Output B, 1 turnECEvan ChaOct 14Build a single-file interactive WebGL toy called "Jelly Bum": a translucent gummy-candy sculpture of a buttocks-and-hips torso (cut flat at the waist and mid-thigh, like a mannequin) wearing a string bikini bottom. It stands on a layered cake board. It wobbles like real jelly when you poke it, pull it or shake the table. TECH - One self-contained HTML file. three.js r160 as an ES module from https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js. No addons, no build step, WebGL 2 only, with a friendly fallback message if WebGL 2 is missing. - Write a custom orbit camera with damping and inertia. Drag on the background to rotate, use the wheel or pinch to zoom. Clamp it so the camera never goes below the board. Arrow keys also rotate. - Build a custom studio environment: a mid-grey gradient dome plus a few bright emissive softbox planes, run through PMREMGenerator, so reflections stay contrasty and the jelly reads as saturated candy. Use ACES tone mapping. The scene background must match the page's CSS background exactly. No shadow maps. SHAPE - Define the body as a signed distance field: smooth-union ellipsoids for the waist, hips and two round cheeks (with a visible cleft), plus two tapered round-cone thighs with a gap between them. The top is cut flat with rounded edges and the bottom sits flat on the board. - Mesh it with surface nets (about 0.0165 grid step, roughly 28k vertices). Snap vertices onto the surface, take normals from the SDF gradient, and drop the bottom cap. SOFT-BODY PHYSICS - Use a spring lattice of about 100 nodes (0.19 spacing) inside the body. The bottom layer is glued to the board. Each node has an anchor spring to its rest position plus springs to its neighbours (only where the midpoint is inside the body). Integrate with semi-implicit Euler in about 10 substeps per frame. The wobble should be visible and decay over about 2–3 s. - Deform the mesh on the GPU. Put node rest positions and displacements in a float DataTexture. In the vertex shader (injected via onBeforeCompile into MeshPhysicalMaterial), blend node displacements with normalized Gaussian weights. Compute the analytic Jacobian so normals stay correct. Smoothly glue the deformation to zero at the base. - Poke (tap or click, or Enter on the focused canvas): apply a strong Gaussian velocity impulse into the surface at the hit point. - Pull (drag on the jelly): spring the nearby nodes toward the cursor's offset on a camera-facing plane. Soft-clamp the offset to about 0.5 units. On release the jelly springs back and wobbles. - Picking: sphere-trace the SDF in the deformed space (sample the field at point p - displacement(p)). Show a grab cursor on hover. - Shake the table: a "Shake the table" button plus the Space key. - All scene objects except the lights live in one stage group, and the shake moves that group. - The motion is a quick hop (about 0.05 units, 0.14 s), then a decaying side-to-side rattle (about 4.2 Hz, amplitude about 0.08, about 1.7 s) in a random mostly-sideways direction. - Feed the board's acceleration (finite differences per substep) into every free node as an inertial force, so the jelly lags behind, sways and keeps wobbling after the board stops. - A "Jelly liveliness" meter shows kinetic plus elastic energy as a percentage. LOOK - Jelly: MeshPhysicalMaterial with transmission 1, thickness about 0.85, IOR 1.36, a saturated attenuation colour (distance about 0.75), roughness about 0.07, clearcoat, a light sheen, a faint emissive glow and DoubleSide, so the back faces refract. About 110 tiny white bubbles are suspended inside and move with the deformation. - Flavours: Lagoon (turquoise), Raspberry (magenta-pink), Lime and Mango. Switching flavour animates the colours smoothly and gives the jelly a small bounce. - Bikini: the same mesh pushed out slightly along the normal. The shape is cut out procedurally in the fragment shader from rest-space coordinates: - A cheeky back panel, high-cut at the sides. - A small front triangle and a gusset. - Thin side strings with a dark trim binding. - Small 3D knots with two dangling cord ends on each hip, deformed with the jelly. - Edges anti-aliased with alpha-to-coverage. - Six procedural prints: Leopard rosettes, Zebra wavy stripes, Polka dots, Breton stripes, Gingham and Classic black. The trim and cords match each print. - Cake board: rounded-rectangle extrusions in three layers (cream frosting top, caramel stripe, chocolate biscuit base). Shift it slightly backward so it sits under the cheeks. - Coloured shadow: planar-project the deformed jelly onto the board top and the floor along the key-light direction, tinted with the flavour colour, using multiply blending and a stencil so it never double-darkens. The shadow follows the stage during the shake. Add soft contact darkening under the feet. - The jelly drops in slightly squashed on load, springs into shape, and the camera eases in. - No candies or balls anywhere. UI (English only) - Top-left: a small mono eyebrow "Confectionery physics · No. 7", a big italic serif title "Jelly Bum." with the full stop in the current flavour colour, and a short italic lede: "Gummy curves, skinny strings and a wobble that always comes back." - Bottom-left: "How to play — Poke the jelly. Pull a curve. Shake the table." with the hint "Space shakes it too. Drag the background to turn it, scroll or pinch to zoom." - Right side: a floating rounded control card containing: - A "WebGL 2 · live" badge with a pulsing dot in the flavour colour, and a Sound toggle (WebAudio boing on poke and release, a low thump plus boing on shake; off by default). - "The jelly": a dark "Shake the table" button with a SPACE key hint, plus the liveliness meter. - "Flavour": four candy-sphere swatches. - "Bikini": six pattern swatches. - Sliders for Firmness and Wobble damping. - Fonts: Literata italic (display), Onest (UI) and JetBrains Mono (labels), all from Google Fonts. - Light and dark themes through CSS tokens; the 3D background follows the theme. - On phones the card becomes a scrollable bottom sheet, and the camera reframes the scene into the free area using setViewOffset. - Respect prefers-reduced-motion, keep a visible keyboard focus state, and use touch-action: none on the canvas. · 2LLMClaude Opus 5.5Gen AI—Skill—Software—6.5M tokens·30m 19s00
Output A, 1 turnECEvan ChaOct 14Build a single-file interactive WebGL toy called "Jelly Bum": a translucent gummy-candy sculpture of a buttocks-and-hips torso (cut flat at the waist and mid-thigh, like a mannequin) wearing a string bikini bottom. It stands on a layered cake board. It wobbles like real jelly when you poke it, pull it or shake the table. TECH - One self-contained HTML file. three.js r160 as an ES module from https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js. No addons, no build step, WebGL 2 only, with a friendly fallback message if WebGL 2 is missing. - Write a custom orbit camera with damping and inertia. Drag on the background to rotate, use the wheel or pinch to zoom. Clamp it so the camera never goes below the board. Arrow keys also rotate. - Build a custom studio environment: a mid-grey gradient dome plus a few bright emissive softbox planes, run through PMREMGenerator, so reflections stay contrasty and the jelly reads as saturated candy. Use ACES tone mapping. The scene background must match the page's CSS background exactly. No shadow maps. SHAPE - Define the body as a signed distance field: smooth-union ellipsoids for the waist, hips and two round cheeks (with a visible cleft), plus two tapered round-cone thighs with a gap between them. The top is cut flat with rounded edges and the bottom sits flat on the board. - Mesh it with surface nets (about 0.0165 grid step, roughly 28k vertices). Snap vertices onto the surface, take normals from the SDF gradient, and drop the bottom cap. SOFT-BODY PHYSICS - Use a spring lattice of about 100 nodes (0.19 spacing) inside the body. The bottom layer is glued to the board. Each node has an anchor spring to its rest position plus springs to its neighbours (only where the midpoint is inside the body). Integrate with semi-implicit Euler in about 10 substeps per frame. The wobble should be visible and decay over about 2–3 s. - Deform the mesh on the GPU. Put node rest positions and displacements in a float DataTexture. In the vertex shader (injected via onBeforeCompile into MeshPhysicalMaterial), blend node displacements with normalized Gaussian weights. Compute the analytic Jacobian so normals stay correct. Smoothly glue the deformation to zero at the base. - Poke (tap or click, or Enter on the focused canvas): apply a strong Gaussian velocity impulse into the surface at the hit point. - Pull (drag on the jelly): spring the nearby nodes toward the cursor's offset on a camera-facing plane. Soft-clamp the offset to about 0.5 units. On release the jelly springs back and wobbles. - Picking: sphere-trace the SDF in the deformed space (sample the field at point p - displacement(p)). Show a grab cursor on hover. - Shake the table: a "Shake the table" button plus the Space key. - All scene objects except the lights live in one stage group, and the shake moves that group. - The motion is a quick hop (about 0.05 units, 0.14 s), then a decaying side-to-side rattle (about 4.2 