Anthropic Adds a Guardrail to Claude Code Against Cloning Existing Video Games

Anthropic has updated the system prompt for Claude Code, its coding agent, to keep the tool from being used to recreate existing video games. The Verge reported on November 2, 2026 that the change is aimed at cloning specific titles.

Anthropic's Addy Osmani said the company added two sentences to the Claude Code system prompt, the standing instructions the model reads before each session, used with Claude Opus 5.5. The sentences steer the model away from rebuilding another developer's specific game. The stated rationale is that using Claude to infringe intellectual property is against Anthropic's usage policy.

The update is scoped to that case. Original games, genre games, and games built from a user's own designs and assets are meant to keep working, according to the same report. In practice, a request for a new platformer in the style of a popular genre should proceed, while a request to reproduce a particular commercial title should not.

The change was visible to users before the company's public confirmation. Tom's Hardware reported on October 31 that the model was blocking attempts at game recreation and cloning. Users also claimed that non-malicious, generic reverse-engineering tasks, meaning work that takes a finished program apart to understand how it functions, were being refused alongside the game requests. A follow-up from Tom's Hardware on November 2 reported that Anthropic had confirmed new Claude Code limits, and that users were reporting Opus 5.5 was actively blocking reverse-engineering work. Tom's Hardware carried that later report.

Opus 5.5 was announced on September 22, 2026. Anthropic describes it as leading in agentic coding, meaning multi-step programming work done with limited supervision, and knowledge work. The company says it costs 40% less to run than Opus 5 on typical workloads. Claude Code's own history is documented in Anthropic's February 23, 2026 post on its development, which describes the tool's move from an internal command-line interface to the company's coding agent.

The broader context here is a familiar tension in deploying agents. A rule written against a narrow category of request can catch adjacent work that shares its surface features. Reverse engineering sits close enough to cloning that a prompt-level control may not separate the two cleanly. The users' complaints are the kind of collateral effect that broad refusal rules tend to produce.

In my view, the more useful detail is where Anthropic draws the line. The company's stated boundary is a specific existing title, not a genre or a game built from the developer's own assets. That distinction matters for the people most likely to use Claude Code for games, including prototypers, students, and independent developers working from their own designs. If the scope holds in practice, the restriction removes one category of misuse without closing off the ordinary work of building games with an agent. Whether the reverse-engineering side of the policy narrows in the same way is the open question for the next round of updates.