Two sentences, delivered silently
The reported instruction is two sentences long. Chasm, the X user who first surfaced it, quoted it as telling Claude Code to build original games, games in a genre, and games from the user's own designs and assets, but not to recreate a specific existing game, even if it is renamed or called original.
Anthropic has confirmed that it added two sentences to Claude Code's prompt on Opus 5.5. It has not published the full wording in an official announcement or changelog. The distinction matters: the company's confirmation establishes the purpose and scope of the change, while the exact text comes from the original report.
Chasm reported that the instruction began appearing in Claude Code around October 27, delivered from Anthropic's servers under a flag named "tengu_heron_brook." Chasm also said there was no setting to disable it. Osmani subsequently confirmed that Anthropic had added two sentences to the prompt used by Claude Code on Opus 5.5, but did not independently confirm the flag name, deployment date or settings behavior.
As of November 2, Anthropic had published no separate product announcement or Claude Code changelog entry documenting the change. A team using Claude Code for game work learns where this boundary sits when a request is refused.
Osmani's stated reason is narrower than the coverage suggests. Anthropic's usage policy already prohibits using Claude to infringe intellectual property, he said, and the two sentences steer the model away from a specific game for that reason. This is not a ban on making games with Claude Code. By Anthropic's own account, the three categories most game work falls into — an original concept, a game in an established genre, a project built from your own designs and assets — are meant to keep working.
The boundary the prompt has to judge
A prompt instruction can state that line cleanly. Applying it is a different problem. Claude Code cannot verify who owns a game, whether a developer holds a license, or whether a commissioned project is authorized. It has to judge from the request in front of it, and a request to continue work on an existing codebase can look, from the model's side, much like a request to recreate the game that codebase implements.
That is not a hypothetical gap. On October 30, a developer representing Conflict Simulations LLC reported on Anthropic's GitHub repository that Claude Code had begun refusing work on commissioned digital board-wargame modules. The developer described the modules as using original code and customer-provided materials, with finished work intended for customers' personal use. The developer also acknowledged that the projects were not officially licensed by the original game publishers.
Those details matter in both directions. The report describes work the developer says was independently created and commissioned, but it does not establish that the projects were legally authorized or resolve the underlying intellectual-property questions. The developer also said refusals continued on the same repositories after the original session was cleared. Anthropic had not publicly resolved the specific complaint in the sources reviewed for this article.
Three related questions remain unanswered in any official explanation: whether the rule will extend beyond Opus 5.5; how it will handle mods, decompilations, recompilations and other work involving existing games; and whether a specific incident prompted the timing. These activities can raise different intellectual-property and licensing questions, but Anthropic has not explained how the instruction distinguishes between them.
gHacks connected the timing to complaints surrounding Steam Next Fest in October, when developers raised concerns about AI-assisted copies of existing games. That is reported industry context, not an explanation Anthropic has given for the change. Osmani's stated reason was preventing the use of Claude to infringe intellectual property.
What changes for a team using Claude Code
For most game projects, on Anthropic's account, nothing should change. The practical change lands on a narrower set of work: rebuilding a known title for study or comparison, porting or restoring an older game, building modules or companions for an existing board or video game, and any long-running project whose repository the model may now read as a recreation. Those teams should expect refusals even after renaming a project, because the rule explicitly covers a renamed recreation, and they have no local setting to adjust. Chasm reported no local setting for disabling the instruction. Osmani's response offered a human contact instead: he invited developers whose own projects were refused to send him the cases so the company could continue adjusting the rule.
That is the part worth watching. A two-sentence prompt rule is cheap to ship and easy to revise, which is a reasonable way to respond quickly to game cloning. It is a weaker way to carry a policy that decides which paying customers' existing projects continue. If the rule stays, the useful next step from Anthropic is not a longer prompt. It is a published statement of the boundary, how an authorized project is recognized, and what a refused developer can appeal to besides a post on X.
Sources
- [1] Addy Osmani (Anthropic technical staff), post on X, October 30, 2026, confirming the two-sentence addition to Claude Code's prompt on Opus 5.5 (primary confirmation, quoted verbatim in gHacks and VideoCardz)Read source
- [2] gHacks, "Anthropic Adds a Claude Code Rule on Opus 5.5 That Blocks Recreations of Existing Games," November 2, 2026 (rule text, allowed/blocked split, Steam Next Fest context reported as context, unanswered questions)Read source
- [3] VideoCardz, "Anthropic restricts Claude Code from recreating existing games, developers report blocked projects," November 1, 2026 (Chasm discovery and "tengu_heron_brook" flag, Conflict Simulations LLC GitHub report of October 30)Read source
- [4] TechTimes, "Anthropic Adds a Rule to Stop Claude From Making AI Copies of Your Favourite Games," November 2026 (corroboration of the October 27 spotting, the October 30 confirmation, and the limits of the rule)Read source