Kits, terrain, and a new model nobody can get yet
Procedural Press — AI + Game Dev • Subscribe
On limits, models, and Astra
Fable 5.1 landed Sep 1 and a lot of people burned their weekly allocation within a day or two, almost always on large Blender jobs. Some numbers from developers willing to share them: one Blender experiment took 6% of a weekly Codex limit; one studio is running three Claude and two Codex accounts through a fundraising sprint; one popular "Fable one-shot" post was, by its author's own account, Opus plus four or five hours of iteration. The workarounds are mostly about routing: use 5.1 to write the plan and hand execution to Opus, or orchestrate long runs so they touch the expensive model rarely.
OpenAI's PlayCo prototyping post for GPT-6 Astra landed into that on Sep 3. Very few people have their hands on it yet, and the prevailing read wasn't that Astra is smarter. As one of them put it: "same intelligence, but cheaper/more available = wins." One caution: several of the "Astra AAA" clips circulating are Astra plus Blender plus Higgsfield, credited to Astra alone. Read the launch post rather than the timeline.
One veteran engineer offered an alternative to chasing budget: structure the project so a cheaper model can keep extending it. Concretely, from his own terminal engine: AI and human players both drive actors through a virtual controller, so swapping in multiplayer or pausing doesn't need special cases; the world clock is a component, so pausing and speed changes are one mechanism; visual effects go through a compositor that stacks and crops layers, instead of being written straight into the output buffer. His other advice was to complain about coupling in plain language when the model produces it, and to have the agent record a profiler trace rather than guess whether a slowdown is GPU or CPU. None of this is new engineering. It's what lets a session stay coherent past the first weekend.
Projects
Architecture Kit — @Lunarsong · repo · announcement
Lunarsong spent the back half of August building a modular medieval kit in Blender with Claude (pieces first shared on Aug 23, then an inn, then more building families), and on Aug 26 released the process itself as a reusable skill. You give it a style, it generates grid-snapping pieces, and a validator and critic pass check the output before it comes back to you. The goal was to turn a personal workflow into something other people could run with Claude or Grok. The earlier hand-run kit is also public.
Two days after the announcement, @SeloSlav ran the skill with Codex for a "Croatian hinterlands" medieval style. His first pass produced a tier-1 residence of 33 meshes and 3,136 triangles out of roughly 600 generated components, then a hunter's camp. The style was his, the recipe was someone else's, and it ran on a different model.
The validator/critic pattern is showing up elsewhere too, usually under the name gauntlet loop: give the agent a verifiable target (usually a reference image), have it submit renders to fresh critic subagents that score against the reference, and iterate, sometimes thirty rounds. It works for props, buildings and UI, where the critic has something concrete to compare against. It works less well for whole biomes or whole games, where the critic can't judge.
Selo Empire — @SeloSlav · repo · clip
The kit run above fed into this: a three.js medieval settlement sandbox in the Manor Lords mold. Procedural hills, forest and rivers; dirt roads that curve and wear; mills, quarries and burgage zones; SpacetimeDB underneath for economy and labor. He'd earlier posted a charcoal-etching overview map that zooms straight into the in-game terrain. On Sep 3 the repo went public alongside a clip pitching it as "7 years / 70 people → 7 days." The replies argued with the graphics, and he agreed with them: the art is rough, and the project is two months old.
Vibe3D, and the kits arriving on it — @alightinastorm · repo · announcement
The other way to ship a kit is as code. alightinastorm launched Vibe3D on Aug 13 as a source-first model registry for three.js: bunx vibe3d add @scifi-kit/pressure-gauge installs the TypeScript that generates the model into your project, not a GLB, so you or your agent can edit it like any other source file. It shipped with a 180-model sci-fi kit (MIT), two reference terrain meshes, and two skills: vibe-model, for authoring your own models from a reference photo, and vibe-terrain. The registry is open, meaning anyone can publish a kit to npm.
