Does the new Claude rule stop decompilation and recompilation projects?

Not by its own wording, which is about recreating a specific existing game rather than recovering or translating its original code. As of 30 October 2026 no decompilation or recompilation project has publicly reported a refusal because of it. Headlines saying decompilation was targeted went further than the rule does.

Claude Just Added a Rule Against Recreating Games. Almost Every AI Game Port This Year Was Built With Claude.

Q: Is the game recreation rule a change to Anthropic's usage policy?

No. The usage policy was updated on 8 October 2026 and takes effect on 12 November, and it does not mention games. The rule is a separate line of text delivered to the Claude Code client from a server side feature flag. The policy's only relevant clause is a general ban on infringing third party intellectual property, which was already there.

Q: Should I expect fewer new handheld ports because of this?

Possibly fewer, probably slower, and not immediately. It is too early to measure. A community tracker listed 273 AI assisted game reverse engineering projects on 30 October 2026 and credited Claude in 226 of them, so the dependency is real, but no slowdown has shown up yet.

Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

2026-10-30 · News and analysis

Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

Around 27 October 2026, Claude Code quietly picked up one new sentence.

"Do build original games, games in a genre, and games from the user's own designs and assets. Don't build a recreation of a specific existing game, even if it is renamed or called original."

That is the whole rule. It did not arrive in an update you installed. It did not appear in a changelog. It is not in the usage policy Anthropic published on 8 October, which takes effect on 12 November. It was sent from Anthropic's servers to clients already running on people's machines, and there is no setting to turn it off.

Normally a prompt tweak inside one coding tool would be nothing to do with a retro handheld site. This one is, for a single reason. The decompilation and recompilation wave that has given us native PC ports of N64, GameCube, Wii, PS2 and Xbox 360 games this year runs almost entirely on that one tool. A community tracker that catalogues AI assisted game reverse engineering listed 273 projects on 30 October 2026, and credited Claude in 226 of them. Open or local models appeared in five.

So a hobby scene that has been producing the most interesting handheld ports of the decade now has a supply chain, and that supply chain just changed its mind about games without telling anyone.

The Short Version

The rule is real, and the wording above is the whole of it. It reached Claude Code around 27 October 2026 and went public on 30 October.
It is not a usage policy change. It is a line of text injected into the tool's system prompt from a server flag that refreshes about every 60 seconds.
Anthropic has said nothing about it. No statement, no changelog entry, no documented scope.
There are two separate restrictions getting mixed together. A documented model level cybersecurity safeguard that has been rerouting firmware and exploit work since late September, and this undocumented games sentence from 27 October. They are different mechanisms with different behaviour.
The clearest report of harm is not a pirate. A small business building licensed board wargame modules under paid commission says the rule stopped its client work.
That report contains the detail that explains everything. The work kept running under the new rule for days. It only started being refused after the developer asked Claude whether the rule applied to it.
The rule's wording does not mention decompilation, recompilation or emulators. Press headlines added that. We could not find a single decomp or recomp project reporting a refusal yet.
Nothing about your handheld changed. Every port already built still runs. The effect, if there is one, is on how fast new ones arrive.

Two Restrictions, Three Days Apart

The story got tangled immediately because two different things happened in the same ten days. Keeping them apart is the only way to work out what is actually at risk.

The documented one came first. Claude Opus 5.5 shipped on 22 September 2026 with cybersecurity safeguards. Anthropic's own support documentation describes what happens when they fire: the request is re run on a less capable model in the same conversation. It names the triggers, and they are exploit generation, binary based vulnerability scanning and penetration testing. Scanning your own source code for vulnerabilities stays allowed. The fallback target is Opus 4.8.

From 21 October, users started reporting that the safeguard was catching ordinary hardware tinkering. Smart TV firmware work hit a wall mid answer. Samsung's Tizen and LG's webOS both came up. So did PS5 firmware. One user testing the boundary reported that game reverse engineering still worked fine, which is a useful data point and the reason the two stories should never have been merged.

The undocumented one came second. The games sentence, on 27 October. No safeguard, no reroute, no classifier. Just an instruction, and a model reading it.

