![Osmani Says Claude’s Game Rule Is 2 Sentences [2026]](https://shattered.io/wp-content/uploads/2026/10/anthropic-osmani-claude-code-two-sentence-game-recreation-rule-2026-gen.webp)
Anthropic’s Claude Code now steers users away from rebuilding existing video games, and as of October 31 the story has moved from a leaked prompt line to a named Anthropic engineer explaining it. Game Developer reported on October 30 that Addy Osmani, listed as a member of technical staff, said the company added two sentences to the Claude Code prompt on Opus 5.5 to push the model away from recreating someone else’s specific game, because using Claude to infringe IP breaks the usage policy. That is the first on-the-record account of what began as a server-side flag.
Our earlier report on the Claude Code game-clone block covered the discovery, when Anthropic had said nothing. This piece picks up from there. It sorts out which claims have a named source, where the reports disagree, and what the rule means for indie developers, modders and anyone who uses an AI coding agent to build games. Readers who follow the wider gaming coverage on shattered.io will recognize the tension: AI tools cut the cost of making games, while rights holders and original studios worry about a flood of lookalikes.
A caution before the details. The fact base here is thin and it comes from a handful of outlets. Where outlets conflict, this article says so. Where a point is analysis rather than reporting, it is labeled as analysis. Nothing below should read as a confirmed company policy beyond what Game Developer, Crypto Briefing and Gameranx attribute to named people or documents.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What Game Developer Reported on Oct. 30 #
Game Developer’s story, bylined to senior editor Bryant Francis, carries a headline that hedges: Anthropic “appears” to have ordered the tool to stop remaking games. The body then does something the first wave of coverage could not. It quotes Osmani explaining the change. According to the outlet, the company added two sentences to the Claude Code prompt on Opus 5.5, and the purpose is to steer the model away from recreating a specific game that belongs to someone else.
Two details in that account matter. First, the change is described as a prompt addition, not a retrained model or a hard filter. A prompt instruction nudges behavior and can be phrased in plain language, which fits the short rule text that users circulated on October 27. Second, Osmani frames the reason as usage policy. He did not describe a new product limit or a technical constraint. He pointed to an existing rule that already forbids using Claude to infringe intellectual property.
Game Developer also relays reader and commenter reaction. Some people noted that anyone determined to clone a game could switch to open-weight models that carry no comparable guardrail. Others questioned how enforcement could work at all, since a model has to judge whether a requested project counts as a recreation of a specific title. Those are open questions, and the outlet presents them as reactions rather than findings.
The Rule as Users Saw It #
The rule surfaced around October 27, when an X user known as Chasm (the handle appears in several outlets) shared that Claude Code had picked up a new instruction. The instruction arrives from Anthropic’s servers under an internal flag named tengu_heron_brook. There is no user-facing setting for it. Nobody can switch it off in a config file or tick a box that says the project is legitimate.
Reports quote the text with small wording differences, which is typical when users copy a prompt by hand. The two shared pieces read as follows. The first says to build original games, games in a genre, and games from the user’s own designs and assets. The second tells the model not to build a recreation of a specific existing game, even if it is renamed or called original. That second clause is the one that drew attention, because it anticipates the obvious workaround of changing the title and calling the result new.
What the rule leaves open
The wording separates a genre from a specific game. A user who wants a roguelike deckbuilder, a top-down shooter or a farming sim is asking for a category, and the rule says to help. A user who asks for a faithful rebuild of one named title is asking for something else. The instruction does not claim to settle every gray case, and no outlet has published a definition of “specific” beyond the rule text.
Timeline: From the Oct. 27 Discovery to the Nov. 12 Policy Date #
The sequence is short, and the dates come from three outlets. Crypto Briefing places the discovery on October 27 and reports that a post on October 30 drew hundreds of thousands of views. Game Developer dates its update to October 30. Crypto Briefing also notes that the rule does not appear in the broader usage policy update that takes effect November 12, 2026. The table lays those points out.
| Date (2026) | Event | Source |
|---|---|---|
| Oct. 27 | Rule spotted in Claude Code, delivered by server flag tengu_heron_brook | Crypto Briefing, Game Developer | | Oct. 30 | Social post about the rule reaches hundreds of thousands of views | Crypto Briefing | | Oct. 30 | Game Developer reports Osmani’s explanation: two prompt sentences on Opus 5.5 | Game Developer | | Oct. 30 | Gameranx frames the rule as a crackdown on vibe-coded ports and decompilations | Gameranx | | Oct. 30 | Crypto Briefing reports no formal company statement and no mention in the policy text | Crypto Briefing | | Nov. 12 | Broader Anthropic usage policy update takes effect | Crypto Briefing |
The two accounts on whether Anthropic spoke are not strictly contradictory. Crypto Briefing describes the absence of a formal, published statement and the absence of the rule in the policy document. Game Developer describes a comment from an individual employee. A named engineer’s explanation is real information, but it is not the same as a policy page, a blog post or a changelog entry. Readers should hold both facts at once.
