Claude Code's Game Recreation Rule: What's Blocked and What Still Works

Claude Code now refuses to build a recreation of a specific existing game, even if the project is renamed or called "original". Original games, genre games and games built from your own designs and assets are still allowed. Anthropic confirmed the change through a staff member on October 30, 2026, and so far it applies to Claude Code running Opus 5.5. Users say there is no setting to turn it off.

If you build games, game tooling or AI "make a game" products, the rule itself is simple. The hard part is the grey zone around it, which has already produced false positives for some developers. This guide covers what changed, what still works, and how to handle a refusal on work you're entitled to do.

What changed and when

Game Developer reported that on October 30, 2026, X user Chasm spotted a new Claude Code instruction, pushed around October 27, under a server-side flag named tengu_heron_brook. Chasm said users can't disable it. The rule has two sentences, quoted the same way by Engadget and others:

• "Do build original games, games in a genre, and games from the user's own designs and assets."
• "Do not build a recreation of a specific existing game, even if it is renamed or called original."

Anthropic didn't publish an announcement. Instead, a spokesperson pointed reporters to a post on X by Addy Osmani, an Anthropic member of technical staff. According to Game Developer, Osmani said Anthropic added two sentences to Claude Code's prompt for Opus 5.5 to "steer it away from recreating someone else's specific game", because using Claude to infringe IP breaks its usage policy. He said original, genre and self-designed games should keep working, that Anthropic will "continue to iterate to ensure we're striking the right balance", and he asked developers whose work gets refused to send it in.

Reports give different dates. tbreak's own article gives both October 27 and October 30. The likely explanation is that the rule shipped around October 27 and was noticed on October 30. Anthropic hasn't said whether other models, the Claude apps or the API are affected. tbreak notes that only Opus 5.5 in Claude Code is confirmed.

Why now

The context is a wave of AI-assisted game clones. Engadget points to vibe-coded browser versions of Halo: Combat Evolved, Call of Duty: Modern Warfare 2 and Grand Theft Auto V. Game Developer reports that Digital Sun Games canceled a public playtest of ReVamp because it feared someone would feed the demo into an AI tool and clone it, and that the developer of Moonlighter delayed a playtest for the same reason. tbreak adds that at Steam Indie Connect in Busan on October 27, Valve acknowledged the clone problem but announced no new rules.

Anthropic has its own legal reasons to care. Engadget notes the company's record $1.5 billion settlement with authors and a recent lawsuit from the Elvis Presley estate.

What's blocked and what still works

The official wording is short, so the table below combines it with what developers have reported. Treat the "grey zone" rows as reported behavior, not policy.

Kotaku's Need for Speed example shows the main ambiguity. The rule permits genre games and blocks recreations of a specific title, but a prompt that names a specific title as the reference can read as a request for a recreation. Neither Anthropic nor any outlet has published criteria for where that line sits.

The false-positive problem

The best-documented complaint so far is GitHub issue #106167, opened on October 30, 2026 by the owner of Conflict Simulations LLC. The company builds personalized digital modules of board wargames for clients. The author says the code is original, written from materials the clients supply, and that the work isn't publisher-licensed but avoids titles that publishers have already digitized. Sessions that had done this work for months started refusing it after the rule shipped.

Two details in that thread matter for everyone:

1. Refusals can persist. The author says the refusals started after they asked Claude whether the new rule affected their work, and then continued across sessions in the same repositories. One commenter suggested the project's memory or Git history had been "poisoned" with context that keeps triggering the rule.
2. There's no exemption path yet. The author says Anthropic's support bot told them the policy "applies to all customers without exception". Their tickets were later escalated to humans, with no timeline. As of the thread's latest updates, no official Anthropic reply had been posted. Other commenters raised abandonware remakes, clean-room decompilation and open-source projects whose licenses allow derivatives.

Kotaku also cites Reddit users who say Claude now resists IP reproduction more broadly, for example refusing to read an unlicensed translation of a web novel. These are anecdotes, not confirmed policy, but they suggest the change may reach beyond games.

What to do if you build games with Claude Code

None of the steps below are about getting a clone past the rule. If your project really is a recreation of someone else's game, the rule is working as intended, and you'd also be breaking Anthropic's Usage Policy, which bans infringing "the intellectual property rights of a third party". These steps are for legitimate work that gets caught by mistake.

1. Describe your design, not someone else's

Write a short design document covering your core loop, mechanics, progression, art direction and the names of your characters and places. Point Claude at it. "Top-down roguelike with deck-based combat and our own enemy set, see docs/design.md" gives the model something original to build. "Make it like [title]" asks it to judge similarity to a specific product, which is the exact judgment the rule exists for.

2. Clean up project context that keeps triggering refusals

Since refusals can carry over between sessions, check what Claude reads automatically:

1. Your CLAUDE.md and any memory files for wording that describes the project as a remake, port or clone of a named title.
2. Your README, design notes and recent commit messages for the same.
3. Old transcripts or notes where you asked Claude about the rule. Those can keep the topic in context.

Edit these to describe the project accurately. If a named title isn't actually part of your project, remove it.

3. Document your rights when they're real

If you own the IP, have a license, or are working from material your client owns, put that in the repository: a LICENSE file, a note on who holds the rights, and where the assets came from. It's no guarantee, since Game Developer points out it's unclear how the software could verify ownership. But an accurate record helps both the model and any human reviewing an appeal.

4. Report false positives through official channels

Osmani asked developers whose original work gets refused to send it in. Use /feedback inside Claude Code, open a support ticket, and if it's a repeatable bug, add a reproducible case to the Claude Code GitHub issues rather than piling onto a thread with complaints.

5. Keep a fallback for blocked days

If game work is central to your business, test now whether a second coding assistant can handle your workflow, so a policy change doesn't stop your sprint. Our AI coding assistants directory is one place to compare options. As Game Developer notes, open-weight models don't carry this restriction. The legal risk of cloning is the same whatever model writes the code.

What it means if you build AI products on Claude

The rule is confirmed for Claude Code's prompt only, not the API. If you run a "describe a game, get a playable build" product on Claude, you're still bound by the Usage Policy's IP clause, and the current policy page lists an effective date of November 12, 2026. A few practical consequences follow.

• Write your own content policy. Don't rely on the model to catch everything. If users can ask for "Tetris but called Blockfall", decide what your product does with that request and say so in your terms.
• Expect server-side prompt changes. This rule reached users through a server-delivered flag, without a changelog entry. The GitHub issue asks Anthropic to document changes this big. Until it does, keep a small regression suite of representative prompts for your core workflows and run it regularly, so you notice behavior changes before your users do.
• Plan for outside validation of the clone problem. Studios cancel playtests and platforms take questions about clones. Tools that help developers prove what's original, such as provenance records, asset tracking and design histories, may become more useful to buyers.

What to watch next

• Anthropic's next adjustment. Osmani said the company will keep iterating. Watch for narrower wording, published criteria, or an exemption path for verified rightsholders. The GitHub issue compares the last one to the existing program for security researchers.
• Scope creep or rollback. Confirmation either way on whether the API, other Claude models, or reverse-engineering tasks fall under the same rule.
• Platform moves. tbreak reports that Valve says it won't police which tools developers use, but asks developers who are threatened with a clone unless they pay to report it.
• Other AI coding tools. Whether other vendors add similar rules. If they do, "build in a genre, not a title" becomes the norm across the industry.

What to do now: if Claude Code refused original work this week, do steps 1 to 4 above in order, then retest. If you ship a game-generation product on Claude, write down your own policy on clones before users test it for you. For more breakdowns of AI tooling changes like this one, see the LaunchBoosts blog.