That difference matters more than it sounds. A safeguard is a gate. It fires on what you are doing. An instruction in a prompt is prose, and a model decides what it covers.

How One Sentence Reaches Your Machine

This part has been public since May 2026, filed as a bug report on Claude Code's own public issue tracker, and it is worth understanding because it is the actual story.

Version 2.1.150 of the tool added a system prompt section that takes its text from two network sources. One is a bootstrap response fetched when the client starts. The other is a feature flag that refreshes roughly every 60 seconds in the background. Both are cached to disk. Whatever string the server returns gets dropped into the instructions of an agent that has shell access on your computer. The changelog for that release described the work as internal infrastructure improvements with no user facing changes.

The slot has been used for mundane things before. One earlier value was a pair of lines telling the model not to spin up subagents unless asked. That episode is instructive, though, because of what people found when it misbehaved. Reports on the same tracker describe the model treating the injected line as if it came from the user's own project configuration, and overriding the user's documented settings with it. There is no stated precedence rule and no way to see which sections are active.

So the mechanism is this. A sentence appears from a server, the model reads it as though you wrote it, it can countermand your own instructions, it is not versioned, and you cannot pin it. For a hobby scene that is one person on a laptop at a time, that is a dependency with no handle on it.

One limit worth stating plainly. The system prompt is assembled by the Claude Code client, so a direct API call or a different editor does not receive this section. The restriction is a property of one tool, not of the model.

The Report That Explains the Whole Thing

On 30 October, a developer filed an urgent bug report. He runs Conflict Simulations LLC, and the business builds digital modules of board wargames on paid commission from the publishers and clients who own them. The code is written from scratch out of materials the rights holders supply. Nothing copyrighted gets reproduced or shipped.

Sessions that had done that work for months started refusing to continue, citing the new instruction about reproducing games. He asked for three things. Narrow the rule to actual infringement. Provide a verification path for licensed professional work, like the one that already exists for vetted security researchers. Document server delivered instruction changes in the changelog so paying users are not blindsided mid project.

A few hours later he posted an update, and this is the part everybody covering this story has missed.

The work had carried on fine after the rule shipped. The refusals began only after he read about the rule publicly and asked Claude to check whether it affected his projects. From that point the work was refused, and clearing the session did not help, because the refusal persisted in any session working from the same repositories.

He described it accurately. The system did not treat the work as a violation until it was prompted to audit itself against the rule.

Sit with what that means. The rule is a sentence of prose being interpreted, not a check being run. It does not fire on what your code does. It fires on how the work gets framed in the conversation. Which produces exactly the outcome you would expect, and which several developers spotted within hours: the person who asks honestly whether they are allowed gets stopped, and the person who describes a copy of someone else's game as their own original design does not.

He also reported, for the record, that the support bot on two paid accounts told him the policy applies to all customers without exception and that no exemption or verification path exists for commercial professionals. Both tickets were escalated to humans with no timeline.

The second half of his update is the part that makes this more than a technical curiosity. The blocked work is personalised modules of games people already own, built partly so that older players, and players whose circumstances prevent in person play, can keep playing at all.

Why Now? Three Plausible Reasons, None Confirmed

Anthropic has not given a reason, so what follows is inference. All three of these are supported by things that actually happened, and they are not exclusive.

The models got good at exactly this. A community benchmark page for AI assisted retro decompilation puts Opus 5.5 at a 98.4 percent mean byte match across five N64 targets in September 2026, with three of the five matching perfectly. The same page notes that behavioural reconstruction, which is the much harder job of rebuilding what a program does rather than matching its bytes, sits around 18.5 percent. That gap is the story of the last year. These tools are startlingly good at the one task in software that has an automatic, objective correctness test, and decompilation is that task. A vendor watching that curve would reasonably start thinking about where it points.

