All MicroEvals
Wingman game benchmark
Create MicroEval

Wingman game benchmark

Prompt

Build a browser-based wingsuit proximity-flying game in Three.js, set in the karst mountains of ancient China. One continuous map, 20 natural/architectural landmarks, extreme visual and physical realism, and a presentation strong enough that captured footage should look implausible for a browser game. Priority order: 1. Map/environment 2. Flight model 3. Player body 4. Sound Everything else supports those. Core constraints * Vite + latest Three.js r1xx + TypeScript + ES modules. * npm install && npm run dev must be enough to run. * Evaluate WebGPURenderer + TSL first in milestone 1, especially for terrain/cloud/vegetation compute. If unstable, switch explicitly to WebGL2 early. * Zero downloaded assets: no models, textures, HDRIs, fonts, or audio. Generate everything in code. Procedurally baked/generated textures are allowed. * Target: MacBook M5 Pro + Chrome, locked 60fps at Retina resolution, including clouds and high-speed proximity flight. * Performance is an architectural requirement from milestone 1. * Also support mobile touch controls in addition to mouse/keyboard and gamepad. Art direction Ancient Chinese karst country inspired by Wulingyuan/Huangshan: enormous vertical pillars, cliffs, caves, arches, cloud seas, mist, pine trees clinging sideways to rock, waterfalls, rivers, lakes, and old Chinese architecture. Architecture must be specifically Chinese: pagodas, cliff temples, upturned tiled eaves, moon gates, carved stone stairs, rope bridges, bronze bells, prayer ribbons, ruined gateways, pavilions. No pan-Asian generic styling. Target realism through atmosphere, scale, geology, lighting, clouds, haze and material response rather than pretending procedural assets equal photoscanned UE5 detail. Tune the world primarily for dawn: low sun cutting across a cloud sea, long pillar shadows, shafts through gorges, stacked atmospheric perspective. Rendering Implement correctly from the beginning: * Linear workflow, sRGB output, ACES filmic tone mapping, physically based lighting units. * Bruneton-style precomputed atmospheric scattering with Rayleigh/Mie, real sun direction, multiple scattering and distance-based aerial perspective. * Generate the sky/environment map in code and PMREM-filter it when sun conditions change. * True volumetric clouds: raymarched layered Worley/Perlin/fBm, Henyey-Greenstein phase, approximate multiple scattering, cloud entry/exit, terrain cloud shadows. Half-res + temporal reprojection. * Volumetric height fog and gorge light shafts. * Terrain must use SDF-based volumetric geometry, not a heightmap. Karst needs overhangs, caves, arches and undercuts. * Generate pillar fields from noise/geological rules, horizontal strata, erosion, SDF cave/arch subtraction, marching cubes or dual contouring, octree/chunk LOD. * Avoid visible terrain/noise seams and LOD cracks. Rock shading: * Triplanar procedural materials. * Geological strata. * Multi-octave detail normals. * Lichen/moss based on slope, moisture and cavities. * Water staining below seeps. * POM/detail treatment close-up. * Distance-blended tiling with no obvious repetition. Vegetation: * Huangshan pines on cliffs, bamboo valleys, scrub/crevice vegetation. * GPU instancing. * Hierarchical wind motion. * Proper LOD/impostors. Water: * Hybrid mesh/particle waterfalls. * Dense base mist with scattering. * Distant river. * Lake landing zone with refraction/depth color. Lighting/shadows: * Cascaded shadows, contact hardening, strong long pillar shadows. Post: * TAA mandatory. * GTAO. * Object/camera motion blur. * Physical-aperture DOF. * Hard-threshold bloom. * Subtle chromatic aberration, grain, vignette/lens character. * Code-generated final LUT. Flight model Simulate real wingsuit aerodynamics, tuned for playability but not faked: * Lift/drag from AoA. * Lift curve with stall beyond critical AoA. * Continuous body input changes wing surface area: spread = more lift/slower, tuck = faster/dive. * Energy management is the main mechanic: altitude ↔ speed ↔ flare/lift. * Skilled players must be able to regain altitude using energy/ridge/thermal behavior. * Glide ratio roughly 2.5–3:1 and realistic speeds. * Ridge lift, thermals, lee-side sink. * Cliff turbulence/rotor/gap turbulence. * Ground effect. * Parachute deployment with canopy inflation, strong deceleration and controllable descent/landing. No score, multiplier, timer, leaderboard, ghost, opponent or