Coding agents build playable browser games in creator demos

• Opus 5 creative coding demos went playable fast: LLMJunky's Rocket Arena post linked a browser game and source code, then the multiplayer follow-up added room codes and controller support.
• The strongest 3D examples were all browser-native: chrisfirst's V8 demo exposed 112 clickable engine parts, while his Colosseum game came from a single Three.js prompt.
• Sol Ultra and Codex Sites were already enough for a Soulslike prototype: MengTo's browser game included shield blocking, dodge rolls, archery, and inventory in one afternoon.
• Grok 4.5 is showing up inside shipped game production: DannyLimanseta's progress post credits Cursor with redesigned enemies, cards, and VFX for a Steam roguelike deckbuilder.
• The most reusable workflow came from video-to-game handoff: kaigani's Seedance concept became a Gemini brief, ChatGPT mockups, and a Claude Code-built QLOBE Kids game.

Anthropic's Opus 5 launch post framed the model as an everyday coding workhorse at $5 per million input tokens and $25 per million output tokens. Claude's context-engineering post says Anthropic removed over 80% of the Claude Code system prompt for newer models. Grok Build's workflow post describes background jobs that fan tasks across hundreds of agents. The creator demos below are messier than launch copy, which is the useful part.

Anthropic shipped Opus 5 into Claude Code and the Claude Platform, and Claude Devs gave developers the API slug claude-opus-5 plus the Claude Code command /model claude-opus-5.

The same thread positioned Opus 5 around long-horizon coding: Claude Devs said it handles architecture, deployment, and design-quality UI across multi-stage work. Claude Devs also said Opus 5 defaults to high effort, with Fast mode running about 2.5x faster at 2x base price.

Day-zero hands-on reports added a workflow caveat. Dan Shipper said existing skills and plugins made Opus 5 argue with instructions and stop early, while deleting old skills and starting fresh improved results in his vibe check. Allie K. Miller also found the model "very eager" and moved from high reasoning effort to medium in her early-access note.

LLMJunky used Opus 5 to build a Rocket League-style browser game, then posted the playable version and code. His first demo said the run consumed 27% of a 5x Max subscription, and one reply put the rough cost at $3.

The implementation details were unusually specific for a social demo:

• LLMJunky said it was pure TypeScript.
• Another reply said it used no assets and was built in Claude Code.
• His prompt follow-up listed the iteration asks: practice mode, infinite boost, tighter steering, reversed steering controls, rounded ceiling corners, audio, orb boosts, and rounded goal interiors.
• A later clip showed Opus controlling the game to test mechanics.
• A reply put the build time at 1.5 hours.

The linked repo, am-will/rl-opus, documents Three.js rendering, Rapier physics, Vite, bots, synthetic audio, and a Cloudflare Worker for online play.

Multiplayer arrived hours later. LLMJunky's multiplayer test showed room code Q6B-F43, Xbox, PS5, and Steam controller support, while his follow-up blamed lag on free Cloudflare servers and closed the room after testers joined.

Chris First pushed Opus 5 into reference-tool territory with a browser V8 engine that exposes 112 clickable parts.

The prompt behind it asked for one self-contained HTML file, Three.js from a CDN, no external asset files, procedural geometry, a 90-degree cross-plane V8, OHV pushrod valvetrain, real slider-crank geometry, valve timing, firing-order indicators, selectable parts, and verification checks in the full prompt.

That prompt is creator gold because it specifies correctness targets before visuals:

• solve piston motion with slider-crank geometry;
• derive cylinder phase from firing order;
• animate valve lift from cam angle;
• expose cutaway, X-ray, exploded, and subsystem views;
• attach plain-language part explanations and telemetry to every named component.

Chris First then posted a second Opus 5 example: a Colosseum arcade fighting game from one prompt, with the player taking over another soldier after death in the demo and the prompt.

MengTo used Sol Ultra through Codex Sites to build a Dark Souls-inspired Three.js game that runs in the browser. His post says he handled the prompts, references, and libraries like img2threejs.

The mechanics list is the story:

• shield blocking;
• dodge rolls;
• archery;
• full inventory;
• stamina-based blocking, according to MengTo's reply;
• low-poly image-to-3D support, helped by img2threejs in another reply.

The demo page itself is hosted as a Codex Sites-style browser app, though the extracted page content is Cloudflare-gated. OpenAI's ChatGPT Work and Codex help page now frames Codex as the software-development surface inside ChatGPT Work.

Danny Limanseta used Grok 4.5 in Cursor for a production pass on Cook The Dungeon, a first-person roguelike deckbuilder with a live Steam page.

His week was not a one-prompt build. It was a structured redesign pass:

• upgraded dungeon visuals;
• redesigned 18 enemies with separate attack and hurt poses;
• redesigned all 86 cards;
• reworked combat VFX;
• used Grok to guide the card redesign card by card.

The art pipeline was separate. Danny said the game art is AI-made in one reply, generated through a custom tool hooked to the ChatGPT Images 2.0 API in another, then manually cleaned up in Photoshop when needed in a later reply.

Godot stayed under the hood. Danny said the game is built in Godot in a reply, and Grok 4.5 in Cursor can work on the Godot project via code without opening the editor in his toolchain note.

Kaigani posted a clean multimodel pipeline for turning a video concept into a playable web game. It starts with a category and one-line idea in Seedance 2, in this case "My Puppet Band" in the first post.

The handoff then becomes mechanical:

1. Gemini watches the generated video and writes a game design brief in the second step.
2. ChatGPT through Codex inspects key screens and creates screen mockups in the same step.
3. Claude Code receives the brief and screens, then builds the game into QLOBE Kids in the build step.
4. The result is playable at My Puppet Band in the final post.

This is the most portable workflow in the batch: video for taste, a vision model for specification, mockups for UI, and Claude Code for implementation.

Amir Mushich and youraipulse used GPT-5.6 with Codex and the Unreal Python API to generate a forest scene inside Unreal Engine.

Amir was explicit that it was iterative production, not a one-shot. His process note says they reviewed the viewport, described problems in normal language, and let Codex translate that feedback into technical changes.

The setup had three pieces in his breakdown:

• OpenAI Codex interpreted requests and wrote scripts;
• Unreal Engine held the environment, lighting, cameras, and animation;
• the Unreal Python API acted as the bridge.

Codex could create and position meshes, adjust scale and rotation, change lighting and fog, modify camera paths, edit Sequencer, save maps, and launch Movie Render Queue according to the tool list. The failure list was just as useful: cameras entered geometry, plants floated, exposure broke, the editor crashed after an unsafe map-copy operation, and the scene became too heavy for the GPU in Amir's failure note.

The final loop was: creative intention, natural-language instruction, Codex-generated Python, Unreal API execution, viewport review, human feedback, another iteration, Movie Render Queue, finished video in the pipeline post.

Emil Kowalski showed a smaller UI pattern that belongs in the same creative stack: ask a coding agent for multiple feature variants, add a switcher, and save the selection as a query parameter for sharing in his prototyping post.

He paired that with Leva controls for live interaction tuning, and Leva's GitHub repo describes it as a React-first GUI with a useControls hook and more than 12 input types.

Two adjacent workflow demos point at the same direction. Min Choi said Codex gained real-time voice mode for talking to coding agents, interrupting them, and sending them to work in his demo, while Greg Isenberg argued that cloud agents in VMs let creators run many Claude Code instances without local worktree collisions in his cloud-agent post.