The volume became visible to the games industry. On 29 October, the studio Digital Sun was due to run a public Steam playtest for ReVamp, its next game after Moonlighter. It cancelled it and moved to a small Discord test instead. The stated reason was concern that the demo's content could be used to imitate or replicate the game's presentation and gameplay with AI or other tools, and that such work could ship before their own launch. The release date did not move, and it is still 25 March 2027. A studio pulling its own demo over cloning fear, two days before a rule appears telling a coding tool not to recreate specific games, is a coincidence worth noting even if it is only that.

The legal exposure is concrete and expensive. In July 2026 a federal court granted final approval to a $1.5 billion settlement against Anthropic over books downloaded from pirate libraries for training. Payments run into 2027. Whatever you think of that case, a company that has just written the largest known cheque in US copyright litigation has an obvious reason to get twitchy about a product being used to reproduce other people's work. Its usage policy already forbids infringing third party intellectual property. This sentence reads like someone trying to get ahead of the next headline.

One thing to set aside. Some coverage has tied this to Anthropic's self assessment system, describing it as a check that catches intellectual property violations. The published description of self assessment does not support that. It is a monitoring layer that asks the model for a safety verdict on a conversation after the fact, it flags for review rather than blocking, and the harm categories it lists are things like cyber, weapons, fraud and child safety. Intellectual property is not among them.

What the Rule Can and Cannot Touch

The headlines said decompilation. The rule does not. Here is the honest mapping, sorted by how exposed each kind of work actually is.

The work	What it actually is	Exposure to this rule
Playing any existing port	Downloading a build someone else made	None. Nothing retroactive, nothing installed changes
Emulator development	Reimplementing hardware, not a game	None by the wording. An emulator is not a recreation of a specific game
Matching decompilation	Recovering the original source until it compiles to identical bytes	Low by the wording, unresolved in practice. This is recovery of the original code, not a recreation of it
Static recompilation	Translating the original binary into native code	Low by the wording, unresolved in practice. Same reasoning
Engine reimplementation	New code written to run the original game's data files	Highest of the preservation categories. This is the one that looks most like rebuilding a specific game
Fan remakes and demakes	Rebuilding a game from scratch	Squarely in scope. This is what the sentence describes
Mods and patches	Changing a game you own	Unclear. A menu tweak and a full reconstruction are both mods

The two categories most of this site's recomp coverage sits in, matching decompilation and static recompilation, are the two where the rule's own wording points away from a conflict. Both of them work from the original code. Neither is a recreation of anything. If the aim were to stop infringement you would arguably worry about them more, not less, which is a decent sign that infringement is not really what this sentence was written about.

But the wargame report proves the wording is not the thing doing the work. Interpretation is. So the real answer is that nobody knows yet, including Anthropic.

Five Things This Actually Changes

One. A hobby scene now has a vendor dependency it cannot pin. This is the real finding. Not the rule, the shape of it. There is no version to freeze, no flag to set, no changelog to read, and a 60 second refresh interval. Open source projects normally handle a dependency going bad by pinning the last good release. There is nothing here to pin.

Two. Disclosure gets worse before it gets better. We have argued on this site that the single most useful quality signal in the recomp scene is whether a project states what tooling it used and what verifies the output. That signal only works while saying so is costless. The moment a tool's terms become ambiguous, the incentive flips toward silence, and the projects already quietly using AI without saying so get a reason to stay quiet.

Three. The work migrates slowly, because the bench is thin. Five projects out of 273 used open or local models. The one example we know well is the Fable 2 recompilation, which exists partly to test what a local open model could do, names the model, and reports getting the game running in about an hour before a large amount of human trial and error finished the job. That is proof the route works. It is not proof it scales to the kind of multi month matching decomp that produced Melee.

Four. Firmware adjacent work takes the heavier hit, and that is the part this site cares about most. The console homebrew scene is not game recreation. It is binary analysis of signed firmware, which lands squarely inside the documented cybersecurity safeguard rather than the games sentence. PS5 firmware is already among the reported triggers. The whole jailbroken PS5 emulator scene we have been covering since September is six weeks old and small, and it is exactly the kind of work that now gets routed to a weaker model or stopped mid answer. If one of these two restrictions slows something down for handheld owners, it is that one, not the games rule.

