Why Claude Code Won't Recreate an Existing Game

Why Claude Code won't recreate an existing game

Claude Code won't recreate an existing game on Opus 5.5 because two sentences were added to its prompt to steer it away from that work. That comes from Addy Osmani of Anthropic, as reported by news outlets. His stated reason is that using Claude to infringe intellectual property (IP) is against Anthropic's usage policy.

Reverse-engineering stops are a separate issue. Users report them, but we found no Anthropic statement that they come from a new rule. The closest documented system is the cyber safety classifier, which is a different mechanism from a prompt. Everything here is the state of the sources on 4 November 2026.

Two issues that the coverage mixes up

Much of the coverage puts these two in one story. They have different evidence and different next steps.

• The game rule: an instruction in Claude Code's prompt. An Anthropic employee stated it on the record. Anthropic has not published the text.
• The reverse-engineering stops: user reports of blocked work. A third-party blog links them to the cyber classifier. Anthropic has not confirmed that link.

A prompt instruction is text that the model reads. A classifier is an automated check that runs on each request. Anthropic documents the classifiers in detail. We found no Anthropic document that describes the game rule.

Each claim below carries one of these source levels:

• Anthropic document: a page on an Anthropic site that we opened on 4 November 2026.
• Employee statement: Osmani's post, as quoted by news outlets.
• News report: something an outlet says, with no Anthropic page behind it.
• User report: one person's experience, relayed by an outlet.
• Third-party analysis: an argument made by someone outside Anthropic.

What is confirmed about the game rule

Employee statement. Tom's Hardware (2 November 2026) quotes Osmani: "We added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game." He also says that original games, genre games and games built from your own designs and assets are meant to keep working.

News report. Game Developer (30 October 2026) says Anthropic confirmed an update to Opus 5.5, and that a spokesperson pointed the outlet to Osmani's post. That post and that referral are the only confirmation we could trace.

User report, not confirmed. The wording that circulates comes from a screenshot posted by an X user called Chasm. Anthropic has not published the two sentences, and Tom's Hardware does not confirm that the screenshot shows the real ones. As described there, the text covers a specific existing game, and renaming the copy does not change that. Game Developer adds that the rule sits under the label tengu_heron_brook, that it was pushed around 27 October, and that Chasm reports no way to disable it. Treat all of this as unconfirmed.

News report. Tom's Hardware says that, as far as it can ascertain, the blocking is limited to Claude Code used with Opus 5.5. If you are new to the tool, start with what Claude Code is. For the model, see Opus 5.5 pricing and plans.

Anthropic document. The Usage Policy page lists infringing the IP rights of a third party as prohibited. That page shows an effective date of 12 November 2026, which is still in the future. It names Claude Code among the apps it covers, and it does not mention games. We could not tell whether this clause differs from the version in force today.

What is reported about the reverse-engineering stops

User reports. Tom's Hardware relays three. Jarred Walton, a former staffer, says his software checks ran with no issue until 31 October and were then blocked. A screenshot shows a refused request to reverse engineer product activation. A Redditor's account was allegedly suspended and then reinstated. These are single reports. Walton's theory of a new rule for reverse engineering is unconfirmed.

Third-party analysis. Capital and Compute (25 October 2026) argues that the cyber classifier causes these stops, and says there is no confirmed Claude downgrade. The post says its evidence is largely user reports. It also says "Anthropic has announced no change to Opus 5.5" and that it is unknown whether the classifier threshold was tightened around 21 October. The post was published before the game rule reports, so it says nothing about that rule.

Anthropic document. The model-switch help article says Opus 5 and Opus 5.5 run classifiers on every user request. The checks also read memory, connector content, web search results and files. Requests flagged as higher-risk offensive security work may fall back to Opus 4.8. The article's examples are exploit generation, binary-based vulnerability scanning and penetration testing. Secure coding stays available, including scanning source code for vulnerabilities.

That article does not name reverse engineering. The 6 October 2026 announcement mentions it once, as "reverse-engineering malware" under Defense Access. The same announcement says generally available models "have conservative cyber safeguards that block most cyber work". So a block on binary analysis fits the documented system, but no Anthropic page we read says that this is the cause of the reported stops.

How to tell which one you hit

Read what Claude Code actually printed. The errors reference documents these cases. It does not mention the game rule, so we cannot say which message a game stop produces.

A cyber flag that ended in an error. On Opus 5.5 and Sonnet 5.5 the message opens like this:

<model>'s safeguards flagged this session

A switch to another model. This is the usual sign of a flag on Opus 5.5, because Claude Code switches automatically unless you change the setting, though it may ask first in an interactive session. A notice appears in the transcript, the request runs again, and the session stays on the new model until you run /model. On Opus 5.5 the target tells you the category: Opus 4.8 means a cyber flag, and Opus 5 means a biology flag.

A Usage Policy refusal. The documented example names a different model, so your text may differ:

API Error: Opus 4.6 can't help with this. Start a new session to continue.

This check reads the whole conversation, which means a new message in the same session usually triggers it again. The same is true after --continue or --resume, because the transcript still holds the content. On Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry, a cyber flag also shows this message, so it does not tell you the category there.

A plain reply. Claude declines in normal text, with no API Error: line and no switch notice. None of the documented checks looks like this. That is consistent with a prompt instruction, but no source confirms it. If the task was to rebuild a named existing game, the game rule is the likely explanation.

To see every flag as it happens, run /config, select Switch models when a message is flagged, and choose Ask each time. Claude Code then pauses at a flagged request instead of switching. This changes what you see. It does not change what gets flagged.