race. Proximity feedback is sensory: * Wind changes near terrain. * Camera/speed cues. * Motion blur. * Very subtle eased slow motion on closest passes. * Real collision/tumble with camera following the crash. * Fast, gentle respawn from a natural recent point with no penalty. Controls: * Mouse/keyboard. * Full gamepad with analog body control. * Mobile touch controls. * First-person default. * Third-person toggle, heavily tuned for shareable footage. Player body At speed, the body itself is a major realism element. Implement: * Wingsuit membrane cloth simulation using PBD or equivalent. * Cloth responds to actual airspeed/AoA/turbulence and visibly tensions, ripples and flutters. * Full-body IK. * Head stabilization and gaze leading turns/gaps. * Continuous arm/leg posture driven by flight input, not canned animation blends. * Turbulence buffeting through body and camera. * Landing sequence: flare, legs down, run-out/tumble depending on speed. Helmet Do not attempt a bare photoreal human face. Pilot wears detailed full-face helmet + tinted goggles: * PBR shell, wear/scuffs, simulated chin strap. * Real-time sky/terrain reflections in goggles. * Mist droplets, condensation and streaking. * Eyes visible through tint: blinking, micro-saccades, target/gap gaze. * Brow/jaw tension based on G/proximity. * Breath fog at altitude. * Airflow vibration. Smoke trail Ankle-mounted smoke canisters, inspired by real wingsuit athletes. Primary color: red. Alternate: white between flights. Must be a real system: * Smoke inherits canister velocity then interacts with airflow. * Curl/vortex effects from pilot. * Trails braid during rolls/banks. * Cliff turbulence shreds/deforms it. * Persistent for many seconds so previous flight line remains visible. * Expands, thins and dissipates naturally. * Participating-media volumetric rendering, not flat red sprites. * Light scattering, backlighting glow, self-shadowing. * Interaction with narrow gaps, waterfalls and clouds. * Faint nearby terrain shadows if performance permits. * Spark/puff on ignition, canister glow, ~60–90 sec burn duration. * Toggle while flying. * Burning hiss in audio. * Cameras/director should intentionally compose shots around the trail. Suit/livery Original sponsored-athlete identity, not copied from any real company. Design: * Deep navy + strong red. * Bold geometric wing panels. * Contrasting chest block and wing-edge trim. * Gloss matching helmet. * Invent original mountain-themed team name, symbol and wordmark. * Carry identity across wingsuit, helmet and parachute canopy. Map One seamless mountain descending from a summit temple to a lake. Straight descent: roughly 3–4 minutes. World must remain freely explorable: * Turn around. * Circle pillars. * Gain altitude in ridge lift/thermals. * Drift in cloud layers. * Multiple viable descent lines. * No “correct route.” Start from temple platform on highest pillar. Place 20 discoverable fly-through landmarks, never floating rings/checkpoints: 1. Stone moon gate 2. Twin-pillar gap 3. Open pagoda upper tier 4. Under rope bridge 5. Natural rock arch 6. Beneath cliff monastery supports 7. Behind waterfall curtain 8. Cave through a pillar 9. Ruined gateway columns 10. Under fallen tree across gorge 11. Bell tower 12. Narrow slot canyon 13. Gap between two ancient pines 14. Through temple main hall front-to-back 15. Under stone stair bridge 16. Gap in collapsed wall 17. Tighter second cave 18. Lakeside pavilion columns 19. Shoreline ceremonial gate 20. Final low pass under stone bridge Difficulty should vary from huge cinematic passages to genuinely tight proximity lines. Landmark feedback: * Audio compression/release. * Camera response. * Brief subtle time dilation. * No score/count popup. * Optional quiet journal records discovered landmarks. Modes: * Normal exploratory flight. * Free flight with no landmark structure. * Optional slow-drift/reduced-descent mode for scenic exploration. Restart: one input instantly returns to summit. No menu/loading/results interruption. Audio Entirely synthesized/generated with Web Audio and HRTF spatialization. Wind is the main sound: * Layered low body rumble + high-frequency edge whistle. * Driven by real airspeed and AoA. * Proximity alters wind via reflection/compression/rush/doppler. * Fabric flutter comes directly from cloth state. * Breathing responds to effort, tight passes and release afterward. * Waterfall distance roar. * Pine wind. * Distant birds. * Rare distant temple bell. * Cave/waterfall