Hz, amplitude about 0.08, about 1.7 s) in a random mostly-sideways direction. - Feed the board's acceleration (finite differences per substep) into every free node as an inertial force, so the jelly lags behind, sways and keeps wobbling after the board stops. - A "Jelly liveliness" meter shows kinetic plus elastic energy as a percentage. LOOK - Jelly: MeshPhysicalMaterial with transmission 1, thickness about 0.85, IOR 1.36, a saturated attenuation colour (distance about 0.75), roughness about 0.07, clearcoat, a light sheen, a faint emissive glow and DoubleSide, so the back faces refract. About 110 tiny white bubbles are suspended inside and move with the deformation. - Flavours: Lagoon (turquoise), Raspberry (magenta-pink), Lime and Mango. Switching flavour animates the colours smoothly and gives the jelly a small bounce. - Bikini: the same mesh pushed out slightly along the normal. The shape is cut out procedurally in the fragment shader from rest-space coordinates: - A cheeky back panel, high-cut at the sides. - A small front triangle and a gusset. - Thin side strings with a dark trim binding. - Small 3D knots with two dangling cord ends on each hip, deformed with the jelly. - Edges anti-aliased with alpha-to-coverage. - Six procedural prints: Leopard rosettes, Zebra wavy stripes, Polka dots, Breton stripes, Gingham and Classic black. The trim and cords match each print. - Cake board: rounded-rectangle extrusions in three layers (cream frosting top, caramel stripe, chocolate biscuit base). Shift it slightly backward so it sits under the cheeks. - Coloured shadow: planar-project the deformed jelly onto the board top and the floor along the key-light direction, tinted with the flavour colour, using multiply blending and a stencil so it never double-darkens. The shadow follows the stage during the shake. Add soft contact darkening under the feet. - The jelly drops in slightly squashed on load, springs into shape, and the camera eases in. - No candies or balls anywhere. UI (English only) - Top-left: a small mono eyebrow "Confectionery physics · No. 7", a big italic serif title "Jelly Bum." with the full stop in the current flavour colour, and a short italic lede: "Gummy curves, skinny strings and a wobble that always comes back." - Bottom-left: "How to play — Poke the jelly. Pull a curve. Shake the table." with the hint "Space shakes it too. Drag the background to turn it, scroll or pinch to zoom." - Right side: a floating rounded control card containing: - A "WebGL 2 · live" badge with a pulsing dot in the flavour colour, and a Sound toggle (WebAudio boing on poke and release, a low thump plus boing on shake; off by default). - "The jelly": a dark "Shake the table" button with a SPACE key hint, plus the liveliness meter. - "Flavour": four candy-sphere swatches. - "Bikini": six pattern swatches. - Sliders for Firmness and Wobble damping. - Fonts: Literata italic (display), Onest (UI) and JetBrains Mono (labels), all from Google Fonts. - Light and dark themes through CSS tokens; the 3D background follows the theme. - On phones the card becomes a scrollable bottom sheet, and the camera reframes the scene into the free area using setViewOffset. - Respect prefers-reduced-motion, keep a visible keyboard focus state, and use touch-action: none on the canvas.LLMGemini 4 ArgonGen AI—Skill—SoftwareGoogle Chrome+11.8M tokens·20m 3s00
Output A, 2 turnsECEvan ChaOct 13Model and render a single 3D object with as much detail and realism as you can: an antique gold mechanical pocket watch, open, lying at a slight angle. Show the enamel dial with numerals, minute track, hands and sub-seconds dial, the crown and bow, the hinged lid with engraving, and a see-through caseback or cutaway revealing the movement: gear trains with real teeth, a balance wheel with hairspring, jewels, screws, bridges with perlage/Geneva striping. The gears and balance wheel should animate plausibly. Everything procedural in code (no downloaded models or textures). Prioritise geometric detail, physically based materials (polished gold, brushed steel, ruby jewels, glass crystal with reflections) and studio lighting with an environment map, soft shadows and tasteful post-processing. Hero framing: it should look like a luxury product shot. Output a single HTML page using three.js. No external assets, model everything manually. Inspect your renders closely and keep raising the level of detail until you are genuinely proud of it. Do 2-3 iterations of this, then stop.LLMClaude Opus 5.5Gen AI—Skill—Software—11.5M tokens·1h 42m 50s00
Output B, 1 turnECEvan ChaOct 13Model and render a single 3D object with as much detail and realism as you can: an antique gold mechanical pocket watch, open, lying at a slight angle. Show the enamel dial with numerals, minute track, hands and sub-seconds dial, the crown and bow, the hinged lid with engraving, and a see-through caseback or cutaway revealing the movement: gear trains with real teeth, a balance wheel with hairspring, jewels, screws, bridges with perlage/Geneva striping. The gears and balance wheel should animate plausibly. Everything procedural in code (no downloaded models or textures). Prioritise geometric detail, physically based materials (polished gold, brushed steel, ruby jewels, glass crystal with reflections) and studio lighting with an environment map, soft shadows and tasteful post-processing. Hero framing: it should look like a luxury product shot. Output a single HTML page using three.js. No external assets, model everything manually. Inspect your renders closely and keep raising the level of detail until you are genuinely proud of it. Do 2-3 iterations of this, then stop.LLMMuse WonderGen AI—Skill—Software—243.9K tokens·6m 3s00
1 turnECEvan ChaOct 13Model and render a single 3D object with as much detail and realism as you can: an antique gold mechanical pocket watch, open, lying at a slight angle. Show the enamel dial with numerals, minute track, hands and sub-seconds dial, the crown and bow, the hinged lid with engraving, and a see-through caseback or cutaway revealing the movement: gear trains with real teeth, a balance wheel with hairspring, jewels, screws, bridges with perlage/Geneva striping. The gears and balance wheel should animate plausibly. Everything procedural in code (no downloaded models or textures). Prioritise geometric detail, physically based materials (polished gold, brushed steel, ruby jewels, glass crystal with reflections) and studio lighting with an environment map, soft shadows and tasteful post-processing. Hero framing: it should look like a luxury product shot. Output a single HTML page using three.js. No external assets, model everything manually. Inspect your renders closely and keep raising the level of detail until you are genuinely proud of it. Do 2-3 iterations of this, then stop.LLMMuse WonderGen AI—Skill—Software—222K tokens·8m 12s00
Output D, 1 turnECEvan ChaOct 8make a demo video about https://evals.ag/ In the video, all the reference example vid must be 3d game, 2d game, videos like hooking in 3 seconds. follow the demo video skill! · 4LLMClaude Opus 5.5Gen AIEleven v3+1Skill/demo-video-creatorSoftwareHiggsfield+140.3M tokens·1h 53m 30s00
Output C, 1 turnECEvan ChaOct 8make a demo video about https://evals.ag/ In the video, all the reference example vid must be 3d game, 2d game, videos like hooking in 3 seconds. follow the demo video skill!LLMClaude Opus 5.5Gen AIEleven v3+1Skill/demo-video-creatorSoftwareHiggsfield+139.7M tokens·1h 9m 49s00
Output B, 1 turnECEvan ChaOct 8make a demo video about https://evals.ag/ In the video, all the reference example vid must be 3d game, 2d game, videos like hooking in 3 seconds. follow the demo video skill!LLMClaude Opus 5.5Gen AIEleven v3Skill/demo-video-creatorSoftwareHiggsfield+159.3M tokens·1h 45m 26s00
Output A, 1 turnECEvan ChaOct 8make a demo video about https://evals.ag/ In the video, all the reference example vid must be 3d game, 2d game, videos like hooking in 3 seconds. follow the demo video skill!LLMClaude Opus 5.5Gen AI—Skill/demo-video-creatorSoftwareHiggsfield39.4M tokens·1h 14m 43s00
Output A, 1 turnECEvan ChaOct 8Build a playable 3D hover-bike racing game in the browser (Three.js). Neon cyberpunk city at night, rain, glowing billboards, bloom. Race through traffic with boost pads, ramps and near-miss sparks; speedometer, lap timer and a dramatic crash + replay screen. Make it look like a premium console game trailer the moment it loads. · 4LLMGrok 4.7Gen AI—Skill—Software—3M tokens·5m 13s00
Output D, 1 turnECEvan ChaOct 8Build a playable 3D flying game in the browser (Three.js): pilot a paper plane through floating islands at golden hour. Soft volumetric clouds, waterfalls spilling off the islands, flocks of birds, glowing rings to fly through, wind trails, a cinematic intro fly-by and a score. Hand-painted storybook look, gorgeous from the first frame.LLMDeepSeek V4.1 FlashGen AIEleven MusicSkill—Softwarefal.ai16M tokens·47m 7s00
Output D, 1 turnECEvan ChaOct 8Build a playable 3D hover-bike racing game in the browser (Three.js). Neon cyberpunk city at night, rain, glowing billboards, bloom. Race through traffic with boost pads, ramps and near-miss sparks; speedometer, lap timer and a dramatic crash + replay screen. Make it look like a premium console game trailer the moment it loads.LLMDeepSeek V4.1 FlashGen AI—Skill—Software—18.7M tokens·1h 18m 3s00
