Two sentences. That is the entire change, and The Decoder filed it on Nov 3, 2026, as a desk account of an employee's X post rather than a lab announcement. Addy Osmani, an Anthropic employee, confirmed on X that two sentences now sit in Claude Code's system prompt for Opus 5.5 and prevent the tool from cloning other people's games. No changelog. No policy PDF. No newsroom post from Anthropic.

The spec table below carries the parameters the document states. What it does not prove is the exact wording of those two sentences, the specific date the rule went live beyond "several days," or the usage-policy clause Osmani cited. The mechanism is a prompt edit, not a classifier, not a code-level block, not a filter in the inference pipeline. Two sentences in a system prompt are a soft boundary that a user can override by writing their own instructions. The document does not state whether the two sentences replace existing prompt text or append to it.

The trigger was a flood. Users had been tapping Opus 5.5 and GPT-6 Astra to produce hundreds of unofficial browser ports of Halo and GTA Vice City. A tool called Reverse Engineer Anything pairs AI agents with reverse engineering to turn native binaries into readable pseudocode. Sony's PlayStation 5 became a target after a jailbreak led to a quick wave of emulators. One developer recreated parts of Adobe's Creative Cloud apps. The rule kicks in even when a clone gets renamed or passed off as an original work, though users can still build their own games, genre titles, and projects with custom assets.

Die Brief's read: the constraint lives in the prompt, not in the weights. Two sentences are a keyword boundary, not a copyright engine. The fact that some users report running into problems with original projects suggests the filter matches on patterns rather than intent. Anthropic is not building a legal system. It is adding a sentence to a prompt and calling it policy. For the buyer, nothing in the model's capability changes. For the user, a two-sentence prompt filter may or may not survive the next model update, and the boundary between "cloning Halo" and "building a first-person shooter" is thinner than two sentences can hold.

What remains unmeasured: the exact wording of the two sentences, the specific date the rule went live, the usage-policy clause Osmani cited, the name of the developer who recreated parts of Adobe's Creative Cloud apps, and the PlayStation 5 jailbreak method. The document gives "hundreds" for the number of unofficial ports but no exact figure. No lab page has published the prompt diff. The document does not state whether the restriction applies to other Anthropic products or only to Claude Code.

The constraint is a prompt-level boundary, not a classifier or a code-level block, and the false positives on original projects suggest it matches on patterns rather than intent. For the buyer, nothing in the model's capability changes; for the user, a two-sentence filter is thinner than the line between cloning and original work.

The exact wording of the two sentences, the specific date the rule went live, and the usage-policy clause Osmani cited are not in any document.

Two sentences in a system prompt are a soft boundary overridable by user instructions. The pattern-matching behavior implied by false positives on original projects means the filter will degrade as model capabilities improve, and the boundary between cloning and original work is not something two sentences can hold.