Claude Code puts a brake on game cloning: what firms must not hand to AI
Anthropic has told Claude not to recreate existing games. For companies using AI to produce content, the lesson is that model rules are a filter, not a judgement.

Anthropic has confirmed it added two sentences to the system prompt of its Opus 5.5 model, instructing Claude not to "recreate a specific game belonging to someone else". The trigger was a wave of users reverse-engineering, porting and cloning existing games with the model, to the point where an indie developer's demo could be fully "cloned" soon after release. For corporate users the warning is direct: what AI can do increasingly depends on the few lines of rules a vendor writes into the model — and the boundaries those rules cannot hold must be guarded by the people using it.
Two sentences in a prompt will not stop anyone determined to route around them
Look first at what the ban says. According to 36Kr, the new Claude Code rule is that users may create original games, games of a specific genre, and games built from their own designs and assets, but must not reproduce an existing specific game — even if the name is changed or it is presented as original.
The wording is clear. The problem is that it only binds Claude itself. According to 36Kr's analysis, someone set on cloning can build the main framework with another model, then break the task into narrow technical jobs that carry no visible context and hand those to Claude — strip out every related name and label and ask it only to "implement some kind of creature-capture and battle mechanic", and it has no idea what it is taking part in. That kind of decomposition defeats the rule.
The trouble at the other end is more practical: a blanket ban catches legitimate work. People making original games, and people doing creative reworking of assets, can both get stuck on this rule.
The essence of the matter is that model vendors are using prompts as soft constraints, and the reach of a soft constraint depends on whether users are willing to play along. It stops someone trying their luck. It does not stop someone deliberately taking a detour.
When firms use AI to produce content, the real risk sits in the process, not the model
Move the view away from games. When a company uses AI to write copy, produce posters or make Xiaohongshu (a Chinese lifestyle and social-commerce platform) image-and-text posts, it faces the same class of problem: what a model can generate and what you may actually use are two different things.
Game cloning runs into copyright and originality. Corporate content runs into smaller, more scattered traps — using someone else's image, saying words that should not be said, treating an AI first draft as a final draft and publishing it. None of these disappear automatically because a model vendor added a sentence to a prompt.
Shu Yi, executive director of Sheyu Jiye, has a judgement on this: what AI writes is a first draft, and the boss's judgement and real experience are the most valuable part of the content. It fits here exactly — AI has raised the speed of output, but the gate on "what to publish and whether it can be published" still has to be kept by a person.
Two things to look at before using AI for content-led customer acquisition
Take SHEYU AGENT. Its official positioning is "the boss's first AI team": 16 industry advisers, each covering a professional area, and 34 zero-threshold tools covering Xiaohongshu image-and-text posts, AI posters and copywriting, with one login working across desktop, phone and web.
What connects directly to this incident is the two settings it applies at the output stage.
First, prohibited words plus a brand profile. Platform regulatory red lines plus company-defined prohibited words are checked before content is produced; the brand profile and product library are shared across the team, and when images are generated the logo uses the real file you uploaded. This addresses the problem of "saying the wrong thing and using the wrong mark", putting content through a screen before it goes out.
Second, Xiaohongshu image-and-text creation. It generates and reviews native multi-image posts and accompanying copy, for the user to confirm before use under platform rules; account permissions and submission processes are opened only after acceptance, and where a platform restricts AI hosting, materials are retained for manual publishing and automatic posting is suspended, with the actual publishing result subject to the original post's receipt.
These two capabilities point to one thing: they leave "review" and "confirmation" in human hands rather than deciding for you. What they cannot do is equally clear — a prohibited-word list will not stop you from deliberately copying someone else's idea, and a brand profile will not stop you publishing an AI first draft as it stands. Rules are a sieve; the judgement is still yours.
What to do now
Back to Claude Code. In the short term, the two sentences Anthropic added are a statement of attitude. In the long term they are a signal: model vendors will embed usage boundaries in their products more and more, and those boundaries are set by each vendor individually and can be adjusted at any time.
For companies this means two things.
First, do not stake all the compliance responsibility for content production on a single model. A way of writing that works today may be blocked by the rules tomorrow; and what is not blocked today is not necessarily fine.
Second, fix the "human confirmation" step into the process rather than relying on good intentions. AI produces fast, people read slowly, and the gap in speed is where the risk lives.
Shu Yi has also said: in doing digitalisation, first find the thing that repeats every day and goes wrong every day, and start there. In content-led customer acquisition, the thing that repeats every day and can go wrong every day is probably "nobody takes a second look before it is published".
Changing from that step is more practical than chasing every new rule from every model.