Output C, 1 turnECEvan ChaOct 8Build a playable 3D hover-bike racing game in the browser (Three.js). Neon cyberpunk city at night, rain, glowing billboards, bloom. Race through traffic with boost pads, ramps and near-miss sparks; speedometer, lap timer and a dramatic crash + replay screen. Make it look like a premium console game trailer the moment it loads.LLMGPT-6 AstraGen AI—Skill—Software—277K tokens·12m 0s00
Output B, 1 turnECEvan ChaOct 8Build a playable 3D hover-bike racing game in the browser (Three.js). Neon cyberpunk city at night, rain, glowing billboards, bloom. Race through traffic with boost pads, ramps and near-miss sparks; speedometer, lap timer and a dramatic crash + replay screen. Make it look like a premium console game trailer the moment it loads.LLMClaude Opus 5.5Gen AI—Skill—Software—8.7M tokens·47m 41s00
Output D, 1 turnECEvan ChaOct 8Create a 12-second cinematic 3D product reveal film in the browser (Three.js) for a fictional smartwatch called ORBIT. Dark studio, dramatic rim lighting, slow camera orbit and dolly, liquid-metal reflections, a particle burst on the reveal, bold kinetic typography, logo end card. Auto-plays and loops like an Apple keynote launch film.LLMClaude Opus 5.5Gen AI—Skill—Software—2.4M tokens·20m 44s00
Output C, 1 turnECEvan ChaOct 8Create a 12-second cinematic 3D product reveal film in the browser (Three.js) for a fictional smartwatch called ORBIT. Dark studio, dramatic rim lighting, slow camera orbit and dolly, liquid-metal reflections, a particle burst on the reveal, bold kinetic typography, logo end card. Auto-plays and loops like an Apple keynote launch film.LLMGrok 4.7Gen AI—Skill—Software—6M tokens·1h 14m 22s00
Output B, 1 turnECEvan ChaOct 8Create a 12-second cinematic 3D product reveal film in the browser (Three.js) for a fictional smartwatch called ORBIT. Dark studio, dramatic rim lighting, slow camera orbit and dolly, liquid-metal reflections, a particle burst on the reveal, bold kinetic typography, logo end card. Auto-plays and loops like an Apple keynote launch film.LLMDeepSeek V4.1 FlashGen AI—Skill—Software—12.5M tokens·1h 0m 35s00
Output A, 1 turnECEvan ChaOct 8Create a 12-second cinematic 3D product reveal film in the browser (Three.js) for a fictional smartwatch called ORBIT. Dark studio, dramatic rim lighting, slow camera orbit and dolly, liquid-metal reflections, a particle burst on the reveal, bold kinetic typography, logo end card. Auto-plays and loops like an Apple keynote launch film.LLMGPT-6 AstraGen AI—Skill—Software—69.9K tokens·4m 7s00
Output D, 1 turnECEvan ChaOct 8Build a playable 3D monster-truck smash game in the browser (Three.js + physics). Drive a giant monster truck through a colourful toy city and smash everything: towers of blocks topple, cars flip, glass shatters into debris with real physics, slow-motion on big crashes, dust and sparks, a destruction score. Should feel explosive and satisfying from the first second.LLMClaude Opus 5.5Gen AI—Skill—Software—11.3M tokens·34m 26s00
Output C, 1 turnECEvan ChaOct 8Build a playable 3D monster-truck smash game in the browser (Three.js + physics). Drive a giant monster truck through a colourful toy city and smash everything: towers of blocks topple, cars flip, glass shatters into debris with real physics, slow-motion on big crashes, dust and sparks, a destruction score. Should feel explosive and satisfying from the first second.LLMDeepSeek V4.1 FlashGen AI—Skill—Software—20.4M tokens·1h 7m 21s00
Output B, 1 turnECEvan ChaOct 8Build a playable 3D monster-truck smash game in the browser (Three.js + physics). Drive a giant monster truck through a colourful toy city and smash everything: towers of blocks topple, cars flip, glass shatters into debris with real physics, slow-motion on big crashes, dust and sparks, a destruction score. Should feel explosive and satisfying from the first second.LLMGrok 4.7Gen AI—Skill—Software—5.4M tokens·38m 23s00
Output A, 1 turnECEvan ChaOct 8Build a playable 3D monster-truck smash game in the browser (Three.js + physics). Drive a giant monster truck through a colourful toy city and smash everything: towers of blocks topple, cars flip, glass shatters into debris with real physics, slow-motion on big crashes, dust and sparks, a destruction score. Should feel explosive and satisfying from the first second.LLMGPT-6 AstraGen AI—Skill—Software—261K tokens·11m 4s00
Output C, 1 turnECEvan ChaOct 8Build a playable 3D flying game in the browser (Three.js): pilot a paper plane through floating islands at golden hour. Soft volumetric clouds, waterfalls spilling off the islands, flocks of birds, glowing rings to fly through, wind trails, a cinematic intro fly-by and a score. Hand-painted storybook look, gorgeous from the first frame.LLMGrok 4.7Gen AIGPT Image 2Skill—SoftwareChatGPT Images5.9M tokens·33m 30s00
Output B, 1 turnECEvan ChaOct 8Build a playable 3D flying game in the browser (Three.js): pilot a paper plane through floating islands at golden hour. Soft volumetric clouds, waterfalls spilling off the islands, flocks of birds, glowing rings to fly through, wind trails, a cinematic intro fly-by and a score. Hand-painted storybook look, gorgeous from the first frame.LLMGPT-6 AstraGen AI—Skill—Software—89.1K tokens·5m 20s00
Output A, 1 turnECEvan ChaOct 8Build a playable 3D flying game in the browser (Three.js): pilot a paper plane through floating islands at golden hour. Soft volumetric clouds, waterfalls spilling off the islands, flocks of birds, glowing rings to fly through, wind trails, a cinematic intro fly-by and a score. Hand-painted storybook look, gorgeous from the first frame.LLMClaude Opus 5.5Gen AI—Skill—Software—3.7M tokens·26m 45s00
Output B, 1 turnECEvan ChaOct 7Make a ~60s 16:9 product demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. ## Story 1. Hook (first 3s must be instantly clear): real AI-made game/video outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." 2. A real game task is typed into the Evals composer: "Build a polished playable Sonic-style 3D platform game for the browser." → Battle → up to 4 setups (Model · Skill · Environment), each landing with real values and moving UI. 3. Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. 4. Blind compare: two moving outputs, "Which would you choose?" → "I prefer A". 5. Run record of that same output: LLM, skill, environment, turn time, tokens. 6. Setup leaderboard: "Setups, not just models." 7. Community wall of real game / video / shorts outputs. 8. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. ## Content - Show only games, videos, 3D scenes, motion pieces and shorts. No websites, no landing pages. - Real Evals UI only, never mocked up. Every name and number on screen is real; only model names that exist in Evals (e.g. Claude Opus 5.5, GPT-6 Astra). - No real-person likeness. ## Voice - One energetic, confident English male narrator: excited on the hook, curious on the question, emphatic on the tagline, quick through the UI steps. - ~170 words in one continuous take that fills the whole film. No pause longer than ~0.35s, no clipped words. Say "Evals" clearly as "E-vals". - The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. ## Music & sound - Modern minimal tech-launch track, 120 BPM: punchy kick, deep 808, crisp hats, glassy synth plucks. Apple / Linear keynote feel. No cheesy EDM risers, no corporate stock vibe. - Voice always leads. Music dips ~6dB only while the voice speaks, on a smooth envelope. - SFX sit just under the music and rise above it only on hits. - 4–6 big hits get a layered reverse swell + sub boom + transient + glitch. Light ticks and whooshes elsewhere, never on every cut. ## Look - One sans font, one accent colour. - Real UI as rounded cards (16–22px radius) with soft shadows. - Sections alternate dark and light. - Never: outlines or hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI. ## Motion: this is the feel - Premium, fast, confident: an Apple / Linear keynote cut at shorts speed. - One continuous camera with velocity-continuous easing. Two easing personalities only: snappy (~25% of remaining distance per frame) and floaty (~12%). - Nothing ever stops dead. Every hold keeps drifting 1–3%. Fast moves decelerate over at least 6 frames. The end card keeps pushing until the last frame. - Exits accelerate into the cut; the next shot carries that velocity through it. - Real whip-pans: both shots travel together over 0.3–0.45s, motion blur peaking mid-move. One whip direction throughout. - Overlapping action: the next element starts while the previous one is ~60% done. No start-stop chains. - Text is screen-locked and arrives together with its container. Words reveal one by one, 2–4 frames apart, and the newest word flashes the accent. In ≤0.35s, readable ≥0.7s, out ≤0.25s. - On busy footage the camera stays calm (slow push ≤3%); the energy comes from cutting on the beat. - Every cut lands on the beat (±1 frame), with the visual hit and the SFX on the same frame. No double cuts. - Big hits (4–6 only): +8.5% zoom punch, ~16px shake at ~22Hz settling in 0.1s, RGB split 16px → 7px over 4 frames, 1–2 frames of white flash. A lighter version on secondary cuts. ## Quality bar Before calling it done, watch it like a motion director who didn't build it: - no dead stops, frozen frames or stutters - every cut on the beat - every word audible - every name and number real Fix the 5 worst moments first.LLMGPT-6 AstraGen AI—Skill/hyperframesSoftware—331.9K tokens·6m 34s00
