Claude Opus 5.5 self-scores coded 3D scenes in creator tests

• A self-scoring loop is the reusable trick: MengTo's ship test says Opus scored the ship, water, landscape, buildings, and smaller parts, then revised them toward an 8/10 bar within a token and time budget.
• The Japanese river-valley prompt treats a browser scene as a production build: MengTo's full brief requires a standalone page, autonomous navigation, weather, a day-night cycle, and browser-based QA.
• A playable sports-game test was an iterative build rather than a clean one-shot: LLMJunky's Rocket League test reported a two-hour initial pass, another 8 to 10 prompts, and imported car meshes and audio.
• The experiments are already moving into explorable worlds: DannyLimanseta's Poseidia tour describes an Atlantis experience made in about two hours, while petergyang's ride experiment turns a family-ride concept into a sequence of playable locations.

Anthropic introduced Opus 5.5 on September 22. MengTo's public brief bans text even on the loading screen; DannyLimanseta's Poseidia tour includes a trailer that its creator says Opus made too.

In MengTo's ship test, the creative direction is divided into five score cards: ship, water, landscape, buildings, and the components within each. The model keeps improving each area until it reaches at least 8/10, or hits the stated time and token limit.

MengTo called scoring important in a follow-up reply. It gives the run an explicit quality floor instead of leaving “more detail” as the only revision instruction.

Anthropic's Opus 5.5 workflow guide similarly frames long runs around a complete task, a definition of done, and a point where the model should stop and ask. MengTo had already applied the pattern to an Astra product demo, where MengTo's single-file product demo described a 186 KB HTML output and his reply said he had started using a score technique.

MengTo's full brief is a scene specification with production and verification requirements attached.

Its requirements break into five parts:

• World: river terrain, temples, bridge, pagoda, boat, villagers, forest, weather, and wildlife.
• Water and atmosphere: reflections, refraction, foam, wake, petals, mist, sunbeams, and a looping day-night cycle with a storm.
• Interaction: an autopiloted boat, keyboard takeover, collision response, chase camera, and a looping journey.
• Packaging: one browser-openable page with embedded assets where practical and no runtime CDN calls.
• QA: desktop-browser testing, screenshots in four lighting states, console inspection, and fixes for visible or technical errors.

MengTo described the riverboat demo as a playable Japanese landscape with dynamic weather, lighting, reflections, physics, scenery, and characters. He said in a Three.js reply that it was all Three.js, whose official manual reduces a renderable scene to a scene, camera, and renderer. In another reply, he said nothing was premade.

LLMJunky reported that Opus drove a browser-based Rocket League-style clone and its cinematic video, with imported Octane meshes and audio plus RocketSim-based physics. The creator put the initial pass at about two hours, followed by 8 to 10 prompts, and said the run used 3% of a weekly x20 allocation.

The public game brief makes the intended harness unusually legible:

• Three.js rendering, a Rapier physics engine, Vite, and separate rendering, physics, controls, state, UI, and assets.
• A controllable car, ball collisions, goals, reset logic, score, timer, boost, jumping, camera modes, and an optional bot.
• Browser verification for movement, collisions, goals, UI updates, camera behavior, console errors, and a nonblank canvas.

The interesting part is the division of labor: the prompt specifies systems, feel, and test cases, while the model is assigned the implementation and repeated tuning.

DannyLimanseta said Opus built an explorable Three.js tour of Poseidia, the last surviving island of Atlantis, in roughly two hours.

That was not literally one prompt. In a clarification, DannyLimanseta said it took fewer than 10 prompts, with most of the work done in one and the rest as tweaks, using vanilla Opus without skills.

Other reported formats show how wide the brief can get:

• petergyang asks in a five-use-case preview whether Opus can rival Astra on 3D scenes; the creator's Golden Gate Bridge test says the bridge video was generated entirely in code.
• petergyang's Soaring experiment spans the Alps, Northern Lights, pyramids, Great Wall, and Paris fireworks as a ride-like sequence.
• bilawalsidhu called a space-and-time reconstruction a first Opus 5.5 experiment, presenting the full journey as a 4D structure.

The hard parts remain specific:

• People and animals are “notoriously hard,” MengTo said after running out of tokens.
• More ambitious work takes longer, according to MengTo's reply.
• Character quality still needs tools such as Meshy or Tripo3D and patience, MengTo said.
• MengTo was optimizing character assets from 12 to 15 MB down to a 5 to 7 MB target in an asset-size reply.
• He put an earlier riverboat run at roughly $10 to $15 in a cost reply and said in another reply that he was on a 20x plan.