Anthropic’s Position: Two Sentences and a Usage Policy #
Osmani’s stated rationale is narrow and legalistic. The company already bans using Claude to infringe IP, so the new lines remind the model of that boundary in a domain where users hit it often. In effect, the prompt change turns an abstract policy into a concrete behavior for one popular use case. That fits a pattern in how coding agents get tuned, where a repeated misuse pattern leads to a targeted instruction instead of a policy rewrite.
The scale of the edit is worth stressing. Two sentences is a small change. Game Developer’s framing, with the verb “ordered” softened by “appears,” suggests the outlet itself saw a gap between the headline-friendly story and the modest mechanics. The user-facing effect could still be large, since Claude Code is a daily tool for many developers and a small prompt shift changes what it will agree to build.
Gameranx adds one more detail from the same coverage cycle. It reports that Claude will run a self-assessment as a safety check on instructions, and it ties the game rule to that broader framing around IP infringement and other misuse. The outlet does not quote a company statement for that point, so treat it as the outlet’s description, not an Anthropic announcement. For background on how the company has tightened Claude Code behavior before, see our report on Claude Code 2.1.311 and its permission-skip option.
Why a Server-Side Flag Matters #
The delivery method shapes the whole story. Because tengu_heron_brook arrives from Anthropic’s servers, the company can add, reword or remove the rule without shipping a client update. Users do not see a version bump. They do not read a release note. They find out when the tool declines a request, or when someone inspects traffic and posts what they found, which is how this one surfaced.
That has trade-offs, and this section is analysis. Server-side control lets a vendor react to a legal concern within hours. It also makes behavior changes hard to audit from the outside. A team that built a workflow around Claude Code last week cannot pin the old behavior, and a studio lawyer cannot point to a dated document that says what the tool will and will not do. Developers who care about reproducibility will likely push for changelog-level disclosure of flags like this one.
The absence of a toggle also removes the argument that a user can take responsibility. Many tools handle risky capabilities with an opt-in and a warning. Here the vendor made the call for every user, including those who own the rights to the game they want to rebuild. Studios porting their own back catalog, for instance, are caught by a rule written for strangers. Nothing in the reports says Anthropic planned an exception for rights holders.
The Copyright Logic Behind a Clone Ban #
Copyright law in most jurisdictions protects specific expression such as code, art, music, characters and text. It generally does not protect ideas, rules or bare mechanics. That is why genre games thrive and why a new match-three or a new platformer does not need permission from its predecessors. The line gets blurry when a project copies enough of the look, the levels, the characters and the feel that it stops being inspired by a game and starts being a copy of it.
An AI coding agent complicates this because it can reproduce structure and behavior quickly from a description or from reference material the user supplies. A person can say, build me that game, and the agent assembles mechanics, menus and logic in an afternoon. The risk for the vendor is not that the model copied a file. The risk is that the vendor’s product became the efficient tool for producing something a court or a rights holder would call infringing. Osmani’s wording about infringement aligns with that concern.
This article does not offer a legal opinion on any specific project, and the reports cited here do not either. Whether a given recreation infringes depends on facts, jurisdiction and what was copied. The prompt rule sidesteps that question by drawing a conservative line at the product level, which is cheaper for the vendor than deciding case by case. Cost and legal exposure drive that choice more than any ruling on clones.
Where the Line Falls: Clones, Genres, Ports #
The rule text implies a simple sorting, but real requests are messier. The table below maps the categories the reports describe against the reported treatment. It reflects the rule’s wording, not tested behavior, because no outlet has published a systematic test of how Claude Code responds to borderline prompts.
| Request type | Reported treatment | Basis |
|---|---|---|
| Original game from scratch | Allowed | Rule text |
| Game in a genre (for example a roguelike) | Allowed | Rule text |
| Game from the user’s own designs and assets | Allowed | Rule text |
| Recreation of one named existing game | Steered away | Rule text, Game Developer |
| Same recreation under a new name | Steered away | Rule text (renamed or called original) |
| Ports and decompilation of existing games | Described as in scope by one outlet, untested | Gameranx |
The last row carries the most uncertainty. Gameranx’s headline ties the rule to vibe-coded ports and decompilations, but the rule text itself talks about recreations. A decompilation project that rebuilds source from a shipped binary is different from writing a lookalike from scratch, and fans do both for preservation, modding and platform ports. If the model treats those requests as recreations, some preservation work loses an assistant. If it does not, the rule has less reach than the headlines suggest. Only testing will tell, and no such testing has been published.
