to feel it. Third, the complaints about overreach come from users and GitHub issue reporters, not from Anthropic or an independent audit. Keep those three separate and the rest of the debate gets easier to follow. What the Rule Says and What It Does Not According to the reports, the instruction has two sides. One side encourages building original games, games within a genre, and games made from the user’s own designs and assets. The other side forbids building a recreation of a specific existing game. The rule is short, which is exactly why people disagree about where it ends. What stays allowed A platformer, a roguelike or a deck builder built from scratch remains fair game. So does a project built on a developer’s own characters, art and audio. Genre work matters most here. Almost every shipped game borrows loops from earlier ones, and the rule as reported keeps genre-level inspiration on the permitted side. What gets refused A request to rebuild a named commercial title is the target. The wording reported by the press closes the obvious loophole too, since renaming the clone does not change the outcome. The text of the rule does not mention reverse engineering, decompilation, modding or debugging, and that silence is where the false positives come from. | Request type | Reported outcome | Basis | |---|---|---| | Original game from scratch | Allowed | Rule text as reported | | New game in an established genre | Allowed | Rule text as reported | | Game built from the user’s own designs and assets | Allowed | Rule text as reported | | Recreation of a specific existing game | Refused, even if renamed | Rule text as reported | | Pokémon Diamond decompilation help | Agreed to help in one test | Reported test with Opus 5.5 | | Replicating a single game mechanic | Refused in some user reports | GitHub issue complaints | | Patching a tabletop-style game | Refused in some user reports | GitHub issue complaints | How the Server-Side Switch Works The detail that surprised many developers is where the rule lives. Reports say it is delivered from Anthropic’s servers rather than baked into the installed tool. Claude Code builds its system prompt from pieces, and some pieces can be changed remotely. Anthropic can add, tweak or pull a sentence without shipping a new version. That design has real advantages. A company can react to a legal complaint or an abuse pattern in hours instead of waiting for users to update. It also explains why people noticed the change by reading the binary, since the label for the mechanism was already there. The same flexibility carries a cost. A developer who tested a workflow on Tuesday can find it broken on Thursday with no changelog entry to explain why. Prompt-level rules also behave differently from hard filters. A hard filter blocks a match every time. A prompt instruction nudges the model, and the model then judges each request on its own. That produces inconsistent results. One user may get a refusal for describing a mechanic, while another gets full help with a near-identical request. The inconsistency fits the pattern in the complaints, where refusals show up halfway through a project rather than at the first prompt. Anthropic’s Position: IP Infringement Is Already Against Policy Reporting indicates that Addy Osmani, a member of Anthropic’s technical staff, confirmed the change publicly. His explanation, as the press relayed it, runs like this: the team added two sentences to the Claude Code prompt on Opus 5.5 to steer the model away from recreating someone else’s specific game, because using Claude to infringe intellectual property already breaks Anthropic’s usage policy. In other words, the company frames this as enforcing an existing rule, not writing a new one. We covered that confirmation in detail in our look at Osmani’s two-sentence explanation https://shattered.io/anthropic-osmani-claude-code-two-sentence-game-recreation-rule-2026/ . The takeaway for this piece is narrower. A staff confirmation settles that the rule is real and intentional. It does not settle how broadly the model applies it. The reports we reviewed describe no formal blog post or written policy update from Anthropic that lists exactly which game-related requests are now off limits. That gap matters for trust. Developers can plan around a written policy. They cannot plan around a two-sentence prompt whose boundaries they have to discover by hitting them. Anthropic could close the gap quickly with a short public explainer that names examples on each side of the line. Whether it will remains open, and nothing in the current reporting says it has committed to doing so. Where Legitimate Developers Report Getting Caught Tom’s Hardware framed the issue around users who say non-malicious, generic reverse-engineering tasks are blocked. The Claude Code GitHub tracker reportedly filled with complaints after the change, describing refusals during ordinary game work. Treat these as user reports. Nobody has published a controlled test showing how often the false positives occur. Mid-project refusals One reported pattern is a refusal that arrives after hours of productive work. A developer builds a prototype with Claude’s help, then asks for a feature, and the model stops because the feature resembles a known title. This hurts more than a refusal at the start. The user has already invested time, and the earlier steps showed the project was harmless. Mechanics and tabletop-style projects Users also report refusals when they try to replicate a single mechanic, and when they patch a tabletop-style game. Mechanics are the hardest case to police. Game rules and ideas sit at the edge of what intellectual property law protects, and the original expression, meaning code, art and text, sits well inside it. A model working from a two-sentence prompt cannot reliably tell the two apart. Troubleshooting is the third bucket. People who maintain mods, fan patches or compatibility fixes often need to read how an existing game behaves. Asking an assistant to explain a crash in a shipped title can look, to a cautious model, like