Output A, 1 turnECEvan ChaOct 7Make a ~60s 16:9 product demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. ## Story 1. Hook (first 3s must be instantly clear): real AI-made game/video outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." 2. A real game task is typed into the Evals composer: "Build a polished playable Sonic-style 3D platform game for the browser." → Battle → up to 4 setups (Model · Skill · Environment), each landing with real values and moving UI. 3. Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. 4. Blind compare: two moving outputs, "Which would you choose?" → "I prefer A". 5. Run record of that same output: LLM, skill, environment, turn time, tokens. 6. Setup leaderboard: "Setups, not just models." 7. Community wall of real game / video / shorts outputs. 8. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. ## Content - Show only games, videos, 3D scenes, motion pieces and shorts. No websites, no landing pages. - Real Evals UI only, never mocked up. Every name and number on screen is real; only model names that exist in Evals (e.g. Claude Opus 5.5, GPT-6 Astra). - No real-person likeness. ## Voice - One energetic, confident English male narrator: excited on the hook, curious on the question, emphatic on the tagline, quick through the UI steps. - ~170 words in one continuous take that fills the whole film. No pause longer than ~0.35s, no clipped words. Say "Evals" clearly as "E-vals". - The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. ## Music & sound - Modern minimal tech-launch track, 120 BPM: punchy kick, deep 808, crisp hats, glassy synth plucks. Apple / Linear keynote feel. No cheesy EDM risers, no corporate stock vibe. - Voice always leads. Music dips ~6dB only while the voice speaks, on a smooth envelope. - SFX sit just under the music and rise above it only on hits. - 4–6 big hits get a layered reverse swell + sub boom + transient + glitch. Light ticks and whooshes elsewhere, never on every cut. ## Look - One sans font, one accent colour. - Real UI as rounded cards (16–22px radius) with soft shadows. - Sections alternate dark and light. - Never: outlines or hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI. ## Motion: this is the feel - Premium, fast, confident: an Apple / Linear keynote cut at shorts speed. - One continuous camera with velocity-continuous easing. Two easing personalities only: snappy (~25% of remaining distance per frame) and floaty (~12%). - Nothing ever stops dead. Every hold keeps drifting 1–3%. Fast moves decelerate over at least 6 frames. The end card keeps pushing until the last frame. - Exits accelerate into the cut; the next shot carries that velocity through it. - Real whip-pans: both shots travel together over 0.3–0.45s, motion blur peaking mid-move. One whip direction throughout. - Overlapping action: the next element starts while the previous one is ~60% done. No start-stop chains. - Text is screen-locked and arrives together with its container. Words reveal one by one, 2–4 frames apart, and the newest word flashes the accent. In ≤0.35s, readable ≥0.7s, out ≤0.25s. - On busy footage the camera stays calm (slow push ≤3%); the energy comes from cutting on the beat. - Every cut lands on the beat (±1 frame), with the visual hit and the SFX on the same frame. No double cuts. - Big hits (4–6 only): +8.5% zoom punch, ~16px shake at ~22Hz settling in 0.1s, RGB split 16px → 7px over 4 frames, 1–2 frames of white flash. A lighter version on secondary cuts. ## Quality bar Before calling it done, watch it like a motion director who didn't build it: - no dead stops, frozen frames or stutters - every cut on the beat - every word audible - every name and number real Fix the 5 worst moments first.LLMGrok 4.7Gen AIEleven v3+2Skill/hyperframesSoftwarefal.ai6.9M tokens·37m 4s00
Output A, 1 turnECEvan ChaOct 7https://www.instagram.com/reel/Dd4VIgTq8e7/ reference this video! you must reference the asset of the video. music, sound effect etc... and what u need to describe is https://evals.ag/ site! · 2LLMGemini 3.8 FlashGen AI—Skill—Software—5.6M tokens·20m 51s00
Output C, 1 turnECEvan ChaOct 7Make a ~60s 16:9 product demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. ## Story 1. Hook (first 3s must be instantly clear): real AI-made game/video outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." 2. A real game task is typed into the Evals composer: "Build a polished playable Sonic-style 3D platform game for the browser." → Battle → up to 4 setups (Model · Skill · Environment), each landing with real values and moving UI. 3. Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. 4. Blind compare: two moving outputs, "Which would you choose?" → "I prefer A". 5. Run record of that same output: LLM, skill, environment, turn time, tokens. 6. Setup leaderboard: "Setups, not just models." 7. Community wall of real game / video / shorts outputs. 8. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. ## Content - Show only games, videos, 3D scenes, motion pieces and shorts. No websites, no landing pages. - Real Evals UI only, never mocked up. Every name and number on screen is real; only model names that exist in Evals (e.g. Claude Opus 5.5, GPT-6 Astra). - No real-person likeness. ## Voice - One energetic, confident English male narrator: excited on the hook, curious on the question, emphatic on the tagline, quick through the UI steps. - ~170 words in one continuous take that fills the whole film. No pause longer than ~0.35s, no clipped words. Say "Evals" clearly as "E-vals". - The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. ## Music & sound - Modern minimal tech-launch track, 120 BPM: punchy kick, deep 808, crisp hats, glassy synth plucks. Apple / Linear keynote feel. No cheesy EDM risers, no corporate stock vibe. - Voice always leads. Music dips ~6dB only while the voice speaks, on a smooth envelope. - SFX sit just under the music and rise above it only on hits. - 4–6 big hits get a layered reverse swell + sub boom + transient + glitch. Light ticks and whooshes elsewhere, never on every cut. ## Look - One sans font, one accent colour. - Real UI as rounded cards (16–22px radius) with soft shadows. - Sections alternate dark and light. - Never: outlines or hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI. ## Motion: this is the feel - Premium, fast, confident: an Apple / Linear keynote cut at shorts speed. - One continuous camera with velocity-continuous easing. Two easing personalities only: snappy (~25% of remaining distance per frame) and floaty (~12%). - Nothing ever stops dead. Every hold keeps drifting 1–3%. Fast moves decelerate over at least 6 frames. The end card keeps pushing until the last frame. - Exits accelerate into the cut; the next shot carries that velocity through it. - Real whip-pans: both shots travel together over 0.3–0.45s, motion blur peaking mid-move. One whip direction throughout. - Overlapping action: the next element starts while the previous one is ~60% done. No start-stop chains. - Text is screen-locked and arrives together with its container. Words reveal one by one, 2–4 frames apart, and the newest word flashes the accent. In ≤0.35s, readable ≥0.7s, out ≤0.25s. - On busy footage the camera stays calm (slow push ≤3%); the energy comes from cutting on the beat. - Every cut lands on the beat (±1 frame), with the visual hit and the SFX on the same frame. No double cuts. - Big hits (4–6 only): +8.5% zoom punch, ~16px shake at ~22Hz settling in 0.1s, RGB split 16px → 7px over 4 frames, 1–2 frames of white flash. A lighter version on secondary cuts. ## Quality bar Before calling it done, watch it like a motion director who didn't build it: - no dead stops, frozen frames or stutters - every cut on the beat - every word audible - every name and number real Fix the 5 worst moments first.LLMQwen3.8 Max (0902)Gen AI—Skill—Software—5.4M tokens·18m 25s00
Output B, 1 turnECEvan ChaOct 7https://www.instagram.com/reel/Dd4VIgTq8e7/ reference this video! you must reference the asset of the video. music, sound effect etc... and what u need to describe is https://evals.ag/ site!LLMGPT-6 AstraGen AI—Skill/product-launch-videoSoftware—1.4M tokens·12m 46s00
