Coding agents ship playable game prototypes in Blender, Three.js, and Claude Code

• Agent-made games moved into playable links: marckohlbrugge's Red Alert clone had a live build and an untested multiplayer mode, while om_patel5's tank shooter rundown described 4v4 matchmaking, progression, bots, and lag compensation.
• The strongest workflows were structured agent runs: AmirMushich's Three.js prototype used Codex GPT-5.6 Sol, Three.js, Leon Lin's repo, and 13 dev phases, while chrisfirst's Fallout-style prompt explicitly fanned work out across subagents.
• Browser-native Three.js was the default canvas: MengTo's Opus 5 post pointed to shared source code and prompt, and MengTo's skills post open-sourced reusable Three.js game-dev skills.
• The polish gap stayed visible: AIandDesign's caveat said most 1-minute demos are shallow, while AIandDesign's Terminus Breach post framed polish as the line between viral demos and finished software.

The Claude of Duty README is the nerdiest artifact: a 55k-line Three.js r180/WebGL2 FPS with no external art, image, audio, or model assets, plus a brutal self-report on frame-rate stalls and critic scores. The related PROMPT.md asks for Vite 7, one bundled JavaScript payload, no game engine, no postprocessing helpers, Playwright checks, and zero runtime network requests. OpenAI's Codex model docs put GPT-5.6 Sol behind the default Power setting and say higher reasoning uses more tokens; Ultra uses subagents for split work. Claude Code's skills docs explain why the skills repos are popping up: a SKILL.md turns a repeatable procedure into a slash-invokable tool.

LLMJunky used Sol in Codex to recreate Rocket League's Octane inside Blender, then moved from vehicle modeling into arena design.

The reference workflow was explicit: LLMJunky's Blender pass says Sol used existing sketches and meshes, while the manual equivalent would have required mesh modeling, topology, symmetry, UVs, and hard-surface shading.

The map work followed the same pattern. LLMJunky's map update said the first pass was created by Codex rather than imported, and his Opus map test showed a separate map-design run.

LLMJunky also drew the IP boundary: his Epic Games IP reply called the project a frontier-capability test and said he would not sell it.

Browser-native games were the common shipping target, partly because a single link can carry the whole prototype.

The weekend builds clustered into a few repeatable shapes:

• RTS clone: marckohlbrugge shipped a Red Alert clone with a playable link, then his multiplayer follow-up added a second multiplayer link he said was not heavily tested.
• Arena shooter map: AmirMushich combined Codex GPT-5.6, Three.js, Leon Lin's repo, Quake 3 and Fortnite references, and 13 dev phases in one day; the linked PROMPT.md specified the browser FPS brief.
• Historical walkable world: levelsio said Opus 5 mostly one-shotted a 1660 New Amsterdam scene from historical maps, with visible defects still listed in the post.
• Liminal-space loop: AIandDesign's Backrooms update added a “descent” mode where the player finds an elevator on each level, while a later reply described the space as endless.
• Photography sim: om_patel5's photography sandbox post described a weekend Claude Code build with manual exposure, multiple camera bodies, interchangeable lenses, and a live histogram.

A Claude Code photography sandbox with manual camera controls.

chrisfirst's Fallout-style prompt shows the new prompt grammar for game builds: name the systems, fan out specialists, loop on each system, and assign a visual critic.

The prompt split the game into named workstreams:

• Terrain and environmental storytelling
• Pip-Boy interface
• Weapon and armor systems
• NPC dialogue trees
• Atompunk 1950s art direction
• 3D design and textures
• VATS slow-motion targeting
• NPCs and enemies
• A harsh visual-review subagent

According to chrisfirst's single-file note, the build took two or three short prompts plus “continue” commands after rate limits, and Claude produced much of the structure in a single HTML file.

The reusable layer is moving into skills. MengTo's Three.js skills post shared a free repo for an isometric action RPG with camera controls, VFX, audio, monster assets, and combat systems; a GitHub commit shows 18 new SKILL.md entries across camera controls, audio feedback, VFX, inventory, combat, assets, optimization, and playable-web-game testing.

emilkowalski put the skills workflow in plain terms: his skills note says people who already know what to build get the most from a skills repo, because they know what to ask for and can delegate the base layer.

The tank shooter was the clearest jump from visual prototype to systems design.

om_patel5 listed a surprisingly backend-shaped feature set:

• 4v4 rounds, one life per round, best of seven
• Tank upgrades before each match, plus salvage upgrades during play
• Six tanks: Panzer IV, Cromwell, Tiger 1, Sherman, T-34, and StuG III
• Desert, grass, and snow maps with destructible terrain
• Ballistic shells, direct hits only, and hit multiplier zones
• Matchmaking, lag compensation, and bot backfill
• Tank customization and XP unlock progression
• Browser entry with no download or install

The claim attached to the demo was blunt: om_patel5 said real-time multiplayer with matchmaking, lag compensation, and hit registration is the part people avoid because it is genuinely hard.

marckohlbrugge's Red Alert clone landed in the same lane at smaller scope, with a multiplayer build shared but not heavily tested.

The pushback came from creators making the same kind of work.

In AIandDesign's shallow-demo caveat, most demos only impressed inside a 1-minute video, while a game worth playing took longer. thekitze had the harsher version: his one-shot fatigue post asked people to make something worth playing.

AIandDesign's polished Godot port of Terminus Breach listed the delta: more enemies, better VFX and lighting, Midjourney-generated backdrops, better difficulty curves, easy and hardcore modes, and Mac, PC, and Android controller support AIandDesign's Terminus Breach post.

Long task latency was part of that gap. AIandDesign's procedural-generation note said changing an endless generated space meant traversing the generator, testing it, smoke testing it, trying it out, and checking the result; a related reply put some tasks around 10 minutes.

Claude of Duty supplied the cleanest non-Twitter postmortem. The README says a static-camera benchmark reported 94 fps while gameplay at Retina DPR ran 12 to 17 fps with 728 to 1236 ms stalls from lazy WebGL program compilation; after optimization, the shipped build reached 28 to 30 fps p50 and 14 to 17 fps p99.

That same README says adversarial critics scored frames 3.59 to 4.14 to 4.05 to 5.05 out of 10, with every critic still picking the real Call of Duty frame in blind A/B.

The process note mattered more than the score: the README says three rounds of six parallel agents moved score +0.46 and defects 60 to 47 to 66, while one sequential single-owner pass moved score +1.00 and cut defects 66 to 26.