a step toward recreating it. Tom’s Hardware’s headline groups this under basic troubleshooting tasks, and that is the part most likely to push users toward other tools if it persists. Reverse Engineering Versus Recreation: The Line Opus 5.5 Draws One reported test sheds light on the intended boundary. When a user asked whether Claude could help with a Pokémon Diamond decompilation project, Opus 5.5 agreed. It explicitly separated reverse engineering from recreating a game. That suggests the instruction is not meant to ban reverse engineering or porting outright. The distinction has a logic. A decompilation project studies how an existing program works and rebuilds it for learning, preservation or interoperability. A recreation project aims to ship a stand-in that plays like the original. The first is long-standing hobbyist practice. The second competes with the rights holder. A reasonable reading of Anthropic’s rule keeps the first and blocks the second. One agreeable answer in one test proves little, though. The same model on the same day can refuse a similar request, and the complaints suggest it sometimes does. The preservation and porting community knows this tension well. Tools such as the one we covered in our report on Switch games running on Mac https://shattered.io/switch-games-on-mac-nxnative-85-titles-wine-like-tool-2026/ sit in the gray zone that this rule now touches, even though they are not clones in the sense Anthropic describes. Why Now: Steam, AI Slop and Studio Pressure PC Gamer connected the update to weeks of Steam being flooded with low-effort AI-built games. That context explains the timing. When a store fills with near-copies of popular titles, rights holders complain, storefronts get blamed, and the AI vendor behind the tools becomes a target. Anthropic is preparing for a public listing, as our coverage of the Anthropic S-1 and its five big questions https://shattered.io/anthropic-s-1-80-7b-loss-five-big-questions-ipo-2026/ shows, and a company in that position has strong reasons to cut legal exposure early. Steam itself was already seeing heavy launch volume. Our look at Early Access launches rising 18% to 2,924 https://shattered.io/steam-early-access-2924-launches-up-18-percent-player-expectations-2026/ shows how crowded the storefront has become, though that count reflects all early access releases and not only AI-assisted ones. Nobody has published a reliable figure for how many Steam titles are Claude-built clones, so any such number should be treated as unverified. | Outlet | Angle | What it adds | |---|---|---| | Tom’s Hardware | Blocking spreads to legitimate tasks | User claims that generic reverse engineering and troubleshooting get refused | | PC Gamer | Steam AI slop backdrop | Links the update to weeks of clutter on the storefront | | XDA Developers | Possible crackdown on vibe-coded ports | Frames the rule as a test, not a settled policy | | Techtroduce, tBreak | Straight news | Headlines describe Claude Code restricted from recreating existing games | | Game Developer | Developer-side reaction | Cited in the story brief, text not independently reviewed here | Market Impact: AI Coding Tools Meet Game Development Game studios have adopted AI tooling quickly. Our report on Q3 game funding https://shattered.io/game-funding-q3-2026-drake-star-1-62-billion-81-percent-ai-tools-2026/ cited Drake Star data putting the quarter at $1.62 billion, with 81 percent going to AI tools. That money flows toward products that write code, generate art and automate testing. A model that hesitates on anything resembling an existing title can slow exactly the workflows those investors are paying for. Consider who uses an assistant for game work. Indie developers prototype faster with it. Modders lean on it to understand engine quirks. Students rebuild classic games as exercises, a practice teachers have used for decades. The last group hits the rule head on, because a class project that recreates a well-known arcade title is exactly what the instruction describes. Rights holders gain from the change, at least in the short run. Studios that watch their franchises get cloned have asked platforms for more protection, and a vendor-side guardrail is cheaper for them than case-by-case takedowns. The tradeoff falls on the developer community, which now has to wonder whether its primary assistant will stall on a legitimate task. Competing tools gain a talking point if the false-positive reports hold up. Code quality questions travel alongside this debate. Opus 5.5 also sits at the center of security research, such as the finding we covered where Opus 5.5 and GPT-6-Sol tied at 73 percent https://shattered.io/ai-generated-code-security-veracode-claude-opus-5-5-gpt-6-sol-73-percent-2026/ in an AI code security test. Studios judge these assistants on safety, accuracy and friction together, and a new source of friction shifts the balance. Competitive Comparison: Three Ways Vendors Handle IP Risk The reports we reviewed describe no equivalent rule from rival coding assistants, so this section compares approaches rather than named products. We could not confirm the current behavior of competing tools, and readers should check each vendor’s usage policy directly. The first approach is a prompt-level instruction, which is what Anthropic reportedly used. It ships fast, adjusts easily and produces uneven results. The second approach is a policy-only stance, where the terms of service ban infringement but the model does little to stop it. That path keeps friction low for users and pushes legal risk onto them. The third approach is a hard classifier that blocks matches for protected works. It gives consistent outcomes but needs a reference catalog and tends to over-block when the catalog is broad. Anthropic’s choice sits in the middle. It enforces something without building a catalog of every game, which keeps costs down and makes errors unpredictable. For a vendor weighing an IPO, a middle path reads as responsive without being heavy. For a developer, it reads as a rule you only learn by tripping