1 turnECEvan ChaOct 7Make a ~60s 16:9 product demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. ## Story 1. Hook (first 3s must be instantly clear): real AI-made game/video outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." 2. A real game task is typed into the Evals composer: "Build a polished playable Sonic-style 3D platform game for the browser." → Battle → up to 4 setups (Model · Skill · Environment), each landing with real values and moving UI. 3. Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. 4. Blind compare: two moving outputs, "Which would you choose?" → "I prefer A". 5. Run record of that same output: LLM, skill, environment, turn time, tokens. 6. Setup leaderboard: "Setups, not just models." 7. Community wall of real game / video / shorts outputs. 8. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. ## Content - Show only games, videos, 3D scenes, motion pieces and shorts. No websites, no landing pages. - Real Evals UI only, never mocked up. Every name and number on screen is real; only model names that exist in Evals (e.g. Claude Opus 5.5, GPT-6 Astra). - No real-person likeness. ## Voice - One energetic, confident English male narrator: excited on the hook, curious on the question, emphatic on the tagline, quick through the UI steps. - ~170 words in one continuous take that fills the whole film. No pause longer than ~0.35s, no clipped words. Say "Evals" clearly as "E-vals". - The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. ## Music & sound - Modern minimal tech-launch track, 120 BPM: punchy kick, deep 808, crisp hats, glassy synth plucks. Apple / Linear keynote feel. No cheesy EDM risers, no corporate stock vibe. - Voice always leads. Music dips ~6dB only while the voice speaks, on a smooth envelope. - SFX sit just under the music and rise above it only on hits. - 4–6 big hits get a layered reverse swell + sub boom + transient + glitch. Light ticks and whooshes elsewhere, never on every cut. ## Look - One sans font, one accent colour. - Real UI as rounded cards (16–22px radius) with soft shadows. - Sections alternate dark and light. - Never: outlines or hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI. ## Motion: this is the feel - Premium, fast, confident: an Apple / Linear keynote cut at shorts speed. - One continuous camera with velocity-continuous easing. Two easing personalities only: snappy (~25% of remaining distance per frame) and floaty (~12%). - Nothing ever stops dead. Every hold keeps drifting 1–3%. Fast moves decelerate over at least 6 frames. The end card keeps pushing until the last frame. - Exits accelerate into the cut; the next shot carries that velocity through it. - Real whip-pans: both shots travel together over 0.3–0.45s, motion blur peaking mid-move. One whip direction throughout. - Overlapping action: the next element starts while the previous one is ~60% done. No start-stop chains. - Text is screen-locked and arrives together with its container. Words reveal one by one, 2–4 frames apart, and the newest word flashes the accent. In ≤0.35s, readable ≥0.7s, out ≤0.25s. - On busy footage the camera stays calm (slow push ≤3%); the energy comes from cutting on the beat. - Every cut lands on the beat (±1 frame), with the visual hit and the SFX on the same frame. No double cuts. - Big hits (4–6 only): +8.5% zoom punch, ~16px shake at ~22Hz settling in 0.1s, RGB split 16px → 7px over 4 frames, 1–2 frames of white flash. A lighter version on secondary cuts. ## Quality bar Before calling it done, watch it like a motion director who didn't build it: - no dead stops, frozen frames or stutters - every cut on the beat - every word audible - every name and number real Fix the 5 worst moments first.LLMGPT-6.1 SolGen AI—Skill/hyperframesSoftware—3.2M tokens·20m 32s00
Output A, 1 turnECEvan ChaOct 7Make a ~60s 16:9 product demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. ## Story 1. Hook (first 3s must be instantly clear): real AI-made game/video outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." 2. A real game task is typed into the Evals composer: "Build a polished playable Sonic-style 3D platform game for the browser." → Battle → up to 4 setups (Model · Skill · Environment), each landing with real values and moving UI. 3. Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. 4. Blind compare: two moving outputs, "Which would you choose?" → "I prefer A". 5. Run record of that same output: LLM, skill, environment, turn time, tokens. 6. Setup leaderboard: "Setups, not just models." 7. Community wall of real game / video / shorts outputs. 8. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. ## Content - Show only games, videos, 3D scenes, motion pieces and shorts. No websites, no landing pages. - Real Evals UI only, never mocked up. Every name and number on screen is real; only model names that exist in Evals (e.g. Claude Opus 5.5, GPT-6 Astra). - No real-person likeness. ## Voice - One energetic, confident English male narrator: excited on the hook, curious on the question, emphatic on the tagline, quick through the UI steps. - ~170 words in one continuous take that fills the whole film. No pause longer than ~0.35s, no clipped words. Say "Evals" clearly as "E-vals". - The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. ## Music & sound - Modern minimal tech-launch track, 120 BPM: punchy kick, deep 808, crisp hats, glassy synth plucks. Apple / Linear keynote feel. No cheesy EDM risers, no corporate stock vibe. - Voice always leads. Music dips ~6dB only while the voice speaks, on a smooth envelope. - SFX sit just under the music and rise above it only on hits. - 4–6 big hits get a layered reverse swell + sub boom + transient + glitch. Light ticks and whooshes elsewhere, never on every cut. ## Look - One sans font, one accent colour. - Real UI as rounded cards (16–22px radius) with soft shadows. - Sections alternate dark and light. - Never: outlines or hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI. ## Motion: this is the feel - Premium, fast, confident: an Apple / Linear keynote cut at shorts speed. - One continuous camera with velocity-continuous easing. Two easing personalities only: snappy (~25% of remaining distance per frame) and floaty (~12%). - Nothing ever stops dead. Every hold keeps drifting 1–3%. Fast moves decelerate over at least 6 frames. The end card keeps pushing until the last frame. - Exits accelerate into the cut; the next shot carries that velocity through it. - Real whip-pans: both shots travel together over 0.3–0.45s, motion blur peaking mid-move. One whip direction throughout. - Overlapping action: the next element starts while the previous one is ~60% done. No start-stop chains. - Text is screen-locked and arrives together with its container. Words reveal one by one, 2–4 frames apart, and the newest word flashes the accent. In ≤0.35s, readable ≥0.7s, out ≤0.25s. - On busy footage the camera stays calm (slow push ≤3%); the energy comes from cutting on the beat. - Every cut lands on the beat (±1 frame), with the visual hit and the SFX on the same frame. No double cuts. - Big hits (4–6 only): +8.5% zoom punch, ~16px shake at ~22Hz settling in 0.1s, RGB split 16px → 7px over 4 frames, 1–2 frames of white flash. A lighter version on secondary cuts. ## Quality bar Before calling it done, watch it like a motion director who didn't build it: - no dead stops, frozen frames or stutters - every cut on the beat - every word audible - every name and number real Fix the 5 worst moments first.LLMGPT-6.1 SolGen AI—Skill—Software—3.8M tokens·13m 12s00
1 turnECEvan ChaOct 7make a demo video about https://evals.ag/LLMGPT-6 AstraGen AIEleven v4+1Skill/demo-video-creatorSoftwarefal.ai4.6M tokens·36m 38s00
Output C, 2 turnsECEvan ChaOct 7reference this video for motion graphic demo https://x.com/samgrows/status/2107571800020222393?s=20 you must describe this site https://evals.ag/LLMClaude Opus 5.5Gen AI—Skill/agent-browserSoftwareFFmpeg17.4M tokens·1h 33m 12s00
Output D, 1 turnECEvan ChaOct 7reference this video for motion graphic demo https://x.com/samgrows/status/2107571800020222393?s=20 you must describe this site https://evals.ag/LLMQwen3.8 Max (0902)Gen AI—Skill/agent-browserSoftwareFFmpeg2.6M tokens·41m 47s00
Output B, 1 turnECEvan ChaOct 7reference this video for motion graphic demo https://x.com/samgrows/status/2107571800020222393?s=20 you must describe this site https://evals.ag/LLMGPT-6.1 SolGen AI—Skill/agent-browserSoftware—398.5K tokens·5m 1s00
Output A, 1 turnECEvan ChaOct 7reference this video for motion graphic demo https://x.com/samgrows/status/2107571800020222393?s=20 you must describe this site https://evals.ag/LLMGemini 3.8 FlashGen AI—Skill/agent-browserSoftware—3M tokens·21m 19s00
1 turnECEvan ChaOct 7Make a ~60s 16:9 product demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. ## Story 1. Hook (first 3s must be instantly clear): real AI-made game/video outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." 2. A real game task is typed into the Evals composer: "Build a polished playable Sonic-style 3D platform game for the browser." → Battle → up to 4 setups (Model · Skill · Environment), each landing with real values and moving UI. 3. Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. 4. Blind compare: two moving outputs, "Which would you choose?" → "I prefer A". 5. Run record of that same output: LLM, skill, environment, turn time, tokens. 6. Setup leaderboard: "Setups, not just models." 7. Community wall of real game / video / shorts outputs. 8. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. ## Content - Show only games, videos, 3D scenes, motion pieces and shorts. No websites, no landing pages. - Real Evals UI only, never mocked up. Every name and number on screen is real; only model names that exist in Evals (e.g. Claude Opus 5.5, GPT-6 Astra). - No real-person likeness. ## Voice - One energetic, confident English male narrator: excited on the hook, curious on the question, emphatic on the tagline, quick through the UI steps. - ~170 words in one continuous take that fills the whole film. No pause longer than ~0.35s, no clipped words. Say "Evals" clearly as "E-vals". - The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. ## Music & sound - Modern minimal tech-launch track, 120 BPM: punchy kick, deep 808, crisp hats, glassy synth plucks. Apple / Linear keynote feel. No cheesy EDM risers, no corporate stock vibe. - Voice always leads. Music dips ~6dB only while the voice speaks, on a smooth envelope. - SFX sit just under the music and rise above it only on hits. - 4–6 big hits get a layered reverse swell + sub boom + transient + glitch. Light ticks and whooshes elsewhere, never on every cut. ## Look - One sans font, one accent colour. - Real UI as rounded cards (16–22px radius) with soft shadows. - Sections alternate dark and light. - Never: outlines or hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI. ## Motion: this is the feel - Premium, fast, confident: an Apple / Linear keynote cut at shorts speed. - One continuous camera with velocity-continuous easing. Two easing personalities only: snappy (~25% of remaining distance per frame) and floaty (~12%). - Nothing ever stops dead. Every hold keeps drifting 1–3%. Fast moves decelerate over at least 6 frames. The end card keeps pushing until the last frame. - Exits accelerate into the cut; the next shot carries that velocity through it. - Real whip-pans: both shots travel together over 0.3–0.45s, motion blur peaking mid-move. One whip direction throughout. - Overlapping action: the next element starts while the previous one is ~60% done. No start-stop chains. - Text is screen-locked and arrives together with its container. Words reveal one by one, 2–4 frames apart, and the newest word flashes the accent. In ≤0.35s, readable ≥0.7s, out ≤0.25s. - On busy footage the camera stays calm (slow push ≤3%); the energy comes from cutting on the beat. - Every cut lands on the beat (±1 frame), with the visual hit and the SFX on the same frame. No double cuts. - Big hits (4–6 only): +8.5% zoom punch, ~16px shake at ~22Hz settling in 0.1s, RGB split 16px → 7px over 4 frames, 1–2 frames of white flash. A lighter version on secondary cuts. ## Quality bar Before calling it done, watch it like a motion director who didn't build it: - no dead stops, frozen frames or stutters - every cut on the beat - every word audible - every name and number real Fix the 5 worst moments first.LLMGPT-6.1 SolGen AIEleven v4+1Skill/hyperframesSoftwarefal.ai5.9M tokens·49m 6s00
1 turnECEvan ChaOct 6reference this video for motion graphic demo https://x.com/samgrows/status/2107571800020222393?s=20 you must describe this site https://evals.ag/LLMGPT-6.1 SolGen AI—Skill—Software—404.3K tokens·6m 6s00
1 turnECEvan ChaOct 6could you make a demo video reference this site https://x.com/jordanarchivess/status/2107536262894915762?s=20. what you need to make it site is https://evals.ag. you must use tts with elevenlabs v4 apiLLMGPT-6.1 SolGen AIEleven v4+1Skill/hyperframesSoftwarefal.ai2.5M tokens·11m 53s00
1 turnECEvan ChaOct 6START NOW. Do not ask me any questions and do not wait for input — every input is below. Make every unspecified decision yourself, log it in DECISIONS.md, and keep going until ONE finished video file exists. The only deliverable is that one .mp4. Task Make ONE ~60s landscape demo video for Evals (evals.ag): battle AI setups on your own task and vote blind. Output: /Users/evan/workspace/video-generator/videos/evals-demo3/renders/evals-demo3.mp4, 1920×1080, 30fps, H.264 + AAC 192k, −14 LUFS. Nothing else to deliver (no vertical version, no variants). Work dir: repo /Users/evan/workspace/video-generator (Claude Code on this Mac). Build in HyperFrames (HTML + one paused GSAP timeline) under videos/evals-demo3/. Render: npx --yes hyperframes@0.8.101 render -o <out> --quiet from the project dir. Use curl for HTTP/API calls (python has no SSL certs). ElevenLabs key: ELEVENLABS_API_KEY in /Users/evan/workspace/video-generator/.env. Story (English voiceover + on-screen type) Hook, first 3s must be instantly clear: real AI-made GAME/VIDEO outputs of the SAME prompt slam in side by side, moving, full-bleed → big type "Which one is better?" → rapid A/B/C/D flips on the beats → "Same prompt. Different setups." → "You can't tell by guessing." Type a real game/video task into the Evals composer → Battle → up to 4 setups (Model · Skill · Environment, each landing with real values and moving UI). Send → velocity-matched zoom-through into the big hit → the 4 results play side by side. Blind compare: two moving outputs, "Which would you choose?" → vote ("I prefer A"). Run record of that same output (real LLM, skill, environment, turn time, tokens). Setup leaderboard ("Setups, not just models."). Community wall of real game/video/shorts outputs. End card: Evals mark + wordmark, "Stop guessing. Battle it.", evals.ag. Content rules Show ONLY videos, games, 3D scenes, motion videos, shorts made in Evals. No websites / landing pages. Main thread = the real 4-setup Sonic battle: prompt "Build a polished playable Sonic-style 3D platform game for the browser.", clips /Users/evan/workspace/video-generator/videos/evals-launch/assets/clips/run_a.mp4 … run_d.mp4, run https://evals-app.vercel.app/runs/run_597Q2CRE4FXNJ2R7HZXMVF4Z03. Capture fresh real UI from https://evals-app.vercel.app with Playwright at 2× DPR: composer, Run-mode menu, setups stepper, compare page, output run record, Leaderboards, Explore → Outputs. If a page needs sign-in, use the public pages. Record extra community game/video outputs from their public "Open preview" links. Screencast 1600×900, 30fps, 6–7s, with keys pressed or orbit so they move. Reuse the pattern in /Users/evan/workspace/video-generator/videos/evals-launch/scripts/rec.mjs. No real-person likeness, no invented features/numbers/names. Only show model names you see in captures (e.g. Claude Opus 5.5, GPT-6 Astra). Voiceover (must flow continuously) ElevenLabs voice Will bIHbv24MWmeRgasZH58o, model eleven_v4, stability 0.15, similarity 0.8, style 0.8, speed 1.1, with emotion tags in the text ([excited] [confident] [curious] [emphatic] [quick]). Use the /with-timestamps endpoint. Write ~165–180 words so the voice fills the whole film. Generate it as ONE take and lay it unspliced — never cut it into phrases or time-stretch pieces. Longest silence inside the VO ≤ 0.35s. No clipped words. The edit follows the voice: each shot change lands on the beat nearest the start of the phrase it illustrates. Verify with Whisper that every word is there and no tags are spoken. If "Evals" sounds like "evils", respell it "E-vals". Music & sound New modern premium BGM, not cheesy. Generate 2–3 candidates with ElevenLabs Music: "modern minimal tech launch, punchy kick, deep 808, crisp hats, glassy synth plucks, Apple/Linear keynote feel, no cheesy EDM risers, no corporate/stock vibe, 120 BPM". Pick the cleanest; detect BPM and the first downbeat. VO leads at about −16 LUFS. Music ducks about 6dB only while he speaks (smooth envelope). SFX bus ~3dB under the music, above it only on hits. At 4–6 big hits, layer reverse swell + boom (HP 60Hz) + transient + glitch. Light ticks and whooshes elsewhere — never on every cut. Look Fresh premium design system, written once in DESIGN.md: one sans font, one accent colour real UI captures as rounded cards (16–22px) with soft shadows dark/light section switches Never: outlines/hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI mockups. Motion (non-negotiable — this is the feel) Reference for feel: /Users/evan/workspace/video-generator/videos/evals-hype/renders/evals-ref2-short-fast2-impact.mp4; rules in /Users/evan/workspace/video-generator/videos/evals-hype/MOTION_FIX.md. One camera = a pure function of time with velocity-continuous easing. Deterministic: no timers or randomness. Two easing families only: snappy ≈25% of remaining distance per frame, floaty ≈12%. Exits accelerate into the cut; the next shot carries the velocity through it. Real whip-pans: both shots travel together over 0.3–0.45s with velocity blur. Overlays are screen-locked. Overlapping action, no start-stop chains. Every hold drifts 1–3%. No dead stops, no frozen frames, no double cuts. Word-by-word reveals 2–4 frames apart; the newest word flashes the accent. Cuts on the beat ±1 frame. Speed by re-rendering: render at --fps 30/speedup (e.g. --fps 95/4 = 1.263×), play back at 30fps. Impact FX in post (ffmpeg) on 4–6 big hits only: +8.5% zoom punch (τ 0.09s) 16px shake at ~22Hz (τ 0.1s) rgbashift 16px for 2 frames, then 7px for 2 frames 1–2 frames of white flash A lighter version on secondary cuts. QA (do it, then deliver) python3 /Users/evan/workspace/video-generator/videos/evals-hype/scripts/motion.py <render> <bpm> <first_beat>: no dead stops, no frozen holds, cuts on the beat. sh /Users/evan/workspace/video-generator/videos/evals-hype/scripts/strip.sh <render> <t> <out.png> at every transition — look at each strip. VO RMS gap check (≤ 0.35s) plus a Whisper pass on the final mix. Text readable and inside the margins; every name and number on screen is real. Fix anything that fails, re-render, and finish with the single file videos/evals-demo3/renders/evals-demo3.mp4.LLMGemini 3.8 FlashGen AI—Skill—Software—27.5M tokens·58m 44s00