Two people have. @CRTkafa's medieval kit went up Aug 27 and hit v0.2.1 on Aug 31: 37 low-poly models sharing one library, 39,518 triangles for the entire kit, with a viewer where you can turn each model over and pull its sliders. Then thaikit landed Sep 1–3: 134 Thai street props built for a browser FPS meant to run on low-end PCs. Each prop is a procedural three.js factory with named parts and pivoted moving parts, and the repo includes the Claude Code skills that generate them.
Put this next to the Architecture Kit and you have two different answers to the same question. Lunarsong's kit generates Blender meshes with an agent and checks them; Vibe3D kits are hand-editable source you install. Both treat an asset as a recipe rather than a file.
He also released Super Terrain around Sep 1: a mesh-terrain editor modeled on Unreal 5.8's Mesh Terrain, running on three.js WebGPU with no WebGL fallback. The demo world is a sparse 4 km × 4 km terrain with sculpt layers, painted materials, live add/subtract CSG (dig a hole, then move the cave), five LODs and worker-based compilation. Forests are drawn as splines and grown from them. Desktop only. As of Friday he was still working on the trees and says they'll be open-sourced too.
Nova Worldforge, a 4000-part aircraft, and a three.js detour — @startracker
Three public artifacts in about a week. A modular aircraft of roughly 4,000 parts, Opus driving Blender, with the workflow in the thread. Then Worldforge, a terrain generator whose ocean erosion now produces islands, open-sourced with a three.js demo promised once biomes are done and it's optimized. Then a three.js level-design pass he described as "update 18."
The three.js choice is deliberate. He's using it for fast math and asset prototyping because agents write it quickly and there's no compile step, and he wants a bridge so that things built in three can be pulled into Unreal for cinematic shots. By Friday morning the status was: authored levels look decent, procedural level generation is still hard, and the model that gets closest runs out of budget before he can make much progress on it.
Nira — @rahul__gh0sh · niraanimates.com
Browser-based rigging and animation driven by an agent: drop in a GLB, rig it, author a shot, export it game-ready. On Aug 24 he added shot authoring. On Aug 25 he showed the agent generating motion and framing the shot, then feeding the output plus a character sheet to a video model to get an anime-style result. Next on his list: a video-provider API call built into the agent, MCP or bring-your-own-key so people can use their own agent, and video-based mocap if he can find an open-source approach that's both good enough and cheap enough to run (MediaPipe wasn't, HMR was too expensive).
The premise is the same as the kit work, applied to motion: encode the animation as rules and code rather than generating frames.
In-browser UFO horror set on a Montana farm, three.js, free to play now. Corn to haul to the silo, a glider off the peak, an inflatable boat, and a lamppost that goes out when the visitors get close. It's pre-alpha and aimed primarily at Steam. Early playtest feedback was a slow load and one to two frames per second on an M4 Pro Mac Mini in Chrome; voidmode had been testing on much larger machines and says he'll tune the LOD and quality auto-detect for lower specs. He's also researching downhill river and waterfall techniques, and plans to open-source the asset-generation harness he built to stop paying Meshy and Tripo. This week he ran a Fable 5.1 pass on the environment's photorealism and posted work-in-progress shots.
Hoops Redraft — @StrengthClub1 · hoopsredraft.com
Free in-browser NBA sim: pick any club from any season since 1976, redraft that year's real class, see if you can beat what the club actually did. Built in a few days with Claude Code, and it feels finished rather than demoed. He also has a drift-racing game and a boat-racing game in progress, neither public yet.
Tabula and xbox360.lol — @tag111111111
Two public dests from one person. Tabula is a desktop-playable game with hats and a level editor, already in the playable catalog, and he's actively looking for playtesters. Then on Sep 3, something else entirely: xbox360.lol, the Xbox 360 Blades and NXE dashboards decompiled and running in Chrome (clip). Not a game, but a public, working thing.
Final Abyss — Cal Woodford · play on Spawn
Gothic browser RPG with sixteen races, dungeons, loot and guilds. Free, no download. Added to the playable catalog this week.
Also public this week: @lofihashbloom's Shai-Hulud sandworm (Unity CLI, Blender MCP, Claude Code), which drew real playtest notes about hook placement and whether you can go under the sand. @cranberryclock's first playable slice of MM-MOW, a mowing MMO where the scythe is the legendary. Neither has a web dest yet.