claude code clones: Opus 5.5 Prompt Blocks Game Remakes

Anthropic just made claude code clones of specific video games a lot harder. Reporting from AI/TLDR and AIpresso says Claude Code on Opus 5.5 now refuses requests to recreate an existing title, even when the ask is rephrased as an "original" remake.

What actually happened

According to the November 2 AI/TLDR digest, Claude Code now declines to recreate someone else's specific game and cites Anthropic's IP rules. AIpresso's November 2 edition matches that account: guardrails were added after complaints piled up, and Osmani confirmed the two-sentence prompt append on Opus 5.5.

That framing matters. This is product-policy text sitting in front of the model, similar in spirit to other Claude Code steering we have watched around unslop claude skill. Weights did not have to change for behavior to change. If you swap models, wrappers or local hosts, you may see different results — AIpresso notes some people claim Opus through other harnesses can sidestep the block — but the official Claude Code + Opus 5.5 path is where the refusal is landing.

The allowed side of the policy is as important as the ban. Genre games (a new platformer, a new puzzle game) and builds from the user's own art, design docs and assets remain in bounds. The bright line is cloning a specific existing title, including renamed knockoffs.

Why game cloning became a flashpoint

Coding agents are finally good enough to scaffold playable slices of famous games from a short prompt. That thrills hobbyists and alarms rights holders. Ports, "demakes," fan remakes and classroom recreations all sit near that line, and developers using Claude Code for those workflows are the ones hitting the new wall mid-project.

AI/TLDR also flags unverified reports that broader reverse-engineering asks are getting refused. Treat that as anecdotal until Anthropic documents it. AIpresso likewise says reverse-engineering refusals are mounting but does not present a formal policy PDF. For now, the confirmed piece is narrow: specific existing video games inside Claude Code on Opus 5.5.

The competitive backdrop is loud too. DeepSeek's open Super-Agent prototype is being pitched as a Claude Code rival — we covered that lane in deepseek superagent prototype — and mobile agent experiments like mika android agent show how fast "agent that writes software" is spreading beyond one IDE.

What still works

If you are prototyping an original game, the claude code clones crackdown should not block you. Describe mechanics without naming a commercial title. Bring your own sprites, maps and design notes. Ask for engines, patterns and debugging help rather than "rebuild Game X from scratch."

If your project is an authorized port or you hold the rights, you may still need to prove that context in-prompt, and even then a blunt safety line can false-positive. Keep local copies of working trees so a mid-session refusal does not strand you.

Teams building agent harnesses should assume IP-sensitive categories will keep getting prompt patches. Game clones are simply the visible example this week.

Two sentences in a system prompt can erase a whole weekend workflow.

Prompt patches versus model cards

Users often learn about claude code clones limits from a refused chat, not from a versioned changelog. That asymmetry will keep producing surprise posts on X whenever Anthropic tightens another category. If your workflow depends on an agent recreating copyrighted mechanics, build a fallback plan now rather than after a deadline slips.

Document the exact model picker setting (Opus 5.5 vs other) when something works. A teammate on a different default can hit a different policy surface and think the project "broke."

What to do as a reader

If you use Claude Code for games, read the November 2 write-ups on AI/TLDR and AIpresso, then retest your own prompts on Opus 5.5 before promising a remake timeline to anyone.

If you follow AI product policy, log this next to other agent guardrail stories: small prompt edits can ship faster than model cards, and they can surprise users who thought "code anything" was the product pitch.

Nothing here is legal advice. Prompt refusals are product behavior, not a ruling on fair use, and other tools may still attempt the same task with different risk.