How to Make a Browser Game With Claude: Yuji Horii’s 30-Minute Method

Last updated: October 14, 2026

Dragon Quest creator Yuji Horii built a playable browser game with Claude in about 30 minutes — and shared it free. On 11 October 2026, Horii posted Komiya in the Interrogation Room, a short mystery game made with Anthropic’s Claude, noting he no longer needs to draw or code to make a game himself. This guide reconstructs his approach as a recipe you can copy: the game shape that makes it possible, five copy-paste starter prompts, and an honest verdict on what “30 minutes” really means.

Quick answer: Yuji Horii built a browser game with Claude in 30 minutes. His method rebuilt: the right game shape, copy-paste starter prompts and a realistic time budget.

What did Yuji Horii actually build?

Horii made Komiya in the Interrogation Room — a free, short browser mystery game you play in one sitting. According to reporting by Kotaku and Inven Global on 12 October 2026, you play a police officer interrogating Komiya, a 72-year-old former locksmith suspected in a pawnshop safe robbery. Komiya lies constantly, so the core loop is catching contradictions: pick a line of questioning, spot the lie, and confront him with evidence. Two meters — Agitation and Resistance — rise and fall with your choices; if Resistance hits 100, Komiya stops answering entirely. Your completion rating depends on how many lies you expose.

The build details matter for what follows: the game was shared via a public Claude Artifact link, per Inven Global — meaning it was almost certainly built in the Claude app as a single-file HTML game, the format Claude’s artifact system generates. Horii said the build took him about 30 minutes, though as the outlet game-scanner noted, it is unclear exactly which tasks were included in that count.

“Now, with AI, even if you can’t draw, or you can’t program, you can create a game by yourself simply by asking the AI to do it for you. It’s incredible.” — Yuji Horii, on X, 11 October 2026 (translation via Kotaku)

The post resonated because of who wrote it: Horii recalled that on the original The Portopia Serial Murder Case (1983, Enix) he handled the story, the art and the programming alone. Forty-three years later, the solo-builder loop is back — this time the AI does the drawing and coding.

Why this kind of game is perfect for AI building

Horii picked the one game shape that AI currently builds best: a dialogue-and-state game in a single room. Think about what the interrogation format eliminates. No scrolling maps, no physics, no sprite animation, no save system, no networking. The entire game is: text on screen, buttons for questions, two numbers that go up and down, and a win/lose condition. That is exactly what a language model can generate reliably in one file — dialogue trees, simple state variables, and basic page layout.

If you want to copy the 30-minute result, copy the shape, not the setting. One-room scenarios that work the same way: a job interview where the candidate is hiding something, a border-control checkpoint where travellers have inconsistent stories, a doctor’s consultation where the patient won’t admit what they ate. The pattern is always the same: a liar, a set of questions, evidence that contradicts the lies, and a pressure system that punishes sloppy play.

The 30-minute recipe: Horii’s workflow rebuilt

Horii hasn’t published his exact prompts or workflow — so what follows is a faithful reconstruction of the only workflow that fits the facts. The game runs as a Claude Artifact (a single self-contained HTML file), it took about half an hour, and it centres on dialogue plus two meters. Here is that workflow as five steps you can repeat:

  1. Pick the one-room scenario. One character, one location, one secret. Write one sentence: who lies, about what, and what the player wants. Example: “A retired locksmith lies about a pawnshop safe robbery; the player must expose three lies before the lawyer arrives.”
  2. Define the win and fail states before anything else. Horii’s game has both: expose enough lies to win; let Resistance reach 100 and the suspect goes silent. Tell Claude these two outcomes in your very first prompt — AI-built games drift without a defined end.
  3. Script the liar, not the story. Give Claude the suspect’s persona, three concrete lies, and the evidence that disproves each one. Ask for a contradiction map: which question exposes which lie, and which evidence confronts it. This is the game’s design document, and it fits in one prompt.
  4. Add the pressure system. Two numbers, like Horii’s Agitation and Resistance meters. Specify exactly what moves them: “Tough questions raise Agitation by 10; presenting wrong evidence raises Resistance by 15; at Resistance 100 the suspect stops talking.” Numbers with rules beat vague “make it tense”.
  5. Polish, play, fix, publish. Run the game, find the first broken thing, describe the bug to Claude in one sentence, and reprompt. Then share it via Claude’s artifact share link — the same distribution Horii used.

Copy-paste starter prompts

Here is the prompt set that builds the whole game — five prompts, in order, each doing one job. Swap the bracketed details for your own scenario. These are original to this guide, written to mirror the structure of Horii’s game.

Prompt 1 — the skeleton: “Build a single-file HTML browser game: an interrogation-room mystery. I play a detective questioning [suspect name], a [age]-year-old [profession] suspected of [crime]. The game ends when I expose [number] lies (win) or the suspect’s Resistance meter reaches 100 and they stop talking (lose). Dark, minimal UI, works on mobile.”

