Digital Sun, an independent studio in Valencia, Spain, had planned to start letting players into a Steam playtest of its next game, ReVamp, on Oct. 29. On Oct. 19 it changed course. The test would run on the studio's own Discord server instead, with access handed out in waves, and the public Steam playtest was pushed to "around January."
"We've been concerned by recent news of games being reverse-engineered, cloned and published before the original is even out," the studio wrote on Steam. "With launch still months away, sharing a build with tens of thousands of players felt like too big a risk for ReVamp."
On Oct. 16, three days earlier, the Japanese company Alliance Arts had scrapped the public demo of Drunken Goddess Reflux that was due on Oct. 19, the opening day of Steam Next Fest, and replaced it with a closed playtest decided by lottery. One of its two stated reasons was the worry that works imitating the game's presentation and gameplay "could be quickly produced using AI or other tools and released ahead of the game's launch." The agreement testers must accept now includes a clause barring them from using the build for AI training or from having any AI analyze its programs or data.
On Oct. 30, the other side of the problem moved. Addy Osmani, a member of Anthropic's technical staff, confirmed on X that the company had added two sentences to Claude Code's prompt on Opus 5.5 "to steer it away from recreating someone else's specific game because using Claude to infringe IP is against our usage policy."
It is a real step from a major AI coding vendor, and a narrow one: a block of text in one client that can be changed remotely. Anthropic says it added the sentences to the Opus 5.5 prompt; XDA found that the copy delivered through tengu_heron_brook applies to every model an account uses in Claude Code, with Opus 5.5 getting the instruction through a separate channel. It leaves the studios' position where it was. The people paying for clone risk are still the original developers, who are giving up the exposure they depend on most and writing their defenses into playtest contracts. Valve, which runs Steam, has said it doesn't want to be the gatekeeper on which tools developers use. With the tool maker offering two sentences and the platform declining to gatekeep, the cost lands on the smallest party in the chain.
Where the two sentences came from
Users found the rule before Anthropic described it. On Oct. 30, an X user named chasm posted a screenshot saying Claude Code had picked up a new instruction around Oct. 27, delivered from Anthropic's servers under a flag called tengu_heron_brook, with no setting to turn it off. The full text: "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." The post has been seen more than 2 million times.
XDA Developers then found the same wording inside the Claude Code v2.1.314 binary. By XDA's reading, tengu_heron_brook is a general mechanism Anthropic has used before to push adjustable system instructions, not a flag built for games, and a separate mechanism tied to an Opus 5.5 prompt bundle can apply the instruction to that model on its own. XDA's tests also suggested the rule is not a ban on reverse engineering: Opus 5.5 agreed to help with a Pokémon Diamond decompilation project and sketched how it would build a static recompilation that runs natively on Windows.
"Original games, genre games, and games built from your own designs and assets are meant to keep working," Osmani wrote. "We'll continue to iterate to ensure we're striking the right balance. In the meantime, if Claude turns down something that's yours, send it my way." The policy Osmani pointed to is a long-standing line in Anthropic's usage policy against using its products to "infringe, misappropriate, or violate the intellectual property rights of a third party."
Who got stopped first
The first refusal to surface publicly came from a small business. On Oct. 30, the owner of Conflict Simulations LLC opened issue #106167 in the Claude Code GitHub repository. They build digital modules of board wargames on paid commission for individual clients, from games those clients own, for their own use, in code written from scratch. They said Claude Code had handled that work for months, on two paid accounts, one Max and one Team, and that Opus 5.5 sessions now refused to continue, citing the new instruction.
The case is contested. The same day the owner corrected the post to say the work is "not publisher-licensed"; clients show proof they own the game. One commenter replied that unless the clients are the rights holders, they didn't see how it differed from any other reverse-engineered port. The owner also wrote that their sessions kept working after the rule arrived and began refusing only after they asked Claude whether the rule applied to their work; another commenter suggested their session memory or Git history had been "poisoned." XDA said it had not reproduced the refusals.
However it is resolved, the case shows the structural limit of a prompt-level rule: the model has to infer from a request whether the user holds the rights, and it has no way to check. The owner said Anthropic's support bot told them on both accounts that the policy "applies to all customers without exception," with no verification or exemption path. One of their three requests was that server-delivered instruction changes of this scope go into the Claude Code changelog, "so paying professional users are not blindsided mid project."
The platform hands it back
Valve laid out its position three days before Osmani's post. According to the Korean outlet Inven, at the Steam Indie Connect developer event in Busan on Oct. 27, a developer said indie studios had been getting extortion emails saying, in effect, pay up or we'll copy your game with AI and release it first, and asked whether Steam had any protections.
Tom Giardino of Valve's Steam business team said it was the first he had heard of such extortion, that it was closer to a matter for law enforcement than for a game platform, and that Valve wanted to look into cases developers reported. On tools, he was firm. "We don't want to be gatekeepers telling developers which tools they should or shouldn't use," he said, in Inven's English translation of the Korean session. Clones predate AI, he argued; AI just speeds them up. The first defense is community and trust, the second is continuous improvement: "Rather than relying on a single mechanism that someone can easily copy, you must continue to grow your game by adding features, personality, and creativity that stem from your own perspective as a developer." He compared it with Valve's long-standing approach to piracy: skip aggressive DRM and make the legitimate product the best version. Inven noted that Valve offered no definitive solution to cloning or to low-quality games.
The argument has force for a game that is already out and has players. ReVamp and Drunken Goddess Reflux are stuck in the gap before launch, when there is no community yet and nothing to update, and a public playtest is one of the main ways an indie game builds that community in the first place. Showing games early, Digital Sun wrote, is how indies "get seen and find the fans who keep us going." It asked platforms to "keep strengthening protections for original work."
What it means for small teams
Put the three moves side by side and the split is plain. Anthropic restricted one product, but Game Developer, relaying users' replies, and XDA both noted that would-be cloners can switch to other AI tools or to open-weight models without such limits, and Game Developer added that safety limits on generative AI have repeatedly been talked around. Valve announced no new policy. The ones who changed how they work were the original studios: smaller tests, contracts that forbid AI analysis, public exposure pushed closer to release day.
For independent developers, the most immediate cost of AI cloning may not be getting copied. It may be the exposure they give up to avoid it. Nobody can put a number on that yet, but both studios' choices point the same way: the stretch of development in which you show the game to players first is getting shorter.
Small teams building on Claude Code get a reminder of their own. This rule arrived from the server and was confirmed only after users found it. An earlier piece on Claude subscriptions returning to OpenCode looked at how Anthropic's client decides which open-source tools can use Claude. Here, what the client is willing to do changed without users updating anything or touching a setting. A business that runs on one coding agent should treat that as a real dependency risk.
Osmani says the rule will keep changing. Until it gets more precise, or the platforms come up with something new, indie studios defending against clones are largely on their own.