
On November 2, 2026, The Verge AI reported, citing Tom's Hardware, that Anthropic has updated Claude Code's system prompt to bar the coding agent from being used to rebuild, decompile or port existing video games. It marks the first time a major model developer has written a game-cloning prohibition into a coding tool's system-level instructions rather than leaving it in general usage terms. Liu Gong believes the real significance lies not in enforcement strength but in moving IP governance
The November 2 report states that Anthropic has added a new constraint to Claude Code's system prompt: the tool must not be used to recreate, decompile or port existing video games. The change sits at the prompt layer and does not involve a Claude model weight update, taking effect immediately for users who call Claude Code through its command-line interface and IDE extensions.
The boundary drawn is partial. Claude Code can still be used to build original games; what is restricted is the cloning path, including decompiling an original ROM into readable source and recompiling it into a PC or mobile port. Anthropic has not published the clause text or specified what happens when it is triggered, such as whether the tool refuses outright or accompanies the refusal with an explanation.
The immediate trigger is the state of the retro gaming scene: AI-generated decompilations,
Decompilation has long been the technical high ground of the retro gaming community. Restoring a 1990s 3D game from machine code to compilable C code historically required dozens of volunteers, two to three years of effort, and function-by-function manual verification. Large models compress that pipeline's unit cost by an order of magnitude through assembly comprehension, symbol inference and cross-file refactoring, turning projects that were "worth three years" into experiments "worth two weeks."
Meanwhile, coding agents have become one of the primary paid entry points for large models. The system prompt is the cheapest control surface available: editing a paragraph of text is cheaper
Liu Gong's take: this change will not stop anyone determined to clone a game, but it will become industry convention within twelve months. Within the next six months, at least one mainstream coding agent at the GitHub Copilot or Cursor tier will add a comparable clause to its system prompt, and will publish the prompt diff as a compliance asset.
From a technical standpoint, the clause is a soft gate. Claude's context window runs to 200K tokens, so a constraint of a few dozen tokens occupies under 0.02% of the window; inference cost and latency are effectively untouched, and that is precisely why the clause carries little force. There is no verification mechanism attached. A user who states in conversation that they hold a legitimate cartridge and ROM licence will likely be waved through. Reliable interception lives on the output side, in code fingerprinting, binary similarity scoring or a dedicated classifier; Anthropic did not touch that layer. In industrial visual inspection, Liu Gong's habit is to ask whether a defect can be caught inside the process rather than at final inspection, and to demand measurable recall and false-positive rates. A prompt clause offers neither.
On the business side, this is risk pricing. Buyers of coding agents are shifting from individual developers to enterprise teams, and "will your tool generate infringing code for me" is a mandatory question in procurement legal review. A paragraph of prompt text buys a contractual answer at a fraction of the cost of defending a claim. The cost lands on users: a porting consultation that once fit in a single conversation now needs multiple turns, self-verification, or a different tool.
For the industry, a compliance premium emerges. Closed agents carry the clause, open-weight models do not, and decompilation workflows drift toward the less constrained end. The result is a split: licensed tools for commercial projects, unconstrained models for grey ones. That split will not eliminate cloning; it will push it outside the visible layer, much as DRM and emulator communities evolved in the past.
- Claude Code's system prompt runs to several thousand tokens by published documentation, and the new prohibition likely accounts for 20 to 50 tokens, under 1% of the total.
- Claude's 200K-token context means a 30-token clause occupies roughly 0.015%, making cost and latency impact negligible; the binding constraint is vendor willingness, not compute.
- Comparable decompilation efforts, such as the fan-made PC port of The Legend of Zelda: Ocarina of Time, historically relied on dozens to hundreds of volunteers over two to three years; AI assistance can compress equivalent workloads into weeks.
- Cross-vendor comparison: OpenAI and Google prohibit infringement in general usage policies but
For the industry, the near-term effect within three months is that the compliance gap between tools becomes public: enterprise buyers add "does it restrict infringing generation" to procurement checklists, and decompilation communities begin routing around constrained closed agents toward local or open-weight alternatives. Over a one-year horizon, system prompts shift from internal implementation detail to auditable compliance documents, with vendors converging on a set of disclosure norms around behavioural clauses. The harder question, whether machine-generated code counts as a derivative work, will still be settled by case law rather than by prompt engineering.
For ordinary users,
Questions worth exploring:
- System prompts are replacing usage policies as the cheapest and least transparent governance lever at model developers, which argues for third-party auditing.
- AI-assisted decompilation has cut the marginal cost of game porting by an order of magnitude, yet no precedent defines how copyright treats machine-generated compilable source.
- A compliance premium will push workflows toward open-weight models, eroding closed vendors' advantage in IP-sensitive scenarios.
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-11-02 18:56:02). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.
