Opus 5 Built a Homeworld-Style Space RTS From One Prompt — $632.65 Later

A week after Matt Shumer's Claude of Duty prompt went viral for one-shotting a browser FPS, someone remixed it for a different genre. On July 27, 2026, indie builder @mikeluan123 posted a 57-second clip of a Homeworld-style, universe-scale space RTS running in the browser — built with Claude Opus 5, Three.js, zero art assets, and a modified version of Shumer's now-famous "Claude of Duty" prompt template. The post crossed 90.9K views, split between people stunned by the visuals and at least one skeptic who did the math on the published token count and didn't buy it.

This is the second data point in what's becoming a pattern this cycle: take Shumer's "fan out sub-agents, /loop until a harsh critic is wowed, blind-compare against the real thing" prompt shape, swap the target genre, and see what a frontier model does unsupervised for hours. Here's what Luan actually posted, what the cost breakdown says, and where the "did this really cost that little" skepticism lands.

TL;DR

The prompt — Shumer's template, retargeted

Luan published the exact prompt he used, explicitly flagged as a remix:

I want you to build a homeworld style universe scale RTS at the level of most recent space RTS games. It should be utterly perfect, visually beautiful, stunning in detail, with every single thing done at AAA quality—from textures to physics to anything you could think of, with the sense of massiveness and scale.

Fan out sub-agents and have sub-agents tackle each one individually so that the game is utterly perfect. You should /loop on each item and have a separate sub-agent check it visually to ensure it looks triple A. That separate sub-agent should be a really harsh critic, and if it doesn't look triple A, it should keep going.

Don't stop until each sub-agent is utterly wowed with the quality when compared with the actual Homeworld game. It should literally compare them side by side blind and say which one looks better. Do this in ThreeJS. /loop until it's utterly perfect. Fan out sub-agents and ultracode.

Only need random mode, not mission, create multiple ship types, great UI, HUD, and game mechanics, and fluent control, max quality.

For ship models, and effects, AAA, max detail needed, use shaders, generative textures, and instancing to ensure best performance.

The structure is nearly identical to Claude of Duty's published prompt: fan-out sub-agents, a dedicated harsh-critic sub-agent, a blind side-by-side comparison against the real reference game, and /loop as the stop condition — not a fixed number of iterations. What changed is the target: FPS mechanics and CoD parity swapped for RTS scale, fleet-scale ship variety, and Homeworld's specific sense of massiveness. The closing paragraph — explicit instructions to use shaders, generative textures, and instancing — is the one addition tuned specifically for a space RTS's rendering load (hundreds of ships on screen) rather than an FPS's single-viewpoint scene.

This is the same loop engineering pattern explainx.ai has now tracked across three separate builders in July — Shumer's FPS, Shumer's voxel Manhattan, and now Luan's RTS: give the model an unreachable quality bar, a critic that can't be talked down, and a comparison target it has to beat blind, then let it run.

What the cost breakdown actually shows

Luan posted exact token figures alongside the clip:

The shape of these numbers tells its own story. Fresh input (68.6k) is tiny relative to output (4.6M) and especially cache-read (837.2M) — that ratio is consistent with a long /loop session where each new sub-agent turn re-reads a large, growing context (prior code, prior critic feedback, accumulated architecture decisions) rather than the model generating 4.6M tokens from a cold start. Prompt caching is what keeps that kind of run affordable: without it, re-reading hundreds of millions of tokens of accumulated context on every loop iteration would cost dramatically more than $632.65.

That's also exactly the number skeptic Samir Alibabic (@samiralibabic) pushed back on: "That's not a lot of tokens for something like that. I would expect billions. Something doesn't add up. A unplayable demo maybe?" Luan's own reply didn't dispute the possibility outright: "I'm shocked too, turns out we have great tooling already (web did heavy lifting). Highly suggest to try the prompt." That's a meaningfully different claim than "the numbers alone produced this" — it credits existing Three.js/web tooling and library maturity for absorbing work that would otherwise show up as more agent-generated tokens, the same "code is the asset pipeline" dynamic covered in explainx.ai's img2threejs coverage.

