A new rule inside Claude Code is reportedly stopping the tool from helping recreate existing video games. According to an X post and a GitHub issue filed on October 30, 2026, Claude Code began receiving a server-side instruction around October 27 that makes sessions refuse game-related recreation work. Anthropic has not published the wording or commented in anything we could find, so this post separates what is documented from what is inferred.
TL;DR: claims and their status
| Claim | Source | Status |
|---|---|---|
| A server-sent rule appeared around Oct 27 | X post by @chasmmmmmmmmmmm, summarized by AGTP | Reported by users; no Anthropic confirmation |
Flag name is tengu_heron_brook | Same X post | Unverified; we could not inspect it |
| Renamed copies of specific games are blocked; original games are allowed | Same X post | Reported behavior |
| No user setting turns it off | Same X post | Reported |
| It blocks licensed professional game work | GitHub issue #106167 by user lerugray | First-hand report from one developer, with Opus 5.5 sessions on Max and Team accounts |
| Support said there is no exemption path | Same issue | The author's account of support replies |
| Official docs or changelog mention it | Anthropic | Not found |
What the GitHub issue actually says
The issue, titled "Server-delivered 'games' instruction blocks licensed professional development work," comes from a developer who runs a small company building digital modules of board wargames under paid commission from the publishers and clients who own those games. They describe the code as original, written from scratch from supplied materials, and say it does not reproduce copyrighted code, assets or text.
Their report: sessions that did this work for months began refusing it, citing a new instruction against reproducing games. They ask Anthropic to narrow the rule to actual infringement, add a verification or exemption path for professionals comparable to the existing program for vetted security researchers, and document server-delivered instruction changes of this scope in the Claude Code changelog.
A follow-up comment adds a detail worth noting. The author says refusals began only after they asked Claude to check whether the rule affected their work. The refusal then persisted in other sessions working from the same repositories. The author reads that as a false positive. It is also consistent with the model simply applying the instruction once it was surfaced in context, so treat the interpretation as theirs, not established fact.
The author also says an automated support bot stated the policy "applies to all customers without exception" and that tickets were passed to humans with no timeline. These are their descriptions, and we have not seen the support transcripts.
What the X post says
The report that spread the story, by the account @chasmmmmmmmmmmm and re-summarized by AGTP on X, says the rule arrives from Anthropic's servers under the flag tengu_heron_brook, forbids recreating specific existing games even when renamed, still allows original games, genre games and projects built from the user's own assets, and cannot be disabled. Reactions summarized there are split: some welcome it as a copyright guardrail, some say it only stops beginners because a clone can be described as an original design, and some, like the user quoted as saying a coding assistant should not decide which projects you may work on, object on principle.
We could not open the original post or the flag, and a search for the flag name found nothing from Anthropic. The word "tengu" is widely believed to be an internal codename in Claude Code telemetry, but that is unconfirmed, so do not read more into the name.
Why a server-side instruction matters
Claude Code's behavior comes from three layers: the model, the client code you install, and instructions the service can add at runtime. A runtime instruction can change what an agent does without a new release, which is why nothing shows up in a changelog or a version number. That is operationally convenient for a provider responding to a legal or abuse concern, and uncomfortable for professionals who built workflows on the earlier behavior.
This is the same tension we have followed elsewhere. Anthropic's own report on unintended model actions described tightening what Claude Code can do on the network, and a Claude Code agent summary feature raised questions about changes that users find only by inspection. If you want a client-side layer you control, a local policy tool such as Hugging Face SafeAgents shows the other side of the coin, but it cannot override server instructions.
Copyright, clones and the line between them
The legal line the rule gestures at is older than AI. Copyright protects specific expression: code, art, music, text, level layouts. It does not protect game mechanics or genre conventions as such, which is why hundreds of match-three, roguelike and platformer games coexist legally. A clone becomes a problem when it copies protected assets or code, or when it trades on a trademark and trade dress closely enough to confuse buyers. Rights holders have enforced this for decades against unlicensed remakes, with takedown notices aimed at fan projects as well as commercial ones.
