Claude Code updated rules: Opus 5.5 no longer replicates specific existing games

Wanting Claude Opus 5.5 to help you make a GTA 5 may no longer be so easy...

On October 30, 𝕏 user Chasm posted that Anthropic's AI coding tool Claude Code had a new rule pushed from the server side around October 27: Claude can help users make original games, but no longer helps people recreate a specific existing game, even if the user renames it or claims it is "original." The post quickly surpassed 2 million views.

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

That is, "You can build original games, games of a certain genre, and games based on the user's own designs and assets. Do not build a recreation of a specific existing game, even if it has been renamed or called original."

This text is stored as a string in the Claude Code local configuration file .claude.json, under the field name tengu_heron_brook, alongside several other configuration items likewise named with random word combinations. In the same configuration block, fields such as "model": "claude-opus-5-5" and "entrypoint": "claude-desktop" can also be seen.

Multiple previous teardown analyses of the Claude Code client have mentioned that Tengu is the internal codename for Claude Code, and its feature flags are generally named in the pattern "tengu_ + two random words" to obscure their specific purpose.

"If Claude rejects something that belongs to you, send it to me"

After this 𝕏 post went viral, Anthropic also responded. Claude Code developer Addy Osmani commented under the post: "We added two sentences to the prompt when Claude Code runs Opus 5.5, guiding it not to recreate someone's specific game, because using Claude to infringe intellectual property violates our usage policy. Original games, genre games, and games made with your own designs and assets should all continue to work normally."

He also acknowledged that the boundaries of this rule are still being figured out: "We will keep iterating to make sure we find the right balance. In the meantime, if Claude rejects something that belongs to you, please send it to me."

This response reveals several noteworthy points:

Anthropic's basis is the existing usage policy, not a newly added prohibition; these two sentences merely "translate" the policy into instructions the model can execute.

The rule is explicitly attached to Opus 5.5; Anthropic did not say whether the same prompt also applies when Claude Code calls other models.

A personalized feedback channel like "send it to me" also indirectly shows that false positives are almost inevitable, and the team already expects legitimate requests to be rejected.

Another detail in the timeline is that Opus 5.5 was released on September 22, focused on coding and long-task capabilities, and on the day of release a large number of users posted playable games it generated in one go. Some reports said that within the first few hours of launch, game prototypes imitating Dark Souls, Microsoft Flight Simulator, and Mario Kart appeared. The better the model gets at making games, the harder it is to avoid the question of "whose game is it making."

Of course, like many similar Anthropic policies, this also drew a wave of mockery:

If you look only at the Claude Code side, two sentences of prompt seem insignificant. But placed back into the game industry over the past month, it looks more like the first response to a growing anxiety.

On October 19, Spanish indie studio Digital Sun, known for the Moonlighter series, announced that it was canceling the large-scale public test of its new game ReVamp originally scheduled to open on Steam on October 29, and would instead invite players in batches for small-scale playtests in its official Discord, with a broader public test tentatively set for around January next year.

The studio said in its statement that recently more and more games have been reverse-engineered and cloned and released before the original; with several months still to go before ReVamp's March 2027 release, putting out a large-scale public version at this time is "too risky."

Digital Sun's statement itself did not directly name AI, but almost all reports linked it to AI coding tools. The reason is not hard to understand: in the past, cloning an indie game required a small team to spend weeks or even months; now, someone can get a playtest build, feed it to an AI coding assistant, and possibly produce a "playable" imitation within days.

Around the time Digital Sun spoke out, the two-person Korean development team YNOT also encountered something similar. On October 17, they learned that someone was using AI to clone their Halloween-themed restaurant management game Spookitchen, and on October 20, on the eve of the Steam Next Fest, they took down the public demo. The team admitted on 𝕏 that taking down the demo was only a stopgap measure, not a fundamental solution.

The more extreme case is extortion. According to Korean media reports, an indie developer received an email in which the sender claimed to have already made a clone with AI and could put it on Steam ahead of the original at any time, demanding payment of Steam items worth about $200 within 24 hours or they would publish it.

How many people can two sentences stop?

Anthropic does not want to become a tool for mass game copying. But how much this rule can actually do is currently subject to many doubts.

The most direct issue is enforcement. This is a soft constraint at the prompt level, not an independent content filter, and whether it ultimately refuses depends on the model's understanding of the user's request. GameDeveloper pointed out that researchers and journalists have repeatedly proven that the safety limits of generative AI can be circumvented with various phrasings; this rule will most likely only stop the least effortful plagiarists.

Someone who wants to make a clone can simply not name the game, and instead describe the gameplay, interface, and numerical values to the model item by item.

The second issue is boundaries. There is no clear dividing line between "a game of a certain genre" and "a recreation of a specific existing game." Making a Roguelike is obviously fine; making a game that is "the same as Slay the Spire, but with the card names changed" is obviously problematic, but there is a large gray area between the two. In many jurisdictions, gameplay rules themselves are often difficult to protect under copyright; what is protected is usually the specific expression such as art, code, and text. It is unrealistic in itself for the model to make legal judgments on the fly in a single conversation.

The third issue is false positives. Chasm's concern is not unfounded: fan remakes, ports of old games, game preservation, and mod development—activities that have long existed in these communities—may all overlap with the literal meaning of "recreating a specific game."

In fact, a developer has already used Claude Code to rebuild, based on preserved script files, an online game launched in 1992 whose source code had long been lost. If such a project is the developer's own work, it should be allowed under the rules; but if it is a community taking over an old game that has already shut down, it falls into a gray area.

Osmani's line "if it rejects something that belongs to you, send it to me" is, to some extent, precisely reserving room for remedies in these cases.

The last issue, and the one most discussed under Chasm's post: alternatives are readily available. As many respondents pointed out, if Claude is too troublesome to use, people planning to clone games through vibe coding can simply turn to open-weight models without such restrictions. Constraints from a single vendor can hardly change the capability boundaries of the entire tool ecosystem.

Reference links

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: the one who plays games, published by 36Kr with authorization.