What's confirmed vs. what isn't

Unlike Claude of Duty, which shipped as a public MIT-licensed repo with a reproducible screenshot/imagediff harness, this demo is a 57-second clip and a token receipt, not open-source code:

• Confirmed: the video exists, the token/cost breakdown was published, the prompt text was published, the model was Opus 5, the engine was Three.js.
• Not independently verifiable: whether the RTS is actually playable end to end beyond the clip, whether "minor gameplay issues" means cosmetic bugs or broken core loops, and whether the published cost reflects the full build or a later, cheaper polish pass on top of unlogged earlier iteration.

That gap is worth naming plainly rather than passing along "utterly perfect" as fact — the same discipline explainx.ai applied to Claude of Duty's own honest README, which admitted its FPS didn't match real Call of Duty despite the same prompt family's instructions to keep looping "until utterly wowed."

Kimi K3 first, Opus 5 second — a cross-model retest

What makes this thread more than a single flashy clip is that Luan ran the same prompt family twice, on two different models, nine days apart:

• July 18, 2026: an earlier space RTS attempt using Kimi K3, which Luan summarized in Chinese as strong on creative/3D "vibe coding" but prone to overthinking, with a recommendation to hand debugging work to a different model and use Claude Code plus sub-agents to speed up the workflow.
• July 27, 2026: the Opus 5 rerun covered here, which Luan says produced noticeably more model detail. Asked directly whether GPT-5.6 Sol could do the same, Luan hadn't tested it but said 3D capability looked "observably strong" across Fable, Kimi, and Opus 5 as a general trend for current frontier and open-weight models, not a clean win for any single one.

That maps onto explainx.ai's own tracking of Kimi K3 as a model that's strong on creative and frontend-adjacent work but different in character from Fable- or Opus-class models on harder engineering tasks — Fireworks' own routing study found K3 and Fable specialize differently rather than one strictly beating the other. A single builder's side-by-side isn't a benchmark, but it's a useful anecdotal confirmation of a pattern showing up in more rigorous evals too.

What builders should take from this

1. The prompt template is now genre-portable. Shumer's FPS structure — goal, fan-out, harsh critic, blind comparison, /loop — transferred to an RTS with only the domain-specific closing paragraph (shaders, generative textures, instancing for fleet-scale rendering) changed. Expect more genre remixes.
2. Publish the token receipt, not just the clip. Luan's cost breakdown is what turned this from "cool video" into an actually debatable claim — it's also what let a skeptic ask a sharp, falsifiable question instead of just disbelieving in general.
3. Cache-read dominance is the tell for a long /loop run. If you're evaluating someone else's "cheap" AI-build claim, check whether cache-read tokens dwarf fresh input/output — that's the signature of an extended agentic session, not a single short completion.
4. "Utterly perfect" is a prompt instruction, not a verified result. Treat viral demo claims the way Claude of Duty's own README treated itself: state what's confirmed, flag what isn't, and don't let the marketing framing stand in for a repo you can actually run.
5. Cross-model reruns are more informative than single demos. Luan's Kimi K3 → Opus 5 retest is a more useful signal than either clip alone — if you're evaluating a model for creative 3D work, run the same prompt across at least two models before drawing conclusions.

Related on explainx.ai

• Did Opus 5 one-shot Call of Duty? Claude of Duty, the prompt this remixes
• Claude Opus 5 launch — benchmarks, charts, Fast mode
• Kimi K3 open weights: 2.8T params, day-0 hosting
• Kimi K3 + Fable 5 routing study — Fireworks AI
• Matt Shumer — GPT-5.6 Sol voxel Manhattan prompting guide
• Opus 5 Rocket League clone — then Opus played it
• Loop engineering — design loops that run while you sleep
• img2threejs — procedural photo to Three.js
• Top 10 Claude Opus 5 use cases

Primary source: @mikeluan123 on X, July 27, 2026 — video, prompt text, and token/cost breakdown as posted.

Cost, token, and prompt details reflect @mikeluan123's July 27, 2026 X thread as published. No public repository or standalone playable build was linked at the time of writing — treat gameplay and completeness claims as the builder's own characterization until independently verified.