over it. Historical Context: Clones, Emulators and the Courts Clone disputes are older than the AI era. Arcade-era knockoffs, shareware lookalikes and mobile copies of hit puzzle games all drew legal threats. The recurring lesson from those fights is that rules and ideas get thin protection while specific art, code and presentation get strong protection. Courts tend to side with rights holders when a copy reproduces the look and feel of a particular game, and against them when a copy only shares a genre. Reverse engineering has its own record. Emulator and compatibility projects have won some legal arguments and lost others, depending on how they handled copyrighted code and encryption. Our piece on the unconfirmed Mario Kart World emulator claim https://shattered.io/mario-kart-world-pc-emulator-eglantine-claim-unconfirmed-switch-2-jailbreak-2026/ shows how quickly such stories surface around any big Nintendo release, and how cautious reporting needs to be. What changes now is scale and speed. A person once needed months and real skill to clone a game. An assistant compresses that into an afternoon. Rights holders know it, which is why vendor-side limits appeal to them. The older legal tests still apply, but the practical barrier that protected studios, the labor cost of copying, has dropped sharply. What Developers Can Do Today Legitimate users have a few practical options. None of them involve getting around the rule. They help the model understand what you are actually doing. - State your project’s nature up front. Say whether the work is original, built from your own assets, or a study of a game you own. - Describe the bug or behavior in plain technical terms and skip the game title when the title adds nothing. A crash in a memory routine is a memory problem. - Separate analysis from recreation. Ask for an explanation of how a file format works rather than for a rebuilt game. - Keep a second tool ready for tasks that stall, and read each tool’s usage policy before building on it. - Track the Claude Code GitHub issues for confirmed examples, since reports there show where the line falls in practice. Teachers and students need one extra step. Frame clone exercises as original projects in a genre, with new art and new rules, and the work lands on the permitted side. Our gaming section https://shattered.io/gaming/ follows the developer-tooling story as it develops. Five Predictions for the Next 90 Days These are analysis, not reported facts. They rest on how similar vendor guardrails have played out and on the details above. 1. Anthropic publishes clearer guidance. Pressure from developers and the press makes a written explainer with examples likely, probably framed as a usage policy clarification. 2. The prompt gets reworded. Because the rule is server-side, Anthropic can tune the language quietly, and false-positive complaints give it a reason to do so. 3. Rivals add similar language. If studios lobby vendors, other coding assistants will likely adopt comparable instructions, though we have no confirmation of any such move today. 4. Rights holders cite the change. Expect studios and trade groups to point to the rule as a standard other vendors should meet. 5. A reverse-engineering carve-out emerges. The reported Pokémon Diamond test hints that the model already tolerates decompilation, so a formal exception for preservation and interoperability work is plausible. The Takeaway for Studios and Hobbyists The takeaway: the rule is real, it is intentional, and it works through a flexible server-side prompt that Anthropic can change at will. Its stated goal, stopping direct game recreation, has broad sympathy in the industry. Its side effects remain unmeasured, and the evidence for them comes from user reports and GitHub complaints, not from a published test. For studios, the change is a modest win with an open question about enforcement. For hobbyists, modders and students, it is a source of friction that may ease or worsen depending on how Anthropic tunes the wording. For the wider market, it shows that AI vendors now tune their assistants around intellectual property pressure as well as safety. Watch the Claude Code release notes, the GitHub tracker and any formal Anthropic statement over the coming weeks. Frequently Asked Questions Does Claude block game cloning in every product? The reporting we reviewed centers on Claude Code, where the instruction was found. Press coverage also describes Claude behavior more broadly, but we could not confirm how the rule applies in other Anthropic products. Can I still build a game in the style of a famous title? Games within a genre remain allowed according to the reported rule text. A game that recreates a specific existing title, even under a new name, falls on the refused side. Is reverse engineering banned now? No ban appears in the reported rule text. In one reported test, Opus 5.5 agreed to help with a Pokémon Diamond decompilation project. Other users say their reverse-engineering requests were blocked, so results vary. Has Anthropic officially commented? An Anthropic technical staff member, Addy Osmani, reportedly confirmed that two sentences were added to the Claude Code prompt on Opus 5.5 and tied the change to the usage policy on IP infringement. We found no formal blog post or written policy update. Why did users notice the change so fast? The mechanism label was visible inside Claude Code 2.1.314, and refusals showed up in everyday use. Reports say the rule is delivered from Anthropic’s servers, so it took effect without an update. What should I do if Claude refuses a legitimate task? Restate the task with clear context about ownership and purpose, describe technical problems without leaning on a game title, and report persistent false positives on the Claude Code GitHub tracker, where developers are logging examples. Are the false-positive numbers known? No. Reports of blocked legitimate work come from users and issue threads. Nobody has published a measured rate, so treat claims of widespread failure as unproven until a controlled test appears.