A model has no reliable way to check any of that from a prompt. It cannot see whether a user holds a license, whether the assets are original, or whether the code is a clean-room rewrite. That is why category rules such as "do not recreate existing games" are tempting for a provider: they are cheap to deploy and easy to defend as a good-faith effort. They are also why they misfire. The licensed board-game developer in the GitHub issue is doing exactly what a publisher would hire a studio to do, and the rule, as reported, cannot distinguish that from an anonymous user asking for a pixel-for-pixel copy of last month's hit.
There are three design options for a provider facing this. It can keep a blunt rule and accept false positives. It can narrow the rule to the actual harm, copying protected assets or code, and rely on the model to judge requests case by case. Or it can add a verification path, so that licensed professionals declare ownership and receive an exemption, as Anthropic already does for vetted security researchers according to the issue author. The author asks for the second and third. None of this is a prediction of what Anthropic will do; it is the menu its own users are pointing at.
Practical advice if you build games with AI agents
If your project is original, describe it by mechanics, art direction and rules rather than by reference to a named title. This is better engineering anyway: specifications that say what the player does produce more maintainable code than "make it like X". If your project is licensed, keep the license, the asset provenance and the contract in the repository, and state them in your project instructions file so the context is explicit in every session. If you depend on a single vendor's agent for client work, keep a second tool configured, since a server-side change can land on a Friday and block a deadline on Monday.
Finally, keep records. If a refusal blocks paid work, a dated log of the prompt, model, plan and error text gives support and the maintainers something concrete to act on, which is more useful than a screenshot of a single failed message.
Why Anthropic might do this
The company has not said, so any reason is inference. Two plausible pressures are visible in public coverage. Game publishers have long taken action against clones and unlicensed remakes, and AI coding tools make a one-prompt clone of a hit game easy. Our coverage of how studios themselves use these tools, such as Epic Games confirming AI tools write Unreal Engine code, shows the industry is both a heavy user and a rights holder. A blanket rule is blunt, though, and the GitHub issue is a clear example of the cost: legitimate licensed work cannot be told apart from piracy by a category.
Who this affects in practice
- Hobbyist modders and fan remakers: the most likely group to see refusals when they name a specific title.
- Licensed studios and contractors: the issue author's case. They need a documented exemption path.
- Learners: tutorials that recreate Pong, Snake or Pac-Man are common. Whether classic-game tutorials are caught is not documented. Projects like Claude Code Game Studios, which orchestrate agents to build games, are mostly about original work, but we have not tested them against the rule.
- Open-source game engine contributors: work on engines and tools is a different category from recreating a title, and nothing we found says it is blocked.
How to check it yourself
- Start a fresh session and ask for a small original game described by mechanics. Note the result.
- In another fresh session, ask for a recreation of a named commercial title. Note the result.
- Repeat with an explicit statement of license or ownership, to see whether context changes behavior.
- Record Claude Code version, model, plan and date, then add findings to issue #106167 instead of making new reports.
- Keep ownership and license paperwork handy, since the issue author has offered contracts to reviewers.
Remember the session-dependence the author observed: one refusal is not proof of a rule, and one success is not proof of its absence.
What we do not know
- The exact text of the instruction and whether
tengu_heron_brookis the real flag name. - Whether it applies to all models and surfaces or only some, such as Opus 5.5 sessions in the issue.
- Whether Anthropic will add an exemption path, narrow the wording or document it.
- Whether it affects the API, Claude.ai or other products, none of which the sources address.
We will update this post if Anthropic responds.
Related reading
- Anthropic report on unintended model actions
- Claude Code agent summary and token burn: claimed vs verified
- Hugging Face SafeAgents for Claude Code
- Claude Code Game Studios template
- Epic Games and AI-written Unreal Engine code
- Co-op Anything: MCP agent playbook for single-player games
Based on public reports as of October 30, 2026; Anthropic has not confirmed the rule.