Output A, 1 turnECEvan ChaOct 6Make a ~60s product launch film for Evals (evals.ag) — the place to battle AI setups on your own task and vote blind. 16:9, 1920×1080, 30fps, music-locked, no narration — on-screen type tells the story. Build it in HyperFrames (HTML + one paused GSAP timeline) under videos/evals-<name>/. 0. Inputs Product: https://evals.ag (app: https://evals-app.vercel.app — use my logged-in Chrome session for anything behind sign-in). Motion reference: videos/evals-hype/renders/evals-ref2-short-fast2-impact.mp4 (approved). Match its feel exactly; change only scenario, design and assets. Core message: "Same prompt, different setups — battle them and vote blind. Stop guessing. Battle it." What Evals really does (only show these — no invented features, numbers, rankings or names) Ask: type what you want to make in the composer ("What would you like to make?"). Run modes: Battle (blind), Side-by-side comparison (pick models and skills), Direct. Battle: up to 4 setups. A setup = model + skill + environment (e.g. Claude Opus 5.5 / GPT-6 Astra, skills like hyperframes / frontend-design / web-design-guidelines, environments like Native Claude / Recorded session). Compare, blind: same prompt, outputs side by side as playable live previews, names hidden until you vote ("I prefer A / B"). Inspect: every output keeps its full run record — LLM, gen AI, skill, software, environment, services contacted, turn time, turn tokens, status. Rank: votes roll up into leaderboards of setups, not just models ("Setups, not just models." — same model, different setups). Explore: community Outputs, Tasks, Skills, Compare, Community, Leaderboards. Real model names visible in the product: Claude Opus 5.5, GPT-6 Astra. Show only names/numbers you actually see in your captures. End card: Evals mark + wordmark, "Stop guessing. Battle it.", URL evals.ag. 1. Reverse-engineer the reference first (frames are the truth) Download the reference and study it frame by frame. Write SPEC.md with measured numbers, not impressions: every cut (frame number), shot order and durations, average shot length, cuts per 10s camera moves per shot (start/end position, scale, curve), easing measured per frame (fraction of remaining distance covered each frame) type-on speed (frames per char), caret behaviour, word-by-word reveal spacing, new-word accent flash duration card/UI choreography (fly-in paths, stacking, depth), light↔dark section switches, split-screen layouts sound map: where every hit/whoosh/click lands relative to the cut Then rebuild shot for shot at the same rhythm. Only brand, copy, design and assets change. 2. Scenario Write SCRIPT.md before building: a fresh narrative arc true to what Evals really does (hook: a new SOTA model every week — which one is actually best at your work? → type a task → Battle up to 4 setups → blind side-by-side, vote → full run record → setup leaderboards → "Stop guessing. Battle it." evals.ag). Don't reuse the previous films' copy. Short punchy lines (2–5 words per beat), English. Every beat gets a timestamp on the music grid. 3. Design Write DESIGN.md once and make every section obey it (no per-section drift): palette (one accent per frame), type pairing (one sans for headlines; weights/sizes in px; tracking −0.02em, line-height 1.0) card treatment: real UI captures, rounded 14–18px corners, soft layered shadow Never: outline/hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI mockups, invented numbers/rankings/features. 4. Assets (all real, all new) Capture the real Evals UI with Playwright at 2–3× DPR: home composer (empty / typed), Run-mode menu (Battle / Side-by-side / Direct), setups stepper (1–4), a live run page (4 agents generating → "Your options are ready"), Compare (blind A/B with "I prefer A/B"), output run record panel, Leaderboards (setup rows), Explore Outputs/Tasks/Skills. Run one real 4-setup Battle on a visual prompt (e.g. a playable browser game) so the four results are genuinely the same task. Record real Evals outputs as clips from their live previews (screencast, 1600×900, 30fps, 6–7s each; press start/arrow keys or orbit so they move). Use outputs not used in earlier Evals films. Typing scenes: put a text layer exactly over the real input field, matching its font, size and colour. New music track (detect BPM and first downbeat with a script) and a fresh SFX set. 5. Motion rules (this is the 감도 — non-negotiable) One camera = one pure function of time cam(t) built from keyframes with velocity-continuous easing; applied from the master timeline. Deterministic: no timers, no randomness, any frame renders alone. Two easing families only: snappy ≈25% of remaining distance per frame (90% in 8–9 frames) and floaty ≈12% (18–20 frames). No overshoot unless the reference has it. Exits accelerate into the cut; the next shot carries the motion through it (velocity-matched). Real whip-pans: outgoing and incoming shots travel together over 0.30–0.45s, with velocity-driven motion blur that peaks mid-move. Never blur switched on over a hard cut. Overlays are screen-locked (never inherit the camera whip). Text arrives with its container in one mask reveal. Overlapping action: the next element starts at ~60% of the previous tween. No start-stop chains. Nothing freezes: every hold carries a 1–3% slow drift (sine.inOut across the whole hold). Fast moves decelerate over ≥6 frames — no dead stops. No double cuts (two visual changes on adjacent frames). Busy footage gets only a slow push; energy comes from the cuts. Word reveals 2–4 frames apart; the newest word flashes the accent, then settles 3–12 frames later. Hero type-on 1 char / 2 frames; in-product typing 0.6–0.9 char/frame; caret solid (never blinks). Every cut on the beat ±1 frame; visual hit and SFX on the same frame. 6. Speed, edit, impact Speed by re-rendering, never frame-dropping: render at --fps 30/speedup and play back at 30fps (e.g. --fps 95/4 → 1.263×). Pick the speed-up so one beat = a whole number of frames. Quantize every segment to 8th notes; round segment start frames up. Impact FX in post (ffmpeg) on 4–6 big hits only (cut to dark, reveal word, flash, key word, logo): zoom punch +8.5%, decaying τ = 0.09s crop shake 16px at ~22Hz, decaying τ = 0.1s rgbashift 16px for 2 frames, then 7px for 2 frames 1–2 frames of white flash Lighter version on secondary cuts (3.5% punch, 6px shake, 1 frame of split). Find hit frames from frame-diff peaks and by looking. 7. Sound Music bed carries the film; no ducking. SFX bus ~3dB under the music overall, above it only on hits and in the music's own gaps. At each big hit, layer reverse swell + boom (high-passed at 60Hz) + short transient + glitch synced to the RGB split. Use a modest total (~40–45 sounds); never spread SFX evenly on every cut. Master at −14 LUFS, AAC 192k. 8. QA before you report (mandatory) motion.py <render> <bpm> <first_beat> must show: no dead stops, no frozen holds, every cut within ±1.5 frames of a beat, no adjacent-frame cuts. strip.sh (16 consecutive frames) at every transition, and look at each one. A good whip shows a smooth velocity ramp with both shots visible mid-move. Contact sheets alone hide motion bugs. Check text readability, safe margins, and that every number/name on screen exists in the real product. Deliver renders/evals-<name>-v1.mp4 plus a report: script, design summary, asset list, shot list with times, BPM and speed factor, hit frames with FX, SFX map, QA results.LLMGrok 4.7Gen AI—Skill/hyperframesSoftware—12.9M tokens·59m 42s00
Output D, 1 turnECEvan ChaOct 6Make a ~60s product launch film for Evals (evals.ag) — the place to battle AI setups on your own task and vote blind. 16:9, 1920×1080, 30fps, music-locked, no narration — on-screen type tells the story. Build it in HyperFrames (HTML + one paused GSAP timeline) under videos/evals-<name>/. 0. Inputs Product: https://evals.ag (app: https://evals-app.vercel.app — use my logged-in Chrome session for anything behind sign-in). Motion reference: videos/evals-hype/renders/evals-ref2-short-fast2-impact.mp4 (approved). Match its feel exactly; change only scenario, design and assets. Core message: "Same prompt, different setups — battle them and vote blind. Stop guessing. Battle it." What Evals really does (only show these — no invented features, numbers, rankings or names) Ask: type what you want to make in the composer ("What would you like to make?"). Run modes: Battle (blind), Side-by-side comparison (pick models and skills), Direct. Battle: up to 4 setups. A setup = model + skill + environment (e.g. Claude Opus 5.5 / GPT-6 Astra, skills like hyperframes / frontend-design / web-design-guidelines, environments like Native Claude / Recorded session). Compare, blind: same prompt, outputs side by side as playable live previews, names hidden until you vote ("I prefer A / B"). Inspect: every output keeps its full run record — LLM, gen AI, skill, software, environment, services contacted, turn time, turn tokens, status. Rank: votes roll up into leaderboards of setups, not just models ("Setups, not just models." — same model, different setups). Explore: community Outputs, Tasks, Skills, Compare, Community, Leaderboards. Real model names visible in the product: Claude Opus 5.5, GPT-6 Astra. Show only names/numbers you actually see in your captures. End card: Evals mark + wordmark, "Stop guessing. Battle it.", URL evals.ag. 1. Reverse-engineer the reference first (frames are the truth) Download the reference and study it frame by frame. Write SPEC.md with measured numbers, not impressions: every cut (frame number), shot order and durations, average shot length, cuts per 10s camera moves per shot (start/end position, scale, curve), easing measured per frame (fraction of remaining distance covered each frame) type-on speed (frames per char), caret behaviour, word-by-word reveal spacing, new-word accent flash duration card/UI choreography (fly-in paths, stacking, depth), light↔dark section switches, split-screen layouts sound map: where every hit/whoosh/click lands relative to the cut Then rebuild shot for shot at the same rhythm. Only brand, copy, design and assets change. 