occlusion and reverb. * Canopy crack at deployment, then dramatically quieter parachute descent. Music: * Sparse guqin/xiao-style procedural/synthesized score. * Rare and quiet. * Disappears at speed. * Wind/silence should dominate. Cameras This game should generate footage worth watching. Press C to cycle cinematic rigs: * Helmet close-up drone. * Chase FPV-style drone with lag/lead/drift. * Wing cam looking back along body. * Landmark fly-by cameras with whip-pan and Doppler. * Long-lens landing-zone ground camera. * Free camera with position/focal-length/aperture controls. Every cinematic rig uses a physical camera model: * Sensor size. * Focal length. * Aperture. * Real DOF. * Autofocus/rack focus. * Subtle distortion, flare, vignette, lens dirt. * Mist/water droplets. * Mild handheld motion where appropriate. Automatic director: Observe flight events: * Landmark pass. * Minimum proximity. * Cloud entry/exit. * Peak speed. * Near miss. * Thermal/ridge climb. * Smoke ignition. * Hard turn with visible trail. * Calm scenic segments. Use editing rules: * Min/max shot duration. * Avoid repeating angles. * Cut on events, not afterward. * Hold close passes. * Allow calm shots to breathe. * Prefer smoke-visible compositions when useful. Opening sequence ~10 seconds: Pilot on summit temple, suit/ribbons moving in wind, push toward helmet, cloud sea reflected in goggles, held breath, step off, camera falls away as terrain drops beneath. Time dilation: * Manual slow-mo. * Automatic subtle eased proximity slow-mo. Replay: * Deterministic full-flight recording. * Timeline scrubbing. * Re-run any camera. * Auto-direct recorded flight. * HUD-free capture. HUD Minimal: * Airspeed * Altitude * Glide ratio No score/timer/objective marker. HUD fades when stable and returns when values change. One key hides it completely. The game must remain readable/playable HUD-off. Development workflow Use maximum parallelism. Define shared architecture/interfaces first, then fan work out to specialist agents/sub-agents: * Atmosphere/clouds * Terrain SDF/erosion/meshing * Rock/material shaders * Vegetation/water * Flight physics * Body/IK/cloth * Audio * Map/landmarks * Smoke/livery * Camera/helmet/goggles/director * Post-processing * Performance Split further if cloth, goggles, clouds, terrain, smoke or auto-director need dedicated owners. Do not optimize for token/time/agent cost. Optimize for final quality. Capture-first validation Build an automated capture harness early using Playwright/headless tooling where practical. It must be able to: * Load the game. * Fly deterministic scripted routes. * Capture Retina stills. * Capture frame sequences/video. * Test first-person and third-person/cinematic angles. Judge output, not code. Use three harsh critic agents: 1. Stills critic: composition, lighting, material response, atmosphere. 2. Motion critic: flight feel, cloth behavior, camera, speed. 3. Realism critic: hunts visible tells such as repeated textures, procedural-noise terrain, fake clouds, incorrect scattering, model-like body motion, temporal artifacts, clipping or LOD seams. Critics report concrete defects and locations, not vague approval. Iterate every subsystem against criticism until no specific major visual/physical defect remains. Runnable milestones 1. Renderer decision, color pipeline, atmosphere 2. SDF terrain + erosion + meshing/LOD 3. Flight physics on raw terrain 4. Rock/material shaders 5. Volumetric clouds/fog 6. Vegetation, water, architecture 7. Body, IK, cloth 8. Full map + 20 landmarks 9. Livery, helmet/goggles, volumetric smoke 10. Cinematic cameras + auto-director 11. Audio 12. Post/grade 13. Performance/final polish Every milestone must be runnable. Report quality honestly: if something is incomplete, unstable, visually weak, cracking, aliasing, too expensive, or only partially implemented, say so numerically/specifically instead of overstating it. Deliverable A clean repo with: * README * Clear architecture * Readable/commented shader code * No external assets * npm install && npm run dev * Desktop mouse/keyboard * Gamepad support * Mobile touch support * Stable 60fps target on MacBook M5 Pro/Chrome * Fully runnable summit-to-lake experience The first ten seconds after stepping from the summit into the cloud sea should be the project’s defining moment. Ultracode.

A system prompt was added to support web rendering

Drag to resize
Drag to resize
Drag to resize
Drag to resize

Response not available

Drag to resize