Skip to content

Renaming the Clone Does Not Make It Original

Anthropic added two sentences to Claude Code telling it to refuse rebuilding someone else's existing game, even under a new name. The rule is narrower than the outrage suggests, and the test behind it is the same one image prompts hit every day.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A game controller beside an open sketchbook with a redrawn character pose, one page loose on a dark desk under cool window light

Somebody at Anthropic added two sentences to Claude Code's instructions, and a corner of the internet read it as the end of an era. The change is smaller than that, and the reasoning behind it is more interesting than the rule.

What actually changed

On 30 October 2026, Addy Osmani, who works on Anthropic's technical staff, confirmed that the tool now carries two extra lines when it runs on Opus 5.5. A user named Chasm had spotted them first. The instruction tells Claude Code to refuse recreating an existing game, even when the user renames it or presents the project as original.

The scope was spelled out in the same reply. Claude should still help with original creations, with games in the same genre, and with projects built from your own assets. The instruction is not a general ban on mods or on browser ports of old titles. Numerama also notes that some users are already reporting refusals on projects they consider legitimate, which is the usual cost of a rule written as a sentence rather than a specification.

Read past the headline and the rule is not "do not make games with AI". It is "do not rebuild this specific game and call it something else."

Why the name is the trigger

An assistant does not see your intent. It sees a description. When someone asks for "a battle royale with the map and characters from a specific 2017 shooter", the only reliable signal of whose work is being copied is the proper noun. Strip the noun and you strip the signal. That is precisely why the instruction says a rename does not count: it is aimed at the workaround, not at the word.

This is also why false positives happen. A brief that describes a mechanic in detail, in the vocabulary of the game that popularised it, looks like a clone request from the inside. The tool has no way to check whether you own the assets, so it errs toward refusing.

Copyright protects expression, not ideas. In games that means characters, art, level layouts, music, dialogue and interface design. The title is a different kind of claim: under US law names and short titles generally are not copyrightable, but they can be trademarked. You can build a battle royale. You cannot build that battle royale's island with that character.

The line is not always clean, and I am not a lawyer. But the shape of it is stable enough to plan around: a mechanic is a design pattern, a specific game is a fixed set of protected choices. Anthropic's two sentences are a crude approximation of that distinction, enforced by a language model reading a prompt. Crude is the honest description. It is also the direction the whole field is moving, because generation is getting cheap and the number of clone requests per day is not.

The same test in an image prompt

If you generate images or video, this is already your problem. Two prompts look similar on the surface:

  • "the character from that anime, standing on a rooftop"
  • "a tired boxer in a late-1990s cel-shaded TV anime look, rooftop at dusk, wind"

The first names a specific work and asks the model to reproduce it. The second names a period, a medium and a mood, and leaves the design to you. The first is the kind of request that platforms increasingly refuse or quietly degrade. The second is the kind of request that usually works, and gives you an image you can iterate on instead of a copy of something that already exists. That is lower-risk prompting, not clearance. Nothing about a prompt settles whether the output is yours to sell.

Style itself is a contested category. Artists have argued for years that a distinctive style can be protected, courts have mostly been skeptical, and the cases are still moving. Franchise characters, logos and living people's likeness are not contested in the same way. Those are the hard lines, and they are hard regardless of which tool you use.

There is a provenance angle too, and it is getting more practical than the legal one. If a client asks where a reference came from, "I described a character from a game" is a bad answer. The royalty model now running for manga characters exists because rights holders want that chain recorded, and the French bill that would presume training use shows where the evidence burden is heading. A prompt log that names your own references is cheap insurance.

How to write a brief that survives the filter

Four habits, all of them useful even when no filter is involved:

Describe the mechanic, not the model. Say what the player does and what the loop feels like. "You hold a shrinking perimeter while scavenging" is a design brief. "It is game, but you are a dog" is a clone request with a costume.

Bring your own assets. If you have a character sheet, a palette, or a location photo, put it in the prompt. The rule explicitly allows work built from your own resources, and it also produces better output, because you are supplying the specificity instead of borrowing it.

Keep the proper nouns out of the prompt and in your notes. The reference belongs in your documentation, not in the instruction. That is not laundering; you are separating your source material from the generation request, which is also how you avoid accidentally training your own pipeline on someone else's design.

Do not count on a rename. If the tool complies and the thing still looks like the original, you have the exposure without the excuse. A new name changes nothing about what is on screen.

What this is a preview of

Two sentences in a system prompt are not law, not a policy page, and not stable. Model behavior will vary, versions will change, and the same request may pass on one tool and fail on another. A refusal is not a legal ruling, and a compliant output is not a license.

What is stable is the question the tool is asking: what are you actually making? For ten years the answer could be a vibe and a reference. As generation gets cheaper, the tools, the platforms and eventually the courts all converge on the same follow-up. If your answer is a proper noun, you are renting someone else's work. If it is a set of decisions you can explain, you have something you can defend as your own. Whether it is also clear to sell is a separate question, and one only a licence or a lawyer can answer.

Sources

Share this guide
X LinkedIn