Anthropic Blocks Claude Code From Recreating Existing Video Games

Anthropic has blocked its coding assistant Claude Code from recreating existing video games. Anthropic employee Addy Osmani confirmed on X that two sentences in Claude Code’s system prompt for Opus 5.5 now prevent the cloning of other people’s games, according to The Decoder.

The change matters to anyone who uses AI coding agents for client or in-house projects, because it sets a boundary on what the tool will build and arrives with a stated intention to adjust the rules further.

• Two sentences in the Opus 5.5 system prompt of Claude Code now stop it from cloning other people’s games.
• Anthropic employee Addy Osmani confirmed the rule on X and said IP infringement violates the usage policy.
• The block applies even if a clone is renamed or presented as an original work.
• Users can still build original games, genre titles and projects with custom assets, though some report problems.
• Claude Pro costs US$20 a month and includes Claude Code; Max starts at US$100 a month.

Anthropic now stops Claude Code from cloning existing games

Anthropic is preventing Claude Code from recreating existing video games. The restriction lives in the system prompt the tool uses with Opus 5.5. A system prompt is the set of standing instructions a model receives before any user request, and it shapes how the model responds to everything that follows.

Osmani, an Anthropic employee, confirmed on X that two sentences in that prompt are responsible for the behavior. He said that using Claude to infringe on intellectual property violates the usage policy. The statement frames the change as an enforcement of an existing rule rather than as a new category of prohibition.

According to The Decoder, the rule has been active for several days. Neither the exact date of its introduction nor the wording of the two sentences is given in the available reporting. How the instruction is phrased, and therefore how broadly the model interprets it, is not known.

The stated target of the Claude Code copyright rule is narrow: reproducing games that already exist. Nothing in the reporting presents the move as a ban on game development, and the section on what remains allowed covers that point in detail.

The confirmation came from an individual employee on X, not from a formal company announcement. The facts available contain no separate statement from Anthropic as a company, no changelog entry and no published Opus 5.5 usage policy text. Everything known about the rule comes from the statement by Addy Osmani on Claude Code, as relayed by The Decoder.

How the block works and what it covers

The rule holds even when a clone is renamed or passed off as an original work. A request that swaps the title, changes the character names or describes the result as a fresh creation does not alter the outcome. The restriction is aimed at the substance of what is being rebuilt, not at the label attached to it.

That detail closes the most obvious workaround. Without it, a prompt such as “make a game just like this one, but call it something else” would be a trivial way around the rule. The reporting does not describe how the model decides whether a request amounts to a clone, and no test or threshold is given.

The location of the rule is worth noting. It sits in Claude Code’s Opus 5.5 system prompt, not in the underlying model. Claude models are Anthropic’s large language models, and they are separate from both the Claude app and Claude Code. A model is the engine; the app and the coding tool are products built around it, each with its own instructions.

Conflicting accounts of where the rule lives

Not every source agrees with the system prompt account. Some outlets, among them Crypto Briefing and BigGo Finance, describe the restriction as a server-side setting rather than plain prompt text. If that is accurate, the rule would be enforced by Anthropic’s own infrastructure and would not appear as instructions anywhere a user can read them.

The wording is disputed too. Several reports quote a version of the rule that they attribute to a user on X, not to Anthropic. Osmani’s confirmation refers to two sentences in the system prompt, and Anthropic has not published either text. Until it does, we suggest you treat any quoted wording as a secondhand account, and the system prompt explanation as the only one confirmed by an Anthropic employee.

The Claude API is a third route: it lets applications integrate Claude models programmatically, meaning a developer’s own software can send requests to the model directly. According to praxis-agents.ai, that is how Claude models can be built into other applications. The reporting on the game rule describes it only for Claude Code with Opus 5.5. Whether the same restriction applies in the app or through API access is not stated.

How Claude Code works inside a project

Claude Code itself is an agentic developer tool. “Agentic” means it carries out steps on its own instead of only returning answers. It works through a loop of gathering context, taking action and verifying results, according to Handicraft Marketing. In practice the tool first reads what it needs to understand a project, then makes changes, then checks whether those changes worked, and repeats the cycle until the task is done.

Projects can also carry their own standing instructions. CLAUDE.md files hold project instructions that Claude Code reads, which lets a team record conventions, preferred tools or constraints once instead of repeating them in every request. Nothing in the reporting says that such a file can override the game-cloning rule, and nothing suggests that it can. Project instructions and the system prompt are different layers, and the rule described here sits in the system prompt.

For teams, the consequence is predictable: a brief that asks an agent to reproduce an existing game, in whole or in part, is now likely to be declined, however the deliverable is named. Work that involves a client’s request to imitate a known title would need to be scoped differently, or handled without the tool.

What developers and players can still build

The block targets copies, not creativity. Users can still build their own games, genre titles and projects with custom assets, The Decoder reports.

A genre title is a game that follows the conventions of an established type, such as its basic mechanics or structure, without reproducing a specific existing game. That category remains open. So do projects built around a creator’s own artwork, sound and design, which is what “custom assets” refers to: material the developer made or has the right to use.

There is a caveat. Some users report running into problems with original projects too. According to several reports, the refusals reach beyond game work and cover crash-dump analysis, anti-cheat code and security audits for a developer’s own game, as well as reverse-engineering requests. Developers have posted such refusals on the GitHub issue tracker for Claude Code. The reporting does not say how many users are affected or whether Anthropic has responded to those particular complaints.

The reported false positives are consistent with a rule written as two sentences in a prompt. Broad instructions can catch work that was never meant to be covered, and that appears to be the experience described by some users. The facts available do not confirm the mechanism, so this remains an observation about the reports, not an explanation from Anthropic.