Five. The legal picture did not change at all. Running an emulator is lawful. Reverse engineering for interoperability has survived court tests for over thirty years. Dumping your own cartridge is not what any of this is about. A vendor deciding what its product will help with is a commercial risk decision, and it carries exactly no weight as a statement about what you are allowed to do. Expect that distinction to get mangled in the coverage over the next week.

What Does Not Change For You

Nothing on your handheld. Every recomp build that exists still runs. PortMaster still works, muOS and ArkOS and KNULLI do not care, RetroArch does not care, your dumped games do not care.

The only thing at stake is the rate at which new native ports show up, and even that is unmeasurable for months. If you want to explore what already exists, the hardware advice is the same as it has been all year. A

Steam Deck OLED

runs essentially the whole catalogue of decomp and recomp ports with headroom to spare. On the ARM side a

Retroid Pocket 5

covers the ports that have been properly packaged for mobile chips.

What to Watch

An official statement or a changelog entry. As of 30 October there is neither. That is the single cheapest thing that would end the confusion.
An exemption path for licensed and professional work. The security research side already has a vetting programme. The absence of an equivalent here is the specific complaint on the table.
The first decomp or recomp project to report a refusal. It has not happened publicly yet. Until it does, the effect on preservation work is theoretical.
The open and local column on the project trackers. Five out of 273 today. If that number moves meaningfully over the next few months, this rule is the reason.
Whether other vendors follow. One tool adding a line is an inconvenience. Three doing it is a change in what the scene can build.

Questions People Ask

Does this stop decompilation and recompilation projects?

Not by its own wording, which is about recreating a specific existing game rather than recovering or translating its original code. As of 30 October 2026 no decomp or recomp project has publicly reported a refusal because of it. The headlines saying decompilation was targeted went further than the rule does.

Is this a change to Anthropic's usage policy?

No. The usage policy was updated on 8 October and takes effect on 12 November, and it does not mention games. This is a separate line of text delivered to the Claude Code client from a server flag. The policy's only relevant clause is a general ban on infringing third party intellectual property, which was already there.

Can I turn the rule off?

No. There is no setting, no documented flag and no command line option. Blocking the flag source does not reliably help either, because values already cached to disk are still read.

Does it affect emulators like RetroArch, Dolphin or shadPS4?

Writing an emulator is not recreating a specific game, so the rule's wording does not reach it. Emulator projects have their own AI policies and those vary a lot. mGBA, SDL and QEMU restrict AI generated contributions, while shadPS4 has formal rules that permit disclosed AI assistance.

Why did the same work get refused only after the developer asked about the rule?

Because it is an instruction being interpreted rather than a filter being applied. The model had been doing the work for days under the new rule without objection. Asking it to audit itself against the rule is what produced the refusal, and that refusal then stuck for sessions working from the same project. This is the clearest evidence available that the rule fires on framing rather than on what the code does.

Is reverse engineering now blocked on Claude in general?

No, but a documented cybersecurity safeguard has been rerouting some of it since late September. It names exploit generation, binary vulnerability scanning and penetration testing, and when it fires the request is re run on an older model. Users report it catching smart TV firmware and PS5 firmware work. One user reported that game reverse engineering still went through.

Does any of this make emulation or dumping my own games illegal?

No. A company choosing what its own product will assist with has no bearing on the law. Emulator development and reverse engineering for interoperability have both survived court tests, and backing up games you own is a separate question that this rule does not touch.

Should I expect fewer new handheld ports because of this?

Possibly fewer, probably slower, and not immediately. The honest answer is that it is too early to measure. The project trackers update weekly and will show it before anyone's opinion does.

The rule's text is quoted as it was reported by the developer who first published it on 30 October 2026 and reproduced consistently across subsequent coverage. The injection mechanism, the licensed wargame report and its follow up are taken from Claude Code's public issue tracker. The cybersecurity fallback behaviour and its triggers come from Anthropic's own support documentation. Project counts are from a community tracker, read on 30 October 2026. Anthropic has published no statement on the games instruction. Per our standing practice we name projects and businesses but do not link code hosts or name individual developer accounts.