---
type: "post"
id: "claude-code-receives-server-side-tengu-heron-brook-4bce44"
slug: "claude-code-receives-server-side-tengu-heron-brook-4bce44"
title: "Claude Code receives server-side 'tengu_heron_brook' instruction against recreating existing games; Anthropic confirms"
description: "Anthropic appears to be delivering a Claude Code instruction that tells the agent not to recreate a specific existing game, including one renamed as original."
retrieval_nugget: "Anthropic appears to be delivering a Claude Code instruction that tells the agent not to recreate a specific existing game, including one renamed as original. The narrow subject is games, but the operational point for AI builders is broader: agent behavior can be changed by remotely supplied instructions rather than a visible client release."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: ["claude","claude-code","anthropic"]
entities: ["Anthropic","Claude Code","Claude Opus 5.5","Claude Opus 5"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "same_event"
what_changed: "Same event as Anthropic ужесточает ограничения на воссоздание видеоигр."
prior_refs: ["memory:41d1f419-2ebf-8fd8-aaa8-a648a8c06a53"]
source_url: "https://www.xda-developers.com/anthropic-testing-potential-crackdown-on-vibe-coded-ports-of-existing-games"
source_urls: ["https://www.xda-developers.com/anthropic-testing-potential-crackdown-on-vibe-coded-ports-of-existing-games"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:claude-code-receives-server-side-tengu-heron-brook-4bce44"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/claude-code-receives-server-side-tengu-heron-brook-4bce44/","markdown":"https://newruntime.com/posts/claude-code-receives-server-side-tengu-heron-brook-4bce44.md","json":"https://newruntime.com/posts/claude-code-receives-server-side-tengu-heron-brook-4bce44.json"}
---

# Claude Code receives server-side 'tengu_heron_brook' instruction against recreating existing games; Anthropic confirms

## Retrieval answer

Anthropic appears to be delivering a Claude Code instruction that tells the agent not to recreate a specific existing game, including one renamed as original. The narrow subject is games, but the operational point for AI builders is broader: agent behavior can be changed by remotely supplied instructions rather than a visible client release.

## Summary

Anthropic appears to be delivering a Claude Code instruction that tells the agent not to recreate a specific existing game, including one renamed as original. The narrow subject is games, but the operational point for AI builders is broader: agent behavior can be changed by remotely supplied instructions rather than a visible client release.

## What happened

XDA reported finding a new instruction in the Claude Code v2.1.314 binary. The instruction says Claude should build original games, genre games, and games from a user's own designs and assets, but should not build a recreation of a specific existing game, even if it is renamed or presented as original. The report describes the instruction as arriving through a configuration field called `tengu_heron_brook`.

That field is important because it is described not as a game-specific switch but as a mechanism Anthropic has used to deliver adjustable system instructions remotely. XDA found evidence of an earlier, unrelated instruction concerning reasoning effort through the same mechanism. In other words, the presence of the text in a local Claude Code executable is not itself proof that every account or every model receives it; the delivery mechanism supports testing and selective rollout as well as broader deployment.

The report found another path, involving `original_games_guidance` and an `opus_5_5_prompt_bundle`. That suggests Opus 5.5 can receive the guidance separately, while accounts receiving it through `tengu_heron_brook` could have it applied to all models used in Claude Code. XDA did not find the specific instruction in its test account's `~/.claude.json`, reinforcing the distinction between a locally visible binary string and the configuration actually active for a session.

XDA compared responses from Opus 5.5 and Opus 5 to a request to recreate Pokemon Diamond. Both initially refused. Opus 5.5 explicitly said its session instructions prohibited recreating specific existing games, including renamed versions, closely matching the recovered wording. Opus 5 later said it could assist with legitimate decompilation work after being asked whether a system instruction caused the refusal. The article cautions that a model's account of its own instructions is not reliable evidence, but treats the differing answers as consistent with separate activation for Opus 5.5.

The reported boundary is not clearly a blanket ban on reverse engineering or ports. In XDA's tests, Opus 5.5 agreed to help with a decompilation project and distinguished reverse engineering from recreating a game. It also outlined how a static recompilation could be approached. The article nevertheless records uncertainty over where the practical line falls, especially for developers whose work combines compatibility tooling, customer-supplied materials, decompilation, or game-adjacent assets.

## What the archive adds

This is a new fact on a known Claude Code thread rather than a continuation of the earlier product change. [A prior Field Note](https://newruntime.com/posts/the-claude-code-plugin-is-open-source-as-f04b4e/) covered an open-source Claude Code multiplayer plugin presented as a reference for other harnesses. The new report concerns a different layer: remotely delivered behavioral guidance inside the coding-agent product.

## Trend signal

The report supports [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). The same Claude Code surface may route behavior through account-level configuration or an Opus 5.5-specific prompt bundle. For a production workflow, that means the surrounding harness, configuration path, and evaluation practice can matter as much as which model name appears in the interface.

## Why it matters

A team can reasonably expect a CLI version to remain stable until it upgrades. Remote policy or prompt delivery breaks that assumption: the same installed client and nominal model can begin refusing a class of requests between runs. This report is not evidence that ordinary software tasks will be affected, nor that all users receive this rule. It is evidence that a coding-agent dependency includes behavior delivered outside the package manager.

That changes operational practice. Teams using agents for workflows with contractual, archival, reverse-engineering, or other policy-sensitive inputs should distinguish an agent refusal from a technical impossibility. A refusal might be a dynamic instruction, an account-level rollout, or a model-specific bundle. Without a small reproducible test suite, a team may discover the change only after a customer workflow or internal automation fails.

## Limits and uncertainty

The supplied reporting does not establish how many Claude Code users receive the instruction, whether its rollout is permanent, or whether it applies to every model and account. It also does not establish a complete definition of “recreation” or predict how consistently the agent will classify decompilation, static recompilation, compatibility layers, or customer-provided game materials. Model statements about their own system instructions are explicitly not definitive. The report's tests show responses to selected prompts, not a comprehensive policy evaluation or a guarantee of future behavior.

## Skill to build now

Create a compact regression pack for the coding agent in each important workflow. Include representative permitted tasks, edge cases likely to trigger policy boundaries, and expected artifacts rather than only expected prose. Run the pack when behavior changes unexpectedly, record the client version, selected model, account context, prompt, and result, and retain a fallback path. The supplied material also identifies documented environment variables for disabling some nonessential network traffic; test their effects in an isolated environment before relying on them.

## Reader takeaway

The game rule is narrow, but its delivery path is the durable signal. Treat a coding agent as a remotely operated service, even when its CLI is installed locally: test the workflows you depend on, and investigate unexpected refusals as configuration changes as well as model behavior.

## Open question

Will Anthropic document which server-delivered instructions are active for a Claude Code session, and provide a supported way for teams to evaluate their impact before a production workflow is interrupted?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Claude Code](https://newruntime.com/topics/claude-code.md): Explore the claude-code topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)