Where the line falls for studios and players

For small studios and hobbyists, the practical line runs between originality and reproduction. Projects that borrow a genre’s structure but bring their own assets are described as allowed, while anything resembling a recreation of a known title is the target. Where a borderline project lands is not spelled out. A developer building a platformer with original characters and levels sits clearly on the allowed side of the description; a project that mirrors a named commercial game closely does not.

The split, as The Decoder describes it, looks like this:

Stays allowed	Is blocked
Original games with your own design	A recreation of a specific existing game
Genre titles that follow a genre’s conventions	The same recreation under a new name
Projects built from your own custom assets	A clone presented as an original work

Players are affected indirectly. The unofficial browser ports described in the section on cloning projects below were created by users with AI models, and the new rule removes one of the tools used for that kind of work in Claude Code. It does not affect games that studios or independent developers make themselves.

What Claude Code is and how to get it

Claude Code is an agentic coding tool from Anthropic built for software development. It reads codebases, edits files and runs commands, according to Handicraft Marketing. A codebase is the full body of source code for a project, so the tool works with the project as a whole rather than with a pasted snippet.

It is a command-line interface (CLI) tool that runs in the terminal, where its feature set is the most complete, and it also works in VS Code, in Cursor and JetBrains IDEs, in a desktop app and in the browser. An IDE, short for integrated development environment, is the editor program in which developers write and test code. The range of surfaces means the new rule applies to a tool that many developers reach through whichever environment they already use.

Access requires payment. Claude Code needs a paid Claude plan (Pro, Max, Team or Enterprise) or an Anthropic Console account billed by API usage. The first route is a subscription with a monthly fee. The second charges according to how much the API is used, so the cost follows consumption instead of staying flat.

Paid plans and access routes

Plan	Price	Claude Code access
Pro	US$20 a month, or US$17 a month billed yearly	Included
Max	Starts at US$100 a month	Listed among the paid plans that work with Claude Code
Team	Not given in the available facts	Listed among the paid plans that work with Claude Code
Enterprise	Not given in the available facts	Listed among the paid plans that work with Claude Code
Anthropic Console account	Billed by API usage	Works with Claude Code

The Pro and Max prices were checked on 22 October 2026, per Handicraft Marketing. Prices can change, and the table reflects that check rather than a current price list from Anthropic. The Team and Enterprise figures are not part of the available facts.

For agencies and content teams weighing cost, the difference between the two access routes matters. A flat monthly plan suits predictable use by one person. Usage-based billing through a Console account ties spending to the amount of work the tool does, which is easier to attribute to a specific client project but harder to forecast. The facts do not say which route suits which kind of team, and the choice depends on how heavily the tool is used.

One more detail concerns records. Claude Code sessions are saved locally as plain-text JSONL files under ~/.claude/projects/. JSONL is a format in which each line is a separate structured entry, which keeps session logs readable and easy to process. Because the files sit on the user’s own machine, a team can inspect what the tool did in a given session without exporting anything from a service.

The flood of AI cloning projects behind the change

The change follows a flood of AI-powered cloning projects. The Decoder describes a pattern in which capable models made it quick to rebuild commercial software.

Users have used Opus 5.5 and GPT-6 Astra to create hundreds of unofficial browser ports of well-known titles such as Halo and GTA Vice City. A port is a version of a game rebuilt to run on a different platform, here a web browser. The mention of GPT-6 Astra shows that the activity was not limited to one company’s models. The reporting does not say whether any other AI company has introduced comparable restrictions.

A second strand of activity involves reverse engineering, the practice of studying a finished program to work out how it was built. Tools like Reverse Engineer Anything pair AI agents with reverse engineering to turn native binaries into readable pseudocode. Native binaries are the compiled programs a computer actually runs, which are not meant to be read by people. Pseudocode is a plain-language outline of a program’s logic. Converting one into the other gives a developer, or an AI agent, a map for rebuilding the original, which is the step that links reverse engineering to cloning.

The pattern extends beyond games. One developer recreated parts of Adobe’s Creative Cloud apps, showing that the practice reaches professional creative software as well as entertainment titles.

Hardware has been affected too. Sony’s PlayStation 5 became a target after a jailbreak, a method of bypassing the restrictions built into a device, led to a wave of emulators. An emulator is software that imitates one system on another, letting programs written for the original run elsewhere. The reporting places this among the examples of cloning activity that preceded the change, not as a result of Claude Code.

How the trend connects to Anthropic’s rule

The reporting connects Anthropic’s rule to this broader trend. It does not say that any single project prompted the change, and it does not describe legal action by any rights holder. The link drawn is one of timing and context: a wave of AI-built imitations of existing software, followed by a rule aimed at one category of them.

For businesses that build software with AI agents, the episode shows how tool providers respond to misuse. They add limits at the product level. A restriction can appear in a system prompt, take effect within days and be confirmed only when an employee mentions it on social media. Teams that depend on a specific tool behavior get no formal notice period, based on the facts reported here. Product-level safeguards also extend beyond coding rules, as shown by Anthropic’s watermark for Claude-generated text across several of its current models.

Anthropic says the rules will keep changing

Osmani said Anthropic plans to keep tweaking the rules going forward. The current two-sentence restriction is therefore not presented as final.

What that means for edge cases, including the original projects some users say are being caught, is not known yet. No timeline, scope or details for further changes were given in the reporting. Whether the adjustments will narrow the rule to reduce false positives, widen it to cover more kinds of copying, or do both is not stated.

Users who build games with Claude Code can expect the boundary to move. Anyone planning client work that sits close to the line between a genre title and a clone is working with a rule that its maker has described as unfinished.

Sources

• The Decoder
• praxis-agents.ai
• Handicraft Marketing

Photo by The Decoder