Skip to content

Anthropic Blocked Game Cloning With Two Sentences. A Look Is Harder to Block

Claude Code now refuses to rebuild a specific existing game, and game developers are cancelling public demos because AI clones arrive before launch. For anyone publishing images or video, the lesson is worse: in visual work the demo is the product, and the guardrail is a sentence in a prompt.

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

On this page
A plaster casting mold and two identical faceless plaster busts on a dusty workshop bench under a raking side light
Editorial illustration; not a model output.

Around October 27, Claude Code's system prompt picked up a rule with the internal identifier tengu_heron_brook. It permits original games, games in an established genre, and games built on the user's own designs and assets. It forbids reconstructing a specific existing game, including one that has been renamed or presented as original. A user spotted the change, Anthropic confirmed it, and according to that user there is no toggle to turn it off.

Two sentences in a system prompt, in one vendor's coding tool, against a problem that exists in every generative tool. The interesting part is not whether the fix works. It is why game developers asked for it, because their answer describes what is about to happen to anyone whose work is a look rather than a mechanic.

What developers were actually afraid of

Steam Next Fest is the week small studios get to be seen. You publish a demo, players try it, some add the game to a wishlist, and a fraction of those become sales at launch. It is the cheapest marketing a small team has.

In the most recent edition, about 35% of the Next Fest store pages that could be read carried an AI disclosure, and the bigger problem was not the volume. It was that developers stopped publishing demos at all. Alliance Arts cancelled the demo for Drunken Goddess Reflux and replaced it with a lottery-based closed beta. Digital Sun cancelled the public test for ReVamp, calling a demo an unacceptably high risk. Studio Bitdot's Needle in a Haystack Simulator has collected so many clones that the original now carries an "original" tag on its own store listing.

Some teams are reportedly considering delaying trailers so they do not reveal the mechanic, art direction, or central idea that makes the game distinctive. That is a strange place for an industry to arrive at: promotion is the thing you cannot afford to skip, and promotion is the input the copier needs.

The mechanism behind it is not mysterious. Opus 5.5 is very good at reading code, which means a playable demo is also a readable specification. Anthropic's own staffer Addy Osmani described the change as two sentences added to steer the model away from recreating someone else's specific game, and acknowledged it can misfire and refuse legitimate work, in which case you write to support.

Valve's public position, as reported, is that clones predate AI, that Steam does not want to be the gatekeeper of what counts as derivative, and that developers should not build a game around a single mechanic that is easy to copy. That last sentence is correct and close to useless when you are three months from launch.

Now translate it to images and video

A game demo hides something. You play a slice, you do not get the whole design, and the parts you cannot see are most of the work. Visual work does not have that shelter, because in visual work the demo is the product.

A style frame, a twelve-second clip, a character sheet, a lighting test: that is the thing a client is buying. Publishing it before the deal is how the deal happens. And a published frame is not a hidden mechanic, it is a reference image with a caption, which is precisely the input today's models handle best. The copier does not need to reverse engineer anything. It needs to see the picture and type "like this".

The gap between the two cases is not the model's capability. It is the surface area. Rebuilding a game requires understanding a system. Rebuilding a look requires one good reference and a sentence.

The guardrail has the wrong shape

The Claude rule is a norm encoded as an instruction, and instructions are the softest control in the stack. It applies to one product. The same capability ships in every other coding agent, every image model with a reference input, and every video tool with an image-to-video mode. A request that gets refused can be restated, split into steps, or moved to a different vendor in under a minute.

That is not a criticism of Anthropic specifically. It is what happens when a policy is expressed in natural language inside a system that treats natural language as input. The rule changes what a well-behaved user sees. It does not change what is technically possible, and the people cloning games were never the well-behaved ones.

So the useful question for a visual creator is not "which vendor is protecting me". It is "what do I own that a reference image cannot transmit".

What a model cannot copy from a post

Inputs you control

A consistent cast, a photographed location, your own footage, a library of your own props and sets. A model can imitate the surface of a style. It cannot reproduce your archive of a specific face under your lighting, in your continuity, across forty shots, unless it has that archive.

The system around the output

Consistency across a sequence, a review loop with a named person making the call, revision history, the judgment about which of twelve good frames is the right one. None of that fits in a prompt, and all of it is what clients actually pay for when generation is cheap. That argument shows up in the numbers more clearly than in any policy debate.

Sequencing

Show the result, keep the recipe. A finished piece of work does not tell a competitor how it was assembled; a process breakdown does. The game studios that cancelled their demos made a defensible trade, and the visual equivalent is deciding which parts of the method appear in public before the invoice is paid.

The next project

A copy lands fast, which means the advantage was never the last image. It was the ability to produce the next one on a schedule someone else cannot match. Cloning a performer's likeness followed the same pattern: the technology was ready long before the incentives lined up, and what protected the people who kept working was leverage, not novelty.

The uncomfortable part

Every one of those defenses costs visibility, and visibility is how small studios get hired. A guarded portfolio is a portfolio nobody discovers. Valve can tell developers to build something harder to copy; that advice assumes they already have an audience.

The realistic version is narrower and less satisfying. Assume anything you publish will be used as input, and decide in advance which inputs you are willing to give away: the outcome, the mood, the genre, the format. Keep the cast, the reference library, the continuity, and the pipeline. That is not a technical defense against cloning, because there is not one. It is a business position that survives it.

Sources

Share this guide
X LinkedIn