On October 30, Anthropic software engineer Addy Osmani confirmed what dataminers had spotted three days earlier: Claude Code's system prompt on Opus 5.5 now contains two sentences telling the model not to rebuild someone else's game. The confirmation, posted on X, came with the quiet admission that does the real talking: "We'll continue to iterate to ensure we're striking the right balance." In prompt engineering, "iterate" is what you call a rule whose boundaries you cannot yet specify.
This is not a small feature request. It is Anthropic trying to encode half of copyright doctrine into two sentences of natural language, deploy it as a system prompt, and hope the classifier in between does not misfire in either direction. It is misfiring in both. Developers are reporting on Anthropic's GitHub that Claude Code now refuses work on their own projects, and, per Ars Technica, some generic tasks like analyzing a crash dump.
The Rule, Timeline, and What Was Actually Shipped
The change was first flagged by an X user going by "chasm," who noticed a new entry in Claude Code's rules file on or around October 27. XDA later verified the rule ships in Claude Code binary version 2.1.314, released October 30, and noted that the internal rule name, tengu_heron_brook, was reused from earlier versions that used it for other purposes. Ars Technica also reports that Opus 5.5 carries its own distinct model-level instructions restricting game recreations, separate from the Claude Code client prompt. Two enforcement points, one in the product, one in the model.
The rule text, as quoted by PC Gamer and Ars Technica:
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.
| Date | Event | Source |
|---|---|---|
| ~Oct 27 | New tengu_heron_brook rule spotted in Claude Code prompt |
chasm on X, via Ars Technica |
| Oct 30 | Rule verified in Claude Code binary v2.1.314 | XDA, via Ars Technica |
| Oct 30 | Osmani confirms "two sentences," cites usage policy | Addy Osmani on X |
| Oct 31 | False-positive reports accumulate on Anthropic's GitHub | Ars Technica, 80 Level |
| Nov 2 | Coverage and analysis of expected ineffectiveness | Ars Technica, 80 Level |
The backdrop is what we covered when Microsoft announced Xbox Labs: a wave of AI-assisted decompilations and ports, AI-built clones flooding Steam Next Fest (the showcase's demo count doubled over the previous event, per PC Gamer), the decompilation community site decomp.dev shutting down because, in maintainer Luke Street's words, newcomers "race to port games, especially utilizing decompilations that they did not contribute to," and indie studios canceling public demos for fear of instant clones. Anthropic's own model, Opus 5.5, was the tool most credited with enabling the wave, which is why this prompt change reads as an apology delivered in JSON.
What This Means: The Enforcement Layer Is the Story
Strip away the copyright politics and this is a systems design decision, and the layer Anthropic chose is the most permissive one available. There is a spectrum of ways to stop a model from doing something, and they differ enormously in cost, coverage, and failure mode.
| Enforcement layer | Example | Coverage | Cost | Bypass cost |
|---|---|---|---|---|
| Training-time alignment | Refusal baked into weights | All products using the model | Retrain | Switch models |
| System prompt rule | The two new sentences | One product (Claude Code) | Near zero | Reframe the ask, jailbreak |
| Output classifier | Scan generated code for asset signatures | Server-side, robust | Inference overhead | Change output format |
| Policy and ToS | "IP infringement is against usage policy" | Legal, not technical | Enforcement effort | All of the above |
Anthropic shipped the cheapest row and cited the fourth. There is no visible output classifier: nothing scans generated code for fingerprints of existing games. Nothing checks whether the user uploaded a real game binary before helping decompile it. The rule is an instruction to a language model that must itself decide, per token stream, whether a request is a "game in a genre" or "a recreation of a specific existing game, even if renamed."
As a classification problem, that boundary is genuinely hard, and Ars Technica's analysis undersells how hard. The model is being asked to implement the abstraction-filtration-function test from copyright law at inference time, with no labels, no threshold, and no ground truth. A block-matching puzzle in the style of Tetris is a genre. Tetris with your logo on it is a recreation. A Mario skin bolted onto Elden Ring, one of the mashups that went viral, is both at once. Ars makes exactly this point: courts cannot reliably draw that line, and we are now asking a sampling process with a temperature to draw it thousands of times a day.
