Day 2 of Claude Opus 5 "one-shotting" games

One day after Matt Shumer's Claude of Duty demo went viral, a second wave of Claude Opus 5 game builds spread across X — some replicating Shumer's exact technique, some remixing it, and at least one roundup thread explicitly trying to separate real builds from fakes.

Here's an inventory of what surfaced on July 27, ranked by how much of it can actually be checked rather than just watched.

The 7 demos roundup — with its own fake-detector built in

Newsletter writer Alex Prompter posted a thread titled “People are building playable games with Claude Opus 5” collecting seven builds from seven different accounts, each attached to a video rather than a bare claim.

These include

• a survivor-like (from a Chinese-language post benchmarking Opus 5 against Fable 5, GPT-5.6 Sol, and K3 on an identical prompt),
• Shumer's original FPS, a “painterly world” procedural-grass showcase that Alex Prompter himself flags as “more of an environment showcase than an actual game,”
• a submarine sidescroller from Pietro Schirano (claiming scratch-built assets, textures, and music),
• a snowboarding physics test,
• a 3D Pokémon clone claimed to come from a one-line prompt, and
• a Minecraft clone Alex Prompter built himself, said to fit inside a 196KB single HTML file.

The more useful part of the thread isn't the list — it's the disclaimer attached to it:

“Some of the Opus 5 game demos going around aren't builds at all. A few are screen recordings of games that already existed. Others are AI video generations of gameplay nobody can play. I traced every entry on this list back to a real build before including it.”

That's a direct admission that fake demos are already circulating in this cycle, which means every claim below still needs the same scrutiny Shumer's original got before its repo was checked.

A "24-hour" build that corrects itself in the replies

The highest-engagement follow-up came from @anshuc, who posted a game called The Long Silence with the framing “I ran Claude Opus 5 in one marathon session for 24 hours straight, and ended up with this game.” .

Unlike most of the roundup entries, this one ships with a live URL (longsilence.anshu.dev) and a public repo (github.com/achimala/TheLongSilence) rather than just a video.

It's also the one demo in this batch where the author walked back their own headline framing in a reply: “This took a few steps, not one-shot. But the process was simple and almost entirely hands-off.” The technique is described as an “adversarial subagent” approach explicitly built on Shumer's original Claude of Duty prompt structure — the same critic-loop pattern, not a new method.

Matt Shumer's own call for submissions

Shumer followed up by asking builders to link anything they'd made with his prompt. Replies surfaced a

• Dario Kart racer (video only, framed by its own author as possibly “the singularity moment @sama spoke of” — hyped enough to discount),
• a Mahabharata-themed battler called “Kuru of Duty” hosted at a real live domain (vedicgame.com/games/kuru-of-duty), and
• a build from Leon Lin, who explicitly didn't use Shumer's prompt at all — he reverse-engineered his own version from watching the video, ran it on “Opus 5 high” with no subagents or orchestration tooling, and published both the repo and the prompt he used.

That last one is the most useful data point in the batch: a comparable result from an independently written, simpler prompt is a stronger signal that Opus 5 has genuinely improved at this class of task than any number of people re-running Shumer's identical original prompt.

What's actually checkable, and what isn't

Ranked by how much can be independently verified without taking the poster's word for it:

• Kuru of Duty and The Long Silence both ship live, hosted URLs;
• The Long Silence and Leon Lin's build both ship public repos;
• Alex Prompter's Minecraft clone comes with a specific, falsifiable file-size claim

Everything else in this roundup — Dario Kart, the submarine shooter, the snowboard test, the 3D Pokémon clone, the painterly-world showcase — is video-only, which means it currently carries the same evidentiary weight Shumer's original demo did before anyone checked its repo: plausible, consistent with a now-repeated pattern, but not independently confirmed.

Why this matters more than any single demo

The interesting story on Day 2 isn't any individual build — it's that Shumer's specific technique (adversarial critic subagents, looped until a harsh reviewer stops finding faults) is now being copied, adapted, and reverse-engineered by independent people getting broadly comparable results.

That's a different and more interesting claim than “one demo went viral”: it suggests the method, not just the model, is doing real work — with the caveat that almost none of the replications have published the cost, wall-clock time, or session logs that were already missing from the original.

And the fact that a thread curator felt the need to publicly filter fake screen recordings and AI-generated video clips out of the same news cycle is its own small data point about how much noise this kind of demo now attracts.