Claude Code got a server-side rule on games: no recreating specific existing titles, original and genre work still fine

Claude Code, Anthropic’s command-line coding agent, began receiving a server-side instruction around October 27, 2026 that forbids building a recreation of a specific existing game — even renamed or relabeled as original — while leaving original games, genre games, and builds from the user’s own designs and assets allowed. The exact wording is now checkable: a user pasted it verbatim from their session’s system prompt into GitHub issue #106591. This is not a client-side setting. The owner of Conflict Simulations LLC, whose paid wargame commissions were refused, wrote in issue #106167 on October 30 that months of routine client work suddenly stopped, citing a “server-side instruction” as the cause.

Key points

• The instruction, verified verbatim in issue #106591: “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.”
• It arrived around October 27 through the session system prompt, with no user-side switch. In #106167, the reporter’s sessions kept working for days — refusals began only after public discussion appeared and he asked the model whether the rule affected his work.
• The blast radius exceeds piracy: paid wargame commissions, tabletop-game patches, common-mechanic re-creations, and a half-finished original game all drew refusals; the three earlier blocks describe their mechanisms differently (#105214 “tagged as cyber,” #104913 a classifier block, #102506 just “getting flagged”), unverified as the same rule.
• Chinese outlets report Anthropic confirmed on X that “we added two sentences” to the system prompt (details via Jiqizhixin; QbitAI says only that Anthropic confirmed) — but as of November 1, neither GitHub issue has any Anthropic member reply.

Background

The rule landed right after the model generation that made one-prompt game clones trivial. Anthropic released Claude Opus 5.5 on September 22, and within hours prototypes imitating well-known commercial games appeared. Per QbitAI, one user produced a playable Pokémon clone in under two hours with a free Claude model and a single prompt (“as close to the original as possible”); cross-IP mods (Mario inside Elden Ring) and AI-assisted console hacks running old titles spread the same way — a wave we covered when a batch of those projects put old games in the browser.

Game studios had been pushing back for weeks. Per Jiqizhixin, on October 19 Spain’s Digital Sun pulled the large public playtest of ReVamp over AI-clone fears, and on October 20 Korea’s YNOT withdrew the Spookitchen demo. One indie developer described extortion: pay up, or an AI clone of your demo ships before your game does. Coding agents had just been welcomed as serious tools — Linus Torvalds had posted a Claude Code-assisted kernel patch series to the mailing list just over two weeks earlier — so the rule tightened exactly as professional reliance was climbing.

The facts

The checkable timeline, with sources:

Date	Event	Source
Sep 22, 2026	Opus 5.5 launches; clone prototypes appear within hours	Jiqizhixin
Oct 19, 2026	Digital Sun pulls the ReVamp playtest over AI-clone risk	Jiqizhixin
From Oct 28, 2026	Game-work refusals pile up after the rule lands	#105589, #106167
~Oct 27, 2026	Rule delivered server-side into session system prompts	#106167, Jiqizhixin
Oct 30, 2026	X user Chasm publicizes it; post passes 2M views; #106167 filed	Jiqizhixin
Oct 30–31, 2026	Tom’s Hardware reports, “no official confirmation”; HN threads at 125/72/176 points	Tom’s Hardware, HN
Nov 1, 2026	#106591 publishes the prompt text; Chinese outlets report X acknowledgment	#106591, Jiqizhixin

The instruction is two sentences, quoted verbatim from the session system prompt in #106591: “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.” It enters the prompt server-side, so deleting local config does not remove it. Jiqizhixin ties the rule to a tengu_heron_brook field in the local .claude.json, but earlier issues on the repo (#87635 and others) show that name belonged to a separate agent tool-dispatch instruction shipped back in August — the attribution is unverified, so we leave it flagged rather than stated.

The false-positive list, assembled from issue bodies and Tom’s Hardware’s roundup: paid digitization of clients’ board wargames (#106167 — clients prove ownership, code written from scratch, personal use only) and re-creating Screeps combat mechanics (#105589) both postdate the rule; the three earlier ones each describe the block differently — #104913 (a tabletop patch) reports “a classifier block,” #105214 (anti-cheat in the user’s own game) says it was “tagged as cyber,” and #102506 (a security review of the user’s own game) only says it kept “getting flagged” — and whether any of them shares a mechanism with the Oct 27 games rule is unverified. The strangest detail sits in #106167’s update: for days after the rule shipped, the reporter’s sessions kept doing the same client work. Refusals started only after he read the public discussion and asked the model directly whether the rule applied to his project.

What others say

Anthropic’s response so far lives on X, not in its issue tracker. Jiqizhixin writes that a Claude Code engineer replied under Chasm’s post, confirming “we added two sentences” to the system prompt when Opus 5.5 runs in Claude Code, citing the usage policy on IP infringement, and asking wrongly-blocked users to send him their cases; QbitAI’s version stops at “Anthropic officially confirmed the two sentences were added.” The X post sits behind a login wall, so we could not verify that reply directly. What we did verify: as of November 1, neither #106167 nor #106591 shows any Anthropic member response, and Tom’s Hardware wrote “no official confirmation” when it published on October 31.

The community pushback runs on two tracks. The first is the false positives themselves: #106167 drew eight comments, some proposing exemption paths for open-source derivatives, clean-team workflows, and verifiable ownership, others arguing the reporter’s own digitization of other people’s games was over the line in the first place — a dispute with no resolution yet. The second is whether the rule works at all: on the 176-point HN thread (110+ comments), a recurring argument is that a prompt-level constraint cannot stop users who switch models; Jiqizhixin’s analysis adds that splitting tasks into context-free technical chunks evades it too. The irony argument — a company trained on copyrighted corpora now policing what users rebuild — appears there as well; Anthropic’s own copyright suit ended in a $1.5 billion settlement agreed in 2025. QbitAI’s framing: gaming’s “Napster moment.”

Our take

If you build games, demos, or reverse-engineering exercises in Claude Code, your workflow just changed: the clone-to-learn path — rebuilding a known game to understand it — is closed under default settings, and no amount of prompt craft gets around it, because the constraint sits in the system prompt delivered by the server, above anything your CLAUDE.md can say. The community’s workarounds so far: describe tasks as pure technical problems with no game context, switch models, or take the model’s own suggestion in #106591 and move to genre-level original work. The first two amount to changing tools; treat that as a signal about where the rule’s costs land.

Nobody can actually state the rule’s boundary. What separates “games in a genre” from “a recreation of a specific existing game” has no definition Anthropic has published, so the model guesses, and guesses wrongly — issue #106167 exists for that reason alone. The mechanism deserves as much attention as the rule: a change that kills paid client projects shipped silently, skipped the changelog, and sat unanswered in the issue tracker for days. That pattern is the real precedent for anyone running AI coding agents in production. If Anthropic follows up with an exemption path or changelog entries — the three requests in #106167 — that follow-up will matter more than the rule itself.