The archive · AI & Models · Product decision · 2026
Claude of Duty bets one prompt plus AI agents can build a CoD-grade FPS; 3.4k stars
mshumer's single-prompt demo: a fleet of orchestrated AI agents wrote a 55k-line, asset-free browser FPS; critics scored it 5.05/10, GitHub shows 3.4k stars.
Claude of Duty
What the business is
Claude of Duty is an open-source browser first-person shooter built with Three.js r180 and WebGL2 — roughly 55,000 lines across 11 subsystems with every texture, mesh, animation and sound generated procedurally at load time — which doubles as a public benchmark of what orchestrated AI coding agents can ship from a single prompt.
How it started
mshumer created the repository on 2026-07-25 with the claim baked into the description: 'A Call of Duty-quality FPS in Three.js, built from a single prompt.' Agents worked under orchestration against ARCHITECTURE.md — subsystem interfaces, directory ownership, cross-subsystem event vocabulary and shared surface types — and produced roughly 55k lines across render, materials, sky, world, physics, player, weapons, fx, ai, ui and audio, with no art assets: every texture, mesh, animation and sound is generated procedurally at load time and the only runtime dependency is three.
What happened
Two measurement findings reshaped the build: a static-camera benchmark reported 94 fps while real gameplay ran 12–17 fps with 728–1236 ms stalls from WebGL programs compiling lazily mid-frame, and screenshot captures were not reproducible because state leaked between shots. The fix set — isolating each capture in a fresh page so imagediff could gate pixel identity, plus shader pre-warming — took worst frames to 66–82 ms and boot to 3.7–4.6 s while keeping output bit-identical. Process also mattered: three rounds of six parallel agents each owning a directory moved critic scores only +0.46 and left frame-ruining defects higher (60 → 47 → 66), while one sequential pass with a single owner per coupled concern moved +1.00 and cut defects to 26.
How it ended up
The maker's own adversarial assessment says the headline bet failed on the AAA bar: eleven independent critics scored frames 3.59 → 4.14 → 4.05 → 5.05 out of 10, two shots reached 'CLOSE' and the rest stayed 'AMATEUR', and in a blind A/B every critic in every round picked the real Call of Duty frame. What shipped is a playable, asset-free 55k-line browser FPS plus a reproducible agent harness, and the repo drew 3.4k stars and 514 forks by the 2026-09-04 crawl.
Background
Claude of Duty is a browser first-person shooter built with Three.js r180 and WebGL2 — roughly 55,000 lines across 11 subsystems, with every texture, mesh, animation and sound generated procedurally from code at load time. No art files exist; the only runtime dependency is three. The repository is by mshumer and was created on 2026-07-25.
The bet is baked into the description: 'A Call of Duty-quality FPS in Three.js, built from a single prompt.' A fleet of orchestrated AI agents worked against ARCHITECTURE.md — subsystem interfaces, directory ownership, cross-subsystem event vocabulary, shared surface types — to produce render, materials, sky, world, physics, player, weapons, fx, ai, ui and audio subsystems. The repo's own process notes are the interesting part: parallel fan-out made coupled rendering systems worse, while sequential single-owner passes fixed them.
The maker's honest assessment says the goal was not met: eleven adversarial critics scored frames from 3.59 up to 5.05 out of 10, two shots reached 'CLOSE' and the rest stayed 'AMATEUR', and in a blind A/B every critic in every round picked the real Call of Duty frame. Engineering claims are equally concrete — a benchmark that hid 12–17 fps gameplay behind 94 static fps was replaced by per-frame profiling, and an optimization pass cut worst frames from 728–1236 ms to 66–82 ms while keeping renders bit-identical.
Traction ran ahead of the AAA verdict: by the 2026-09-04 crawl the repo showed 3.4k stars (3,365 in crawl metadata), 514 forks and 36 watchers about six weeks after creation. No company, funding or revenue appears in the material, so Claude of Duty is best read as a solo, open-source capability demo whose public test of 'one prompt to AAA game' failed honestly — and became the product.
What has to be true
- Concrete traction: mshumer/Claude-of-Duty, created 2026-07-25, showed 3.4k stars (3,365 in crawl metadata), 514 forks and 36 watchers at the 2026-09-04 crawl.
- Falsifiable bet: the description promises a CoD-quality FPS from one prompt, and the README publishes the scores showing the bar was not met (5.05/10; real CoD wins every blind A/B).
- Documented engineering narrative: reproducible capture, per-frame profiling, and the parallel-versus-sequential agent process comparison all come with before/after numbers, not claims.
- Transferable process lesson: parallel agent fan-out made coupled systems worse (60 → 47 → 66) while sequential single-owner passes fixed them (66 → 26) — evidence for orchestrating AI coding agents.
What can be applied
Publish the benchmark and the failure: a demo that openly scores itself 5.05/10 against the real thing still earns 3.4k stars — the harness and the honesty become the product.
Aftermath
As of 2026-09-05 the repo is live at 3.4k stars (3,365 in crawl metadata), 514 forks and 36 watchers from the 2026-09-04 crawl. The README keeps the single-prompt origin while concluding the game does not match modern Call of Duty, and points readers to the author newsletter. No company, funding, revenue or pricing appears in the material, so Claude of Duty remains a solo open-source demo: the bet — one prompt, an agent fleet, the AAA bar — was tested in public and partially failed, and the reproducible harness became the more interesting result.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card