Who Gets Hurt and Who Gains #
The reaction split into camps, according to the coverage. Some developers welcomed the rule because they expect fewer low-effort clones competing with original games on storefronts. Others said they were weighing a move to alternative AI models for game work. Both reactions are understandable, and both are forecasts as much as facts.
Winners and losers, as analysis
Original studios and small indies gain the most from a rule that raises the cost of cloning their work. A one-person team that ships a distinctive game used to worry that a copycat could appear within days. The rule does not stop copycats who write code by hand or use other tools, but it removes one fast path. Hobbyists who learn by rebuilding classic games lose a useful exercise, since a faithful remake of a famous title is a common teaching project. Modders and preservationists sit in the middle, and their outcome depends on how strictly the model reads the rule.
The wider backdrop is a games market already adjusting to cheap generation. Our look at Q3 game funding and the share that went to AI tools shows investor money flowing toward the same tooling that makes cloning easy. Capital for AI-assisted production and anxiety about AI-assisted copying are two sides of one trend.
Digital Sun, Roblox and the Clone Pressure #
Two concrete examples show why clone pressure feels real to developers. Crypto Briefing reports that Digital Sun, the studio behind Moonlighter, scaled back public playtesting for its game ReVamp because of cloning concerns. Open playtests give developers early feedback, and pulling back from them costs a small studio information it needs. That is a tangible price paid by a real team, though the outlet does not tie that decision to any AI tool specifically beyond the general concern.
Gameranx points to a platform-side precedent. It reports that Roblox removed ports of Undertale, Halo and Left 4 Dead. Platforms have long policed unlicensed recreations, and the removals show that rights holders and hosts act once a copy becomes visible. A coding agent that declines to build the copy moves the control point earlier in the chain, to creation, rather than waiting for takedown after publication.
Storefront volume adds to the worry. Our report on Steam Early Access launches rising 18 percent to 2,924 shows how many new games already compete for attention. Any tool that lowers the cost of a lookalike pushes more titles into that pile, which is the scenario original developers fear most.
The Open-Weight Escape Hatch #
Game Developer’s commenters raised the most practical objection. The restriction lives in a hosted product’s prompt, so a user who runs an open-weight model locally faces no such instruction. For someone intent on cloning, the rule is a speed bump. The reports do not quantify how many users would switch, and no outlet has published usage data on the matter.
That limits what the rule can achieve, and it also clarifies the purpose. Anthropic is not trying to end game cloning on the internet. A vendor-level instruction manages the vendor’s own exposure and signals to rights holders that the company does not want its product used this way. Whether the signal counts for much in a dispute is a legal question nobody has tested yet. For readers weighing local options, our open-weight model license comparison explains how licenses differ on acceptable use, which matters here because some licenses carry their own restrictions on infringing outputs.
Market Impact for AI Coding Tools and Game Studios #
No outlet reports a measurable market reaction to the rule, and this section does not invent one. What can be said is directional. AI coding vendors now face a choice between maximum permissiveness, which attracts hobbyists and risks legal heat, and targeted restrictions, which protect the vendor and irritate a slice of users. Anthropic picked a narrow restriction in one domain, which suggests it sees game recreation as a high-friction area worth a specific line.
For studios, the news is mildly positive but not decisive. Large publishers have legal teams and takedown processes, so the marginal protection from one vendor’s prompt is small for them. Small teams benefit more in relative terms, though they also lack the resources to monitor copies. Production economics still pressure everyone. As Ken Kutaragi argued in the coverage we summarized, game development costs keep climbing, which is exactly why cheap AI-assisted production tempts both legitimate developers and copycats. Anthropic’s own corporate position adds context. The company is heading toward a public listing, as covered in our piece on the Anthropic S-1 and its five big questions. Companies in that position tend to reduce legal exposure where they can, and a small prompt edit is a cheap way to do it. That reading is analysis, not something any outlet has reported as a motive.
Competitive Comparison: Three Ways to Handle Clone Risk #
The reports do not benchmark rival products, so this comparison is by approach, not by named vendor. It describes the three broad postures a tool or user can take toward recreating existing games, with the strengths and weaknesses that follow from each. It draws on the rule text, the Game Developer commentary about open-weight models, and general reasoning.
| Approach | Who controls it | Strength | Weakness |
|---|---|---|---|
| Hosted agent with server-side prompt rule (Claude Code today) | Vendor | Fast to change, covers all users at once | No user toggle, hard to audit, edge cases unclear |
| Hosted agent with no game-specific rule | Vendor policy only | Maximum flexibility for users | Higher legal exposure for the vendor |
| Open-weight model run locally | User | No vendor guardrail on this use | User carries all legal risk, license terms may still apply |
| Platform takedown after publication | Storefront or host | Acts on real, visible copies (as with Roblox ports) | Reactive, copies spread before removal |
| Rights-holder licensing | Publisher | Clear permission for ports and remakes | Slow, costly, unavailable to hobbyists |
The comparison shows why a prompt rule is attractive to the vendor even though it leaks. It costs almost nothing, it can change overnight, and it shifts the first decision away from the user. Its weakness is that it only governs one product. Rights holders gain real protection only when several layers work together, from the tool to the platform to the courts.
