Sumiro

The game

On November 4, 2026, developer Shikhar released Sumiro — an ink-wash samurai adventure that runs in the browser for free. A wanderer in a straw hat crosses misty cliffs, torii gates and stone stairways; climbing, riding lifts, working through chapters, with a 30–40 minute playtime. The launch thread passed 2.6 million views in two days.

His own line was: "Anyone can make games now." The Chinese write-up we're drawing from spends its length on what that line leaves out — the actual 24 hours, and the conditions behind them.

How the 24 hours were laid out

Six steps, six off-the-shelf tools (Opus 5.5, Flux 3 Max, GPT Image, Tripo, Blender, Three.js). Each step feeds the next:

1. Start with a video, not code. Flux 3 Max generates a short reference clip — this video is the project's art bible. World, character, palette, camera: everything later derives from it, following OdinLovis's Video2Game approach.
2. Break the video into scenes. Screenshots become assets rebuilt in Blender and placed into Three.js. The easy-to-miss detail: dynamic workflows inside Claude Code, with multiple agents building and checking each other in parallel — "a group of agents working separately, verifying each other," not one agent end to end.
3. Build the character. Four side views generated with GPT Image from a screenshot, Tripo turns them into a 3D model, Opus adds cloth physics and animation.
4. Play it first. The whole step is one sentence: play the prototype before adding more, confirm it actually runs. The only move in the pipeline that runs against AI instinct.
5. Let agents fill the levels. GPT Image generates concept art for each level in the locked art style, up to eight images per level; each concept becomes one playable 3D level.
6. Loop. Play, report bugs, let agents fix, play again.

What the hype line leaves out

The article insists on two preconditions before repeating "anyone can make games now":

24 hours is a third-game number, not a first-game number. Before Sumiro, Shikhar had already shipped Spiderbench (a browser Spider-Man game — that one took three days plus a 5x Claude subscription) and Walden, a katana duel he calls his first prototype. The hard-won lessons — aligning Blender assets with Three.js, keeping parallel agents from fighting, cloth physics not clipping — were all paid before this 24-hour clock started.

Two parts the pipeline doesn't cover.

Music is the one unsolved part: models can't hear, so a separate music-generation toolchain is required — the seventh step outside the six.

The other gap is controls: the demo video looks stiff because it was agent-operated and script-recorded. A human hands-on plays much better — but that isn't the part the 24 hours produced.

The two places copying it actually breaks

• Judging the reference video. Anyone can open the tools. The hard part is knowing whether the video is good enough — whether the style can carry a level, whether the character's silhouette is readable, whether the palette turns muddy in 3D. Pick wrong here and the other five steps are wasted.
• Step 4 is human-only. AI never tires of adding more — assets, concept art, levels. But playing the prototype before continuing is a judgment only a person can make; the author made it its own step because he had paid for skipping it.

Verdict

Read it as a compression story, not a democratization one. The scary number is real — 24 hours, six tools, one playable game — but it belongs to someone who had already built two games and stepped in every pit first. The tools raise the floor very high; the ceiling stays with whoever keeps practicing. Of the six steps, step 1's judgment is the one that can't be copied; the rest can be followed as written.

Attribution

Case assembled from the 挥码天涯 (大黑猫) Chinese write-up "用 AI 做游戏的完整流程", Shikhar's own launch thread, and third-party coverage. The 24-hour figure is the developer's own claim, echoed (not independently verified) by AIGamer and other outlets.

Images in this case come from the breakdown article and Shikhar's launch thread; copyright stays with the original publishers.