2. Scenario Write SCRIPT.md before building: a fresh narrative arc true to what Evals really does (hook: a new SOTA model every week — which one is actually best at your work? → type a task → Battle up to 4 setups → blind side-by-side, vote → full run record → setup leaderboards → "Stop guessing. Battle it." evals.ag). Don't reuse the previous films' copy. Short punchy lines (2–5 words per beat), English. Every beat gets a timestamp on the music grid. 3. Design Write DESIGN.md once and make every section obey it (no per-section drift): palette (one accent per frame), type pairing (one sans for headlines; weights/sizes in px; tracking −0.02em, line-height 1.0) card treatment: real UI captures, rounded 14–18px corners, soft layered shadow Never: outline/hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI mockups, invented numbers/rankings/features. 4. Assets (all real, all new) Capture the real Evals UI with Playwright at 2–3× DPR: home composer (empty / typed), Run-mode menu (Battle / Side-by-side / Direct), setups stepper (1–4), a live run page (4 agents generating → "Your options are ready"), Compare (blind A/B with "I prefer A/B"), output run record panel, Leaderboards (setup rows), Explore Outputs/Tasks/Skills. Run one real 4-setup Battle on a visual prompt (e.g. a playable browser game) so the four results are genuinely the same task. Record real Evals outputs as clips from their live previews (screencast, 1600×900, 30fps, 6–7s each; press start/arrow keys or orbit so they move). Use outputs not used in earlier Evals films. Typing scenes: put a text layer exactly over the real input field, matching its font, size and colour. New music track (detect BPM and first downbeat with a script) and a fresh SFX set. 5. Motion rules (this is the 감도 — non-negotiable) One camera = one pure function of time cam(t) built from keyframes with velocity-continuous easing; applied from the master timeline. Deterministic: no timers, no randomness, any frame renders alone. Two easing families only: snappy ≈25% of remaining distance per frame (90% in 8–9 frames) and floaty ≈12% (18–20 frames). No overshoot unless the reference has it. Exits accelerate into the cut; the next shot carries the motion through it (velocity-matched). Real whip-pans: outgoing and incoming shots travel together over 0.30–0.45s, with velocity-driven motion blur that peaks mid-move. Never blur switched on over a hard cut. Overlays are screen-locked (never inherit the camera whip). Text arrives with its container in one mask reveal. Overlapping action: the next element starts at ~60% of the previous tween. No start-stop chains. Nothing freezes: every hold carries a 1–3% slow drift (sine.inOut across the whole hold). Fast moves decelerate over ≥6 frames — no dead stops. No double cuts (two visual changes on adjacent frames). Busy footage gets only a slow push; energy comes from the cuts. Word reveals 2–4 frames apart; the newest word flashes the accent, then settles 3–12 frames later. Hero type-on 1 char / 2 frames; in-product typing 0.6–0.9 char/frame; caret solid (never blinks). Every cut on the beat ±1 frame; visual hit and SFX on the same frame. 6. Speed, edit, impact Speed by re-rendering, never frame-dropping: render at --fps 30/speedup and play back at 30fps (e.g. --fps 95/4 → 1.263×). Pick the speed-up so one beat = a whole number of frames. Quantize every segment to 8th notes; round segment start frames up. Impact FX in post (ffmpeg) on 4–6 big hits only (cut to dark, reveal word, flash, key word, logo): zoom punch +8.5%, decaying τ = 0.09s crop shake 16px at ~22Hz, decaying τ = 0.1s rgbashift 16px for 2 frames, then 7px for 2 frames 1–2 frames of white flash Lighter version on secondary cuts (3.5% punch, 6px shake, 1 frame of split). Find hit frames from frame-diff peaks and by looking. 7. Sound Music bed carries the film; no ducking. SFX bus ~3dB under the music overall, above it only on hits and in the music's own gaps. At each big hit, layer reverse swell + boom (high-passed at 60Hz) + short transient + glitch synced to the RGB split. Use a modest total (~40–45 sounds); never spread SFX evenly on every cut. Master at −14 LUFS, AAC 192k. 8. QA before you report (mandatory) motion.py <render> <bpm> <first_beat> must show: no dead stops, no frozen holds, every cut within ±1.5 frames of a beat, no adjacent-frame cuts. strip.sh (16 consecutive frames) at every transition, and look at each one. A good whip shows a smooth velocity ramp with both shots visible mid-move. Contact sheets alone hide motion bugs. Check text readability, safe margins, and that every number/name on screen exists in the real product. Deliver renders/evals-<name>-v1.mp4 plus a report: script, design summary, asset list, shot list with times, BPM and speed factor, hit frames with FX, SFX map, QA results.LLMGPT-6.1 SolGen AI—Skill/hyperframesSoftware—289.4K tokens·8m 18s00
Output B, 2 turnsECEvan ChaOct 6Make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464 You must reference this site https://evals.ag/. You must use font sanserif. Motion must be speedy. Never make ai slop.LLMClaude Opus 5.5Gen AI—Skill—Software—16.7M tokens·55m 23s00
Output B, 1 turnECEvan ChaOct 6Make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464 You must reference this site https://evals.ag/. You must use font sanserif. Motion must be speedy. Never make ai slop.LLMClaude Opus 5.5Gen AI—Skill—Software—8.5M tokens·33m 27s00
Output A, 1 turnECEvan ChaOct 6Make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464 You must reference this site https://evals.ag/. You must use font sanserif. Motion must be speedy. Never make ai slop.LLMClaude Opus 5.5Gen AI—Skill—Software—7.5M tokens·24m 48s00
Output A, 1 turnECEvan ChaOct 6Make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464 You must reference this site https://evals.ag/. You must use font sanserif. Motion must be speedy. Never make ai slop.LLMClaude Opus 5.5Gen AI—Skill—SoftwareHiggsfield5.2M tokens·15m 3s00
1 turnECEvan ChaOct 6Evals landscape demo-video prompt (approved "감도") Copy everything below the line into a new session. Make a ~60s landscape product demo video for Evals (evals.ag) — the place to battle AI setups on your own task and vote blind. Horizontal 16:9, 1920×1080, 30fps (for YouTube / website / X timeline). Not a vertical Short/Reel — no 9:16 output, no vertical crops, no phone safe-zones. Music-locked, no narration — on-screen type tells the story while the real product is demoed end to end. Build it in HyperFrames (HTML + one paused GSAP timeline) under /Users/evan/workspace/video-generator/videos/evals-demo3/. 0. Inputs Product: https://evals.ag (app: https://evals-app.vercel.app — use the logged-in Claude in Chrome session if connected; otherwise public pages). Motion reference (local file, already downloaded): /Users/evan/workspace/video-generator/videos/evals-hype/renders/evals-ref2-short-fast2-impact.mp4 (approved). Match its motion feel exactly (it is also 16:9); change only scenario, design and assets. Its 26s cut-down is just the reference for feel — this video runs the full ~60s. Core message: "Same prompt, different setups — battle them and vote blind. Stop guessing. Battle it." Run it autonomously — do not ask me for input Work in Claude Code with the repo /Users/evan/workspace/video-generator as the working directory. Every path below is absolute; everything you need is already on this machine. Do NOT stop to ask questions or request inputs. Every input is specified here; for anything not specified, pick a sensible default, write the decision into DECISIONS.md, and keep going until the final file is rendered. If something is unavailable, use the fallback and continue: Chrome not connected or a page needs sign-in → use the public pages (https://evals-app.vercel.app home, /compare, leaderboards, explore) with Playwright, and record outputs from their public "Open preview" links. Can't run a new Battle → use public outputs that share the same prompt (the /compare pages show pairs). Music: pick an unused track from /Users/evan/workspace/video-generator/videos/evals-hype/music3/ or music2/. SFX: /Users/evan/workspace/video-generator/videos/evals-hype/sfx2/, ref2/sfx3/, ref2/sfx4/. QA scripts: /Users/evan/workspace/video-generator/videos/evals-hype/scripts/motion.py and strip.sh (also beats.py for BPM). If they fail, write equivalents. Render: npx --yes hyperframes@0.8.101 render -o <out.mp4> --quiet from the project dir (--fps 95/4 etc. for speed re-render). API calls via curl (python has no SSL certs here). What Evals really does (only show these — no invented features, numbers, rankings or names) Ask: type what you want to make in the composer ("What would you like to make?"). Run modes: Battle (blind), Side-by-side