Prompt 2 — the liar: “Now write the suspect’s persona and exactly three lies about [crime], plus one piece of evidence that disproves each lie. Build a contradiction map: which player questions can expose each lie, and which evidence confronts it.”

Prompt 3 — the pressure system: “Add two visible meters: Agitation and Resistance, starting at 20 each. Tough or repeated questions raise Agitation by 10; presenting the wrong evidence raises Resistance by 15; presenting correct evidence lowers Resistance by 10 and raises Agitation by 5. If Resistance hits 100, the suspect refuses to answer and the game ends in failure. Show both meters as bars at all times.”

Prompt 4 — the look: “Style it like a noir interrogation room: dark background, single lamp glow, the suspect as a simple silhouette, typewriter-style text for dialogue. No external images — CSS and emoji only, so it runs from one file.”

Prompt 5 — the fix loop: After playing once, paste this with your bug: “Playtest issue: [describe what broke in one sentence, e.g. ‘the Resistance meter never decreases’]. Fix it without changing anything else.”

“30 minutes” — what’s the realistic version?

Horii’s 30 minutes is real, but it is not the number a first-timer should plan around. Here is an honest budget for where the minutes go — and what the headline number likely leaves out.

Phase Horii (~30 min) Your first build
Scenario + rules (prompts 1–3) ~10 min ~30 min
Meters + UI polish (prompt 4) ~10 min ~30 min
Playtest + fix loop (prompt 5, ×2–3) ~10 min ~60 min
Total ~30 min ~2 hours

The gap is experience, not tools. Horii has shipped games for four decades — he knows exactly which rules a game needs, so his first prompt was probably precise. A beginner spends the extra time on iteration: the game soft-locks, a meter behaves oddly, the layout breaks on a phone. None of that is a Claude problem; it is the normal cost of learning what to ask for. Budget two hours, and you will finish with a playable game. The verdict: the 30-minute claim is credible for the build, but the real lesson is the form factor — pick a game this small and anyone can ship it in an evening.

Where Claude trips up on game builds (and the fixes)

Three failure modes show up in nearly every AI-built browser game — all fixable with one follow-up prompt each.

1. Broken state after a few turns. The game forgets a meter value or lets you win by spamming one button. Fix: prompt “After every player action, re-check all meter values and win/lose conditions before rendering. Add a small debug line showing the current state.” Then remove the debug line once it works.

2. Vibe-coded bloat. Ask for “more features” and Claude rewrites the file into something fragile. Fix: never ask for two changes at once. One prompt, one fix, test, repeat — the same loop from step 5 above.

3. Desktop-only layout. AI-generated game UIs often assume a wide screen. Fix: prompt “Make the layout responsive: stack the meters above the dialogue on screens under 600px wide, and make all buttons at least 48px tall for touch.”

Horii’s game sidesteps all three by staying small — which is precisely the point. The smaller the game, the fewer places for the AI to go wrong.

Frequently asked questions

Do I need to know how to code to build a game with Claude?

No — that is the entire point of Horii’s experiment. You describe the game in plain language and Claude writes the HTML, CSS and JavaScript. You will still need to read the game as a player and describe bugs clearly, but you never have to write or edit code yourself.

Is Claude free for building games like this?

Claude’s free tier includes artifact creation, but heavy back-and-forth iteration can hit usage limits. For a single small game built in one session, the free tier is usually enough; if you iterate for hours, you may need a paid plan. Check Anthropic’s current plan limits before a long build session.

How do I share the game once it’s built?

Horii shared his via a public Claude Artifact link — the Claude app lets you publish an artifact to a shareable link, so anyone with the link can play it in a browser. No hosting, no app store, no build step.

Can I sell a game made with Claude?

That depends on Anthropic’s current terms of service and usage policy, which govern commercial use of Claude outputs — check them before monetising anything. Ownership and licensing of AI-generated content is still evolving, so treat a first build as a portfolio piece, not a product, until you’ve read the terms.

Can I build this on my phone?

You can prompt Claude from the mobile app, but iterating on a game — reading the output, spotting layout bugs, testing taps — is much easier on a desktop browser. Prompt on the go if you like; do the playtest-and-fix loop at a computer.

Sources and methodology

Story facts verified 13 October 2026 against: Kotaku (12 Oct 2026 — Horii’s post, the 30-minute figure, the translated quote); Inven Global (12 Oct 2026 — Claude Artifact distribution, meter mechanics, lie-exposure scoring); VGC (12 Oct 2026 — launch coverage). Social buzz confirmed via Instagram/Threads posts, 12–13 Oct 2026. The workflow, prompt set and time budgets are this guide’s original reconstruction — Horii has not published his exact prompts — and were desk-designed, not hands-on tested.

Internal reading: our photo-to-video guide and Figma Agent guide cover more build-with-AI workflows in the same spirit.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!