Anthropic confirms Claude Code's rule against recreating existing games; it arrived through a remote prompt slot, tengu_heron_brook

Anthropic has confirmed that Claude Code now tells its model not to rebuild other people's games. "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," Addy Osmani, a member of Anthropic's technical staff, wrote on October 30. An Anthropic spokesperson pointed Game Developer to the same statement.

The two sentences are the ones the X user chasm posted that morning: "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." Chasm said the text reached Claude Code around October 27 under a server-side field called tengu_heron_brook, with no setting to turn it off. The post passed two million views. Osmani added that original games, genre games and games built from your own designs and assets "are meant to keep working," that Anthropic will "continue to iterate," and: "if Claude turns down something that's yours, send it my way."

The IP argument has been covered plenty. The delivery mechanism has had less attention, and it's the more interesting part for anyone who runs Claude Code.

What heron_brook is

tengu_heron_brook isn't a games flag. It's a general slot in Claude Code's system prompt whose text Anthropic can change from its servers, without shipping a new version.

Users have been taking it apart in Anthropic's own issue tracker since May. Issue #62061 reported that version 2.1.150 added a function that reads a string from two network sources, the client_data field of the CLI's bootstrap response and a GrowthBook feature flag named tengu_heron_brook, and inserts it verbatim into the system prompt as a section of its own. The changelog for that release said "Internal infrastructure improvements (no user-facing changes)."

In July the slot was used for something unrelated to games. Issue #82371 quotes a debug payload from an experiment called claude_code_canal_plateau_experiment, in which tengu_heron_brook was set to "Do not call the AgentTool unless the user requested it." The reporter's project had a CLAUDE.md rule requiring a code-reviewer subagent after every change, and Claude Code stopped running it; asked why, the model cited the injected line. Issue #80988 found the same text built into version 2.1.219 as a fallback for Opus 5. That one is still open.

So the game rule is the third known use of a channel that can change how an agent with shell access behaves overnight, while the version number stays put. XDA Developers found the exact game text inside the Claude Code 2.1.314 binary. It also found a second route, an original_games_guidance entry in an opus_5_5_prompt_bundle, which matches Osmani's "on Opus 5.5". By XDA's reading, accounts that get the text through tengu_heron_brook instead have it applied to every model they use in Claude Code. Anthropic hasn't said which accounts get which route.

Where it bites

XDA's tests suggest the rule is narrower than the headlines. Opus 5.5 refused to recreate Pokémon Diamond and cited its session instructions. But it agreed to help with a decompilation of the game and outlined a static recompilation to run it natively on Windows, telling reverse engineering apart from recreation.

Not everyone has been that lucky. In issue #106167, filed on October 30, lerugray of Conflict Simulations LLC says Claude Code stopped doing the company's paid work. It builds personalized digital modules of board wargames for clients who own the physical games and supply the materials. The account of how it started is the interesting part. The sessions kept doing this work after the rule reportedly shipped around October 27. They refused only after lerugray asked Claude whether the new rule applied, and the refusals then persisted in fresh sessions working from the same repositories. One commenter suggested the refusal had been written into the project's memory or Git history and could be scrubbed. lerugray said they would try. Anthropic's support bot told them on two accounts that the policy "applies to all customers without exception." As of November 1 at 01:00 UTC, nobody from Anthropic had replied in the issue.

Nothing is technically wrong with an AI vendor changing its prompts. But a system prompt section that a server can rewrite at any time is a moving part in every evaluation, benchmark and long-running workflow built on Claude Code. Since May the only reliable way to see it has been to read the binary or the debug output. Users asking Anthropic for a changelog entry when that text changes, which #106167 asks for too, are not asking for much.