The false positives follow mechanically. A developer debugging their own game will describe it the way anyone describes code, by analogy ("make the jump feel like Mario," "fix this crash in my Tetris clone I built for a jam"). Prompt-level rules do not compose: each additional instruction in a system prompt degrades adherence to the others and shifts how the model weighs ambiguous requests, which is why "analyze this crash dump" can get caught in a rule written about cloning. The rule name being recycled from a previous unrelated rule (tengu_heron_brook predates this purpose, per XDA) reinforces that this is prompt plumbing, not a capability boundary.
The Liability Read, in Numbers
Here is the reading the coverage mostly states in passing but does not price out. Osmani's operative sentence is not the game rule at all: it is "using Claude to infringe IP is against our usage policy." The prompt change is best understood as building a contemporaneous, public record that Anthropic instructed its tool against infringement, in an industry where image-model plaintiffs are already establishing what happens to labs whose terms said nothing.
Two numbers make the incentive concrete. First, the rule costs Anthropic essentially nothing to deploy: two sentences, one client release (2.1.314), zero retraining, and it is reversible in the next prompt push, which is precisely what "we'll continue to iterate" promises. Second, it removes essentially nothing from the market. The task of cloning a game is a prompt plus a model, and the switching cost of a coding-agent task is one config change away. Anthropic's own Kimi K3 comparison we ran is a reminder that open-weight frontier-adjacent models now run 30 percent cheaper on a cost-per-task basis while trailing only slightly on capability. A copier who hits a refusal does not abandon the project; they move to a competitor's hosted model or a local open-weight one, and Anthropic keeps the switching-cost loss and none of the goodwill.
That is the asymmetry: the expected legal benefit accrues whether or not the rule works in practice, because a documented instruction is evidence, while the false-positive cost lands on paying developers doing legitimate work today. Any rational issuer ships that trade. The question for builders is which side of the false-positive line their own prompts land on, and that answer is now a dependency you did not choose.
Outlook: Whose Inbox Is the Exception Handler
Osmani closed his confirmation with an unusual offer for a frontier lab: "if Claude turns down something that's yours, send it my way," directing users' false positives to his personal inbox. That is a manual exception queue for an automated policy, which tells you Anthropic has no programmatic way to distinguish an original developer from a renamed clone, in either direction.
Expect three things from here. The rule will be reworded repeatedly, and each reword is a free diff of Anthropic's internal uncertainty about the genre-versus-recreation boundary; watch the Claude Code changelog and the rules files, because dataminers will publish every version. Second, if false positives stay loud enough, enforcement will move one layer up the stack, either into model-level refusals tuned on actual examples (the Opus 5.5-level instructions Ars reports already exist suggest this has started) or into input-side checks like verifying asset ownership before assisting with decompilation. Third, and least conveniently for the policy's intent, the cloning boom will not notice. It runs on many models, many platforms, and a user base that treats model selection as a dropdown. Valve has said it does not want to be "gatekeepers telling developers which tools they should or shouldn't use," so the distribution layer will not enforce either.
The realistic end state is a two-tier market: mainstream tools that refuse the obviously infringing ask to keep their legal exposure clean, and an uncensored long tail of open-weight models that keep the actual work going, which is the same equilibrium the image-generation market settled into. For game developers, the practical takeaway from the decomp.dev shutdown stands regardless of anyone's prompt: the protection now comes from distribution, secrecy around demos, and legal action against shipped clones, not from the guardrails of a coding tool you did not pick.
The two sentences are real, the policy behind them is real, and neither stops a determined copier for longer than it takes to change a model name in a config file. What they do is draw a line, in public, at the cheapest place Anthropic could draw it, and let a probability distribution guard the border.