On 27 October, Claude Code quietly stopped doing a category of work it had been happy to do for months: rebuilding existing video games. No release note, no changelog entry, no version bump. The instruction arrived from Anthropic's servers under a feature flag called tengu_heron_brook, and it is two sentences long.
"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."
There is no setting to turn it off. For most developers, the first they knew of it was the tool politely declining something it had cheerfully done the day before. The rule surfaced publicly on 30 October, and by the following morning trade press had picked it up: Game Developer ran "Anthropic orders AI tool Claude to stop remaking video games", and the thread underneath the original post had already passed 1.7 million views.
What the rule actually bans
The wording is narrower than the panic suggests, and broader than the reassurance does. Original games are explicitly fine. Working inside a genre - roguelike, platformer, city builder - is fine. Building from your own designs and assets is fine, and always was.
What is out is recreation: a specific existing game, rebuilt, and the clause doing the heavy lifting is the trailing one - "even if it is renamed or called original." Ask for "a block-stacking game with seven shapes" instead of Tetris and you have not found a loophole, you have described the thing the rule is designed to catch. It is aimed at intent, not vocabulary.
The genuinely awkward territory is the middle, where most actual work lives. A mod loader that reads a game's own data files is arguably the "user's own assets" side of the line. A demake, a fan remake, a faithful reimplementation of a 1985 arcade cabinet for a modern display pipeline - those are recreations by any honest reading. And that middle is exactly where the preservation and recompilation scene this site documents spends its time.
The crybully problem: policy delivered by remote control
Here is the part that should bother people more than the rule itself. Whatever you think of the copyright logic - and game rewrites are a genuine minefield - the rule was not announced. It was not documented. It did not ship with a version number you could pin, or a note you could read, or a switch you could flip. An instruction was pushed from a server, and the behaviour of a tool thousands of professionals depend on changed underneath them.
That is not a legal strategy. A lab worried about copyright exposure publishes terms and lets you plan around them. This is something closer to a crybully move: impose a restriction without admitting you imposed it, dress it as safety, and leave the user with no recourse except to discover the change by hitting it. The model did not get dumber. It got supervised, by someone who did not think to mention it.
And there is a cost that never appears on the balance sheet: every one of these silent adjustments teaches your users that the tool they rely on is provisional. You cannot build a workflow on behaviour that can be retracted between breakfast and lunch.
Every silent restriction is a DeepSeek advertisement
Which brings us to the strategic error, and it is a large one.
Anthropic's competitive position rests on capability, and on the assumption that users will accept supervision in exchange for the best model. That assumption holds right up until the supervision starts costing people work they were being paid for. Meanwhile the open-weight side of the field has one structural advantage that no amount of benchmark leadership cancels out: a model you download cannot be instructed from someone else's server.
This is the whole argument, and it is embarrassingly simple. DeepSeek and its open-weight peers can be run on your own hardware, pinned to a specific version, and used for whatever you want - including the ugly grey-area work that labs would rather not be seen facilitating. Nobody pushes a flag to your local weights. Nobody decides on a Tuesday that your workload is now a liability. There is no tengu_heron_brook in a file you control.
So every time a closed lab adds an invisible constraint, it does two things at once. It makes its own product slightly worse for the people who use it professionally, and it hands the open model a fresh argument that requires no marketing budget at all. "Your tool would not have done this to you."
Capability moats are temporary. The direction of travel for open weights has been one-way for three years, and the gap at the frontier is measured in months, not generations. What does not close on that schedule is trust. Right now the loudest voice arguing for DeepSeek is not DeepSeek - it is the pattern of behaviour coming out of the labs that keep raising the walls.
What this looks like from the developer's chair
The replies to the original post split predictably. One camp argued that "open-source abliterated models will take over instead" - the standard prediction that any capability withdrawn from a hosted model migrates to a local one that cannot be withdrawn. Another put it more sharply:
"Eventually Claude's entire context window will be taken up with prissy prohibitions and instructions to report the user to the authorities for unauthorized whistling"
Both are the same observation from different angles: the friction is being added at exactly the layer users can neither see nor remove. And friction at that layer is the one thing that reliably pushes people toward rigs they own.
If this just interrupted your project
- Keep your own assets and design documents. The rule explicitly protects work built on your own designs. Provenance is your defence, and it is also just good practice.
- Describe mechanics, not titles. Not as a trick - it is the honest framing, and it is what the rule asks for.
- Keep a local record of prompts and outputs. If behaviour can shift overnight, the export is the only copy you actually control.
- Do not build a pipeline on an instruction you cannot pin. Hosted models are rented behaviour. If the work matters to your income, keep a local or self-hosted path warm.
The rule is two sentences and will probably be forgotten inside a month. The precedent will not be. Behaviour delivered by remote instruction, invisible to the user and unappliable by them, is a governance model - and it is one a lab cannot credibly call "the safe choice" while an open competitor waits, unswitchable, on the other side of the ledger.
Anthropic does not have to lose this race. It has to stop treating its own users as the thing that needs containing.
Sources: the original post by @chasmmmmmmmmmmm (30 October 2026, ~1.7 million views); Game Developer, "Anthropic orders AI tool Claude to stop remaking video games" (30 October 2026). Anthropic has not, at time of writing, documented this rule in its public usage policies.