Threat assessment: moderate. Mostly for people cloning someone else’s game. Low for everyone else.
Anthropic has changed how Claude Code handles game development requests. According to The Verge AI, the company added two sentences to Claude Code’s system prompt on Opus 5.5. Those sentences steer the model away from rebuilding a specific existing game. The reason is simple: using Claude to infringe intellectual property already breaks Anthropic’s usage policy.
The statement came from someone at Anthropic and was short. “Original games, genre games, and games built from your own designs and assets are meant to keep working,” it reads. The company also says it will “continue to iterate to ensure we’re striking the right balance.” Builders who get blocked on their own work are asked to report it.
Situation report
Here’s what changed, as reported by The Verge AI:
- Scope. The change only affects Claude Code running on Opus 5.5.
- Mechanism. It’s a system prompt edit. Two sentences. No retrained model and no separate filter layer were announced.
- Target. Requests to recreate a specific game made by someone else.
- Protected zone. Original games, genre games (think “a roguelike” or “a match-3 puzzler”), and projects built from your own designs and assets.
- Feedback loop. Anthropic says it will keep tuning the line and wants to hear about false refusals.
Why this matters
The news is in how small the change is. Anthropic didn’t ship a classifier or a new policy document. It changed the model’s standing instructions, and that alone shifts behavior across every Claude Code session on Opus 5.5.
That tells you two things. First, system prompts are now a real policy lever for agentic coding tools. A tweak lands everywhere at once, and nobody gets a changelog in their IDE. Second, it’s a blunt lever. A model reading a prompt has to judge where “inspired by” ends and “a copy of” begins. That’s hard for lawyers. It’s harder for an LLM in the middle of a task.
The pressure behind this is easy to see. Agentic coding tools are now good enough to build playable clones of well-known games from a one-line prompt. That’s a great demo. It’s also a legal mess for the company that supplies the model, since game studios guard their characters, levels and mechanics closely. Anthropic is already in the IP spotlight across the industry, and it clearly doesn’t want Claude Code to become the easiest way to knock off a hit game.
Context: the status quo before
Until now, the usage policy banned IP infringement, but enforcement mostly sat with the user. Claude would often go along with “build me a copy of [famous game]” and leave the legal risk to whoever shipped it. The new prompt moves some of that judgment into the model. Claude now acts as a first-pass gatekeeper.
This matches a wider trend. AI labs are moving away from “we wrote a policy” toward “the model enforces the policy in real time.” Image generators went through this with artist styles and trademarked characters. Coding agents are next.
Tactical guidance for builders
If you build games with Claude Code, plan for the following:
- Frame requests by mechanics, not titles. “A grid-based falling-block puzzle with line clears” reads as genre work. Naming a specific franchise and asking for a faithful copy is what triggers the new behavior.
- Bring your own assets. Anthropic explicitly protects games built from your own designs. Say so in the prompt when it applies.
- Expect false positives early. A two-sentence instruction will misfire sometimes, especially on parody, homage, or remakes you hold the rights to.
- Report wrongful refusals. Anthropic is asking for this feedback directly. Edge cases you report will likely shape the next version.
- Watch other vendors. If this works for Anthropic, expect similar guardrails in other coding agents.
Outlook
This is version one, and Anthropic says as much. The open question is whether a prompt-level rule can tell the difference between a homage and a clone often enough to keep legitimate game developers happy. If refusals pile up on original work, expect more tuning, or a move toward more formal tooling. If they don’t, this light-touch approach could become the template for how AI coding tools deal with IP.
Full details are in the original report from The Verge AI.