Historical Context: Clones, Remakes and Tooling #
Game cloning is not new. Popular mechanics have spawned imitators for as long as the industry has existed, and the line between homage and copy has been argued in forums, storefront disputes and courtrooms for years. What changed in 2025 and 2026 is the speed. AI coding agents compress weeks of work into hours, which turns a long-standing nuisance into a volume problem.
Tooling vendors have also moved from loose permissiveness toward targeted rules as their products matured. Early coding assistants rarely carried domain-specific instructions about IP. As agents gained the ability to run whole projects, vendors added guardrails for cyber misuse, for secrets handling and now, reportedly, for game recreation. The shift tracks the growth of capability. A tool that autocompletes a line needs fewer rules than a tool that builds a finished game.
Other recent games-industry stories show the same pressure from the producer side. The shutdown coverage in our report on Mass Effect’s cancellation and 400 EA job cuts illustrates how fragile large studios can be. When established teams lose jobs and franchises stall, the argument over who may reproduce their work, and with what tools, gets sharper.
Predictions: What Happens Next #
These are analysis, not reporting, and they carry real uncertainty.
- The rule gets written down. With the usage policy update due November 12, expect pressure for Anthropic to document the game rule or its principle in public text. A flag with no paper trail invites questions.
- Testing reveals the edges. Developers and journalists will probe borderline prompts such as fan remakes, ports and decompilation help. Those tests will show whether the model reads the rule narrowly or broadly.
- Rights-holder exceptions get requested. Studios that want AI help porting their own catalog will ask for a path around the rule. Whether Anthropic offers one is open.
- Rivals pick a side. Competing coding tools will have to say whether they plan similar rules. Some will market permissiveness, others will match the restriction.
- The conversation shifts to output, not intent. Because the model cannot always tell a recreation from a homage, disputes will move toward what the finished project actually copies.
What Developers Should Do Now #
Teams building original games lose nothing under the rule as reported. If Claude Code declines a request, rephrase it around the genre, the mechanics and your own assets, which is what the rule says it supports. Keep design documents and asset provenance on file, because a record that shows a project came from your own designs is useful if anyone later questions it.
Teams that port or remake their own catalog should expect friction and plan for it. Check the November 12 usage policy text when it goes live and watch for any rights-holder carve-out. Hobbyists who rebuild classics for learning can switch to teaching variants, for example a game in the same genre with different art and rules, which stays inside the rule and teaches the same skills. Anyone considering a local open-weight model should read its license first, since acceptable-use terms vary.
FAQ: Claude Code and the Game Recreation Rule #
What exactly did Anthropic change in Claude Code?
Game Developer reports that Addy Osmani said the company added two sentences to the Claude Code prompt on Opus 5.5. The goal is to steer the model away from recreating someone else’s specific game. Users first noticed the text around October 27, delivered through a server flag named tengu_heron_brook.
Did Anthropic issue an official statement?
It depends on the definition. Crypto Briefing reports no formal statement and notes the rule is absent from the usage policy update taking effect November 12, 2026. Game Developer reports an explanation from Osmani, a member of technical staff. No published policy page or blog post covers the rule, based on the coverage reviewed.
Can I still make a game in the style of an existing one?
Per the rule text, yes. It supports original games, games in a genre and games built from your own designs and assets. It tells the model not to build a recreation of a specific existing game, even if renamed or called original.
Is there a setting to turn the rule off?
No. Crypto Briefing and other outlets describe the flag as server-side with no user-facing toggle. Anthropic controls it, and no report says rights holders or other users can opt out.
Does the rule cover ports and decompilations?
Unclear. Gameranx frames the rule as a crackdown on vibe-coded ports and decompilations of existing games, but the rule text speaks of recreations. No published test shows how Claude Code handles those specific requests.
Can people get around it with open-weight models?
Commenters quoted by Game Developer said yes, since open-weight models running locally carry no comparable safeguard. The reports give no data on how many users would switch. Users still carry legal responsibility for what they build, and some model licenses add their own restrictions.
Why would Anthropic add this now?
Osmani’s stated reason, per Game Developer, is that using Claude to infringe IP violates the usage policy. Any broader motive, such as legal caution ahead of a public listing, is analysis and not something the outlets report.