A task can touch both issues, for example analysing the binary of a commercial game in order to rebuild it. We found no source on how the two interact.

What to do if your work is legitimate

This guide gives no way around either restriction. Use the routes below, which Anthropic or Osmani describe.

Your own game

Original games, genre games and games from your own designs and assets are meant to keep working, according to Osmani. Keep your design document and your assets in the project, so that Claude works from your material. If the goal really is a copy of someone else's specific game, the rule is meant to apply. Whether a project infringes IP is a legal question, and neither this guide nor Claude settles it.

If Claude turns down work that is yours, report it. Osmani invited these reports, saying "if Claude turns down something that's yours, send it my way". Inside Claude Code, run /feedback.

Security research

The Cyber Verification Program (CVP) grants access for legitimate cybersecurity work. Its tiers are Defense Access, Red Team Access and Specialized Access. The facts below come from that page, marked "Updated October 2026":

• Apply through the Verification Portal at portal.anthropic.com/programs.
• Anthropic aims to send a decision or a request for more information within seven business days. That is a target.
• Independent researchers, maintainers and bug bounty hunters can apply as individuals, for Defense Access only, on a paid plan. The page does not list which plans count.
• The program requires data retention, so read that section before you apply.
• If you have a grant and are still blocked, check that your admin provisioned it, that you are in the correct Workspace, and that the work falls within your tier.

Two cautions apply. First, Anthropic's pages disagree. The CVP page lists Opus 5.5 and Sonnet 5.5 as covered, but the model-switch article, dated 24 September 2026, says Opus 5.5 "isn't currently available in the Cyber Verification Program". The CVP page looks newer, but check it for the current list before you apply. Second, the only reverse-engineering example we found for the program is malware. Nothing we read says it covers other reverse engineering, so do not assume approval will unblock your task. For the setup side of this kind of work, see running REA for reverse engineering on a VPS.

A wrong flag or refusal

Run /feedback in Claude Code, or use Send feedback in the Claude apps. Anthropic says these reports help it narrow the safeguards. A Usage Policy refusal also prints a Request ID and a Message ID, so quote both if you contact support. The CVP page itself has no process for false positives.

Getting the session working again

Press Esc twice or run /rewind to return to a checkpoint before the turn that triggered the check, then take a different approach. If you cannot find that turn, run /clear. The old conversation stays available in /resume, and resuming Claude Code sessions explains the details. These commands are for moving on to other work. Sending the same task again will not make it allowed.

What is still unknown

These items are open as of 4 November 2026:

• The exact sentences. Anthropic has not published them.
• What a game stop looks like. No source shows a documented message for it.
• Whether Claude.ai and the API apply the game rule. No source says either way. Osmani's statement refers to Claude Code's prompt.
• Whether other models carry it. The reports name Opus 5.5 only. For how the models differ otherwise, see which Claude model to choose.
• Whether the reverse-engineering stops are new. We found no Anthropic statement that names a new rule, or that links the stops to the classifier.

The cyber side is better documented across surfaces. The CVP page lists Claude.ai, Claude Code and the API (application programming interface). On the API, automatic switching is off by default, and the refusals documentation says you receive a normal response with stop_reason: "refusal". That page does not mention games or IP. Do not apply this behaviour to the game rule.

We checked these sources and found no statement of the game rule: the Claude Code changelog, the errors and model configuration pages, the CVP and model-switch articles, the 6 October announcement, the Usage Policy and the refusals documentation. We also ran one search across Anthropic's documentation and help sites. In the changelog we read the first 200,000 of about 1.17 million characters. That range starts at 2.1.318 (4 November 2026) and runs back past 2.1.304 (19 October 2026). Its flagged-message entries cover model switching, and no entry mentions games. A prompt change may not appear in a changelog at all, so this absence proves little.

Expect the balance to change. Osmani wrote: "We'll continue to iterate to ensure we're striking the right balance." A task that is stopped today may work next month, and the reverse is also true, so test your own case again before you rely on this page.

FAQ

Why won't Claude Code recreate an existing game?

Addy Osmani of Anthropic said two sentences were added to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game. He said the reason is that using Claude to infringe IP is against Anthropic's usage policy. This is reported by Tom's Hardware and Game Developer. Anthropic has not published the sentences.

Can Claude Code still help me build a game?

Yes, according to Osmani. Original games, genre games and games built from your own designs and assets are meant to keep working. If Claude turns down a project that is yours, report it with /feedback in Claude Code. Osmani said Anthropic will keep adjusting the balance, so behaviour may change.

Is the reverse-engineering block the same as the game rule?

No source says so, and the evidence is different. The game rule is a prompt instruction that an Anthropic employee stated on the record. The reverse-engineering stops are user reports. A third-party blog links them to Anthropic's cyber classifier, and we found no Anthropic confirmation of that link as of 4 November 2026.

How do I know whether the cyber classifier stopped me?

Look for a message that opens with <model>'s safeguards flagged this session, or a transcript notice that the session moved from Opus 5.5 to Opus 4.8. Both are documented signs of a cyber flag. A plain text reply with neither sign is not one of the documented checks. No source shows what message the game rule produces.

Does the game rule also apply in Claude.ai and the API?

This is unknown. Tom's Hardware reports that the blocking appears limited to Claude Code with Opus 5.5, and Osmani's statement refers to Claude Code's prompt. We found no source that says how Claude.ai or the API behave for this rule. Anthropic's cyber safeguards are a separate system, and they do apply on all three.