Two Sentences, One Scorecard
Anthropic did not ship a new model or a press release for this one. Around October 27, Claude Code quietly picked up a server-delivered instruction, tracked by users under the flag name tengu_heron_brook, that tells Opus 5.5 not to rebuild games it recognizes. There is no settings toggle to switch it off.
The rule itself is short: do not build a recreation of a specific existing game, even if it is renamed or called original. Anthropic technical staff member Addy Osmani confirmed the change on October 30, saying two sentences were added to the Opus 5.5 prompt because using Claude to infringe intellectual property already violated the company's usage policy.
Stripped down, that is a defensible position. The trouble starts where any policy meets a prompt: the model has to guess intent from wording, and it has no documented way to verify who owns the project in front of it. What follows is a comparison of the promise against the field reports.
Category by Category: Where Claude Wins and Loses
Original games: pass. Developers building something new, with their own art, code and design documents, report no change. This is the category Anthropic explicitly protected, and it is the one that matters most for the tool's commercial future.
Genre games: pass, with a fuzzy border. Osmani said titles "in a genre" keep working, but a survival-crafting game that shares a loop and an interface with a famous release is exactly where a language model's judgment gets slippery.
Your own commissioned work: fail, at least once. In a GitHub issue filed October 30, Conflict Simulations LLC reported Claude Code refusing to touch digital board-wargame modules it builds for paying customers from original code and customer-supplied materials. The refusals, the developer says, began right after asking Claude whether the new rule applied to the project.
Reverse engineering, decompilation, crash-dump analysis, anti-cheat work, security reviews: fail in multiple reports. None of that is cloning. One user reported a refusal while reverse engineering an LG smart TV, which may belong to a different filter entirely. Either way, developers cannot tell which guardrail fired.
Modding and performance work: unresolved. Mods that patch an existing game using the player's own files sit in a legal gray zone, and the reports so far are too thin to call. Expect this to be the next fight.
The instruction did not appear in a vacuum. At Steam Next Fest in October, indie studios cancelled public demos and postponed playtests over fear of AI copycats, and some reported extortion attempts threatening an early clone release. Valve knows about it but has declined to ban AI development tools outright.
The Real Head-to-Head: Claude Code vs. the Escape Hatch
Here is where the comparison gets uncomfortable for Anthropic. A restriction enforced inside one vendor's model is only as strong as its competitors' willingness to enforce nothing. Developers on Reddit and X are already naming alternatives: open-weight models, locally run checkpoints, and cheaper hosted rivals such as OpenAI's GPT-6.1 Terra.
The timing explains the caution. Anthropic's IPO prospectus, filed October 26, lists liability from third-party misuse as a risk and books a $1.5 billion settlement over pirated training books as a 2025 expense. Ahead of a Nasdaq debut, a two-sentence prompt tweak is cheap insurance against a screenshot of Claude rebuilding a protected title.
It also collides with a line Mashable landed on: the least ethical AI might be poised to become the most popular. Anthropic has reversed course on over-restrictive filters before, loosening cybersecurity guardrails that were blocking harmless work. The game rule looks like the same learning curve.
Verdict. Anthropic wins on policy clarity and loses on workflow coverage. Open models win on flexibility and lose on accountability. The practical scorecard for developers is blunt: keep a second tool available for IP-adjacent work, file false refusals on the GitHub tracker instead of arguing with the model, and treat any single-vendor coding agent as a policy risk, not just a productivity bet.
Comments