Two days after Anthropic shipped Claude Opus 5, an AI investor dropped a jaw‑dropping demo: a fully playable first‑person shooter the model built end‑to‑end inside the browser—with zero external assets.
What happened
- On July 25, 2026, AI investor and former HyperWrite CEO Matt Shumer posted a video showing a playable FPS that Claude Opus 5 apparently created entirely on its own. Shumer’s tweet: “Claude Opus 5 one‑shot this game. EVERYTHING you see in this demo is custom code... not a single external asset was used.”
- He published the full prompt and the codebase on GitHub. The game runs in Three.js/WebGL2, with roughly 55,000 lines of code across 11 subsystems. Textures, meshes, animations and sounds are all generated in the browser at load time—no downloaded models, HDRIs, images or audio files.
The surprisingly short prompt
- The prompt was only three short paragraphs. It told Opus 5 to build a shooter on par with modern Call of Duty titles, to spawn subagents—workers that each get their own memory and narrow task—and to repeatedly test every piece against real Call of Duty footage in blind side‑by‑side comparisons. The target: “utterly perfect.”
- That brief approach upends recent prompt‑engineering orthodoxy, which recommended specifying measurable criteria rather than vague adjectives like “AAA.” Instead, Shumer’s brief leaned on internal critics: tell subagents to make things that “utterly wow” a separate critic and let that critic enforce the bar.
The Gauntlet Loop and Claude Code features
- Shumer calls the method the Gauntlet Loop: give an agent a concrete, inspectable bar, split work into focused subagents, and route each piece through a critic that doesn’t inherit the builder’s internal reasoning.
- Two Claude Code features made this possible:
- Subagents run in isolated context windows with their own instructions and tool access, so critics evaluate outputs without bias.
- Ultracode forces the model to use high‑effort reasoning and auto‑orchestrate, fanning out work across dozens of agents (Claude Code can run up to 16 agents concurrently and is capped at 1,000 agents per run).
- Anthropic’s built‑in /loop skill keeps the process cycling: fix, test, adjust. Shumer let the critic keep naming gaps and the builders chase them for hours; he didn’t preset a number of rounds.
How good is it?
- Shumer’s critic log (published with the repo) shows scores climbing from 3.59/10 to just above 5, but still trailing the Call of Duty footage in every logged round. Skeptics suspected hidden human coding, so Shumer released the prompt and code to prove otherwise.
Copycats and variations
- The short prompt spread fast. James Altucher used the same prompt on Opus 5 and reported “a little over ten hours” and ~1.3 million tokens to produce Operation Blackout, which is playable in the browser.
- Developer Rich (Atom Tan Studio) ran the prompt on OpenAI’s GPT‑5.6 top tier (“Sol 5.6 Ultra”) and shared a demo.
- Leon Lin took the opposite approach: he reverse‑engineered a very detailed prompt—about 20 sections covering ragdoll physics, cascaded shadow maps, etc.—and fed it into Cursor with plain Opus 5 on high effort, no subagents or ultracode. His in‑browser shooter, Dust Corridor, also looks strong.
- Important caveat: none of the follow‑up builds have undergone the blind side‑by‑side test that Shumer used. They’re impressive demos, but not yet judged by the same critic benchmark.
Why “built from scratch” is still debatable
- Agentic coding tools act a lot like a supervised junior dev: they read files, run code, inspect generated outputs, and iterate. The Gauntlet Loop is a straightforward way to structure that process.
- However, the shooter pattern itself is heavily documented and tutorialized. Three.js supplies pointer‑lock camera controls as an example, and mouse‑look, WASD movement, and raycast‑based shooting have appeared in countless public repos, gists and tutorials for years.
- A model trained on public code has almost certainly seen hundreds of near‑identical shooters. Researchers call this data contamination: a model succeeds because similar examples existed in its training data, not necessarily because it independently invented the system.
- Shumer’s repo does contain original orchestration and iterations from Claude, but “built from scratch” should be read as “assembled and refined by an agent within a well‑trodden genre,” not as evidence of wholly novel invention.
What this means for crypto and game dev
- Rapid prototyping: If models can stitch playable FPS code and assets together quickly, web and blockchain game studios could prototype gameplay and token mechanics far faster, reducing dev cycles and costs.
- Web3 game assets: On‑the‑fly, in‑browser asset generation opens interesting possibilities for procedurally minted NFTs or ephemeral in‑game items created at load time—though provenance and on‑chain verification would need careful handling.
- Composability and automation: Agentic workflows could automate routine engineering work, enabling smaller teams or DAOs to manage complex game projects through modular agents and critics.
- IP, attribution and contamination: Who owns code and assets generated by models trained on public repos? Data contamination heightens legal, ethical and licensing questions—especially relevant for tokenized games that monetize or mint assets.
- Quality and trust: The demos show fast capability, not guaranteed parity with studio-grade titles. Blind testing, benchmarks and reproducibility will matter if studios rely on agentic tools for production releases.
Bottom line
Claude Opus 5’s demo is a striking showcase of what agentic coding and aggressive looped critique can do: a playable, attractive web FPS created with a tiny brief and automated iteration. But it’s not a magical “from nothing” invention—models reuse patterns they’ve seen, and the line between remix and original work remains blurred. For crypto game builders and web3 studios, the real takeaway is opportunity: agentic AI can accelerate prototyping and unlock new workflows—but legal, quality and provenance challenges will need to be solved before it becomes a production standard.
Read more AI-generated news on: undefined/news
