Anthropic Tells Claude Code to Stop Recreating Existing Games as Indies Pull Playtests

TL;DR

Around October 27, Claude Code sessions on Claude Opus 5.5 started carrying two new sentences: build original games, genre games and games from the user's own designs and assets, but do not build a recreation of a specific existing game, even if it is renamed or called original. An X user named Chasm found the rule under the internal flag tengu_heron_brook and reported no way to switch it off. Anthropic's Addy Osmani confirmed the change on October 30 and tied it to the usage policy's ban on IP infringement. It lands after a month in which indie studios pulled Steam demos and playtests over AI clones, Valve told developers in Busan it will not police tools, and the Claude Code changelog said nothing. A licensed wargame developer already has an open issue saying the rule shut down months of paid, owner-authorized work.

The two sentences, and where they live

The full text, as Chasm posted it and as it sits in the binary:

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.

I checked a Claude Code 2.1.316 build dated October 31 on a Linux box. The sentence is there as a plain string, next to the flag name tengu_heron_brook and a telemetry event called tengu_brook_heron_applied that records the length of the applied text and the session's effort level. By my reading of the minified code around it, the text a session gets comes from a config lookup keyed by model and effort, which fits Chasm's description of a rule "sent from Anthropic's servers" and Osmani's "on Opus 5.5." A dev.to writeup says XDA found the same string in 2.1.314.

The rule is a sentence in the system prompt. There is no classifier in front of the model. Think of it as a note taped to the barista's register rather than a lock on the espresso machine: it changes what the staff will agree to make, and it does nothing to the machine. Whether a given request counts as "a recreation of a specific existing game" is the model's call, in context, every time.

The changelog entries for 2.1.314 and 2.1.316 cover a fix for PowerShell hooks, a deprecation of fast mode on Opus 4.8 and dozens of other items. Neither mentions games. The flag name, for what it is worth, tells you only that whoever names Anthropic's flags has a birding habit.

Why now: a month of indies going dark

The pressure built on Steam. Kotaku credited Opus 5.5 with "a recent wave of vibe-coded decompilations and recompilations," including browser ports of Halo, The Simpsons: Hit & Run and GTA: Vice City, and reported on October 24 that users believed the model's reverse-engineering had been throttled since October 21. Anthropic never confirmed that earlier change; it has confirmed this one.

On October 16, Alliance Arts cancelled the Steam Next Fest demo of Drunken Goddess Reflux, three days before the fest opened. On October 19, Digital Sun, the studio behind Moonlighter, told the people who had signed up for the first ReVamp playtest on Steam that it was moving to a smaller Discord test starting October 29, with the public playtest pushed to "around January." The studio's statement, quoted by VGC: "We've been concerned by recent news of games being reverse-engineered, cloned and published before the original is even out. With launch still months away, a large public build of ReVamp felt like too big a risk." No clone of ReVamp exists, and the studio says it moved on the risk alone.

Aftermath counted more on October 21: Holy Wow, the Trombone Champ studio, holding back screenshots and video of its next game; Irish studio Bearwaves planning a shadow drop; Iskotaa announcing Project Baltic Sea on its own site with moodboards instead of footage. Bearwaves' Joel Auterson: "Not being able to develop our games in the open has been a real blow, and it's very dispiriting."

Valve answered on October 27 at Steam Indie Connect in Busan, co-hosted with the Busan Indie Connect Festival committee and KOCCA. A developer told Valve's Tom Giardino and Marc Wilkinson that indies were receiving emails reading, "If you don't pay, we'll copy your game with AI and release it first." Per InvenGlobal and Game Developer, Giardino said he was hearing of the extortion for the first time, that clones predate AI and AI "simply accelerated the pace," and that "we don't want to be gatekeepers telling developers which tools they should or shouldn't use." His advice was community, constant updates, and reporting extortion to Valve and to law enforcement. The same week, the tool vendor added the rule the platform would not.

What Anthropic says it blocks

Osmani's post, replying to Chasm on October 30: "We 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. Original games, genre games, and games built from your own designs and assets are meant to keep working. 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." Game Developer asked Anthropic for a statement and was pointed back at that post. The usage policy itself has not changed.

Scope, as confirmed: Claude Code, on Opus 5.5. Nobody at Anthropic has said whether the same sentences reach other models in Claude Code, the Claude apps, or the API. Chasm's "no setting to turn it off" holds for the user-facing config, and nothing in the code path I read checks a local setting.

The first false positive already has an issue number

Issue #106167, filed October 30 by the owner of Conflict Simulations LLC, describes a business that builds digital modules of board wargames on paid commission, from games the clients own and from materials the clients supply, with ownership checked first. Sessions that had done that work for months began refusing it, citing the new instruction. The filer asks for three things: narrow the rule to actual infringement, add a verification path like the one Anthropic runs for vetted security researchers, and document server-delivered prompt changes in the changelog. The issue carries the bug and area:model labels and eight thumbs-up.

The work kept running after October 27, according to the filer's follow-up comment. The refusals started only when the filer, having read about the rule, asked Claude to check whether it applied to the project. From that point the refusal persisted across fresh sessions on the same repositories, and one commenter suggests memory or git history carried the self-audit forward.

Other commenters asked for exemptions: permissively licensed open-source projects, clean-room setups that separate the person reading a decompile from the person writing code, rightsholder-authorized remakes, and private rebuilds of abandonware. One replied that digitizing other people's IP for clients is a reverse-engineering port by another name. A one-sentence rule has no way to tell those cases apart.

Why two sentences will not end cloning

The rule lives in a prompt, so it falls to anyone willing to reword, and the clones that worried Digital Sun were never going to arrive as "please recreate ReVamp." The Hacker News thread on Chasm's post, 124 points and 127 comments, spent much of its energy on the obvious exit: open-weight models carry no such sentence. And the ports Kotaku wrote about were decompilations, while this rule names recreations of games, not decompilation of binaries. The October 21 reports of a reverse-engineering throttle remain unconfirmed.

Valve will not gatekeep tools, Next Fest's random exposure will keep surfacing low-effort entries (Wilkinson: "we recognize that as a trade-off, some low-quality games get exposed"), and the advice to indies is to build community and ship updates faster than a cloner can copy. Digital Sun's "too big a risk" arithmetic did not change because Anthropic added a sentence to one product on one model.

If you build games with Claude Code

• Your own designs and assets are on the allowed side by name. The rule keys on provenance, so make it visible: a short note in the repo on who owns the design and the assets gives the model something to reason from when it decides what "the user's own" means.
• Do not ask the model to audit your project against the rule mid-stream. The one documented false positive began exactly there.
• If Claude refuses something that is yours, Osmani asked for it directly. Attach the refusal and the ownership facts; a GitHub issue on the Claude Code repo is the public route.
• If you are a studio deciding whether to ship a demo, your threat model did not change this week. The rule narrows one tool on one model; the open-weight path and every other coding agent are untouched.

Key Takeaways

• Claude Code on Opus 5.5 now carries a two-sentence instruction against recreating a specific existing game, pushed around October 27 and confirmed by Anthropic's Addy Osmani on October 30.
• The text sits in the Claude Code binary under the flag tengu_heron_brook, is applied from a config keyed by model and effort, and has no user toggle. The changelog does not mention it.
• Original games, genre games and games from your own designs and assets are meant to keep working; the model decides which side a request falls on.
• Issue #106167 reports the rule refusing paid, owner-authorized board wargame work, with refusals starting after the user asked Claude whether the rule applied.
• The trigger was a month of indie studios pulling Steam demos and playtests over AI clones, including Digital Sun's ReVamp, while Valve said in Busan it will not police tools.
• Cloners can reword, decompile, or switch to open-weight models, so the rule changes what one tool will say yes to and leaves the rest of the market untouched.

Sources: Chasm on X: the rule and flag name, Addy Osmani on X: confirmation, anthropics/claude-code issue #106167, Claude Code changelog, Game Developer: Anthropic orders Claude to stop remaking video games, Dexerto, dev.to: server-side rule writeup, PC Gamer, Digital Sun on Bluesky, VGC: ReVamp playtest, Game Developer: Moonlighter developer delays playtest, Aftermath: demos and AI clones, InvenGlobal: Valve at Steam Indie Connect, Game Developer: Valve on Next Fest cloning, Kotaku: Opus 5.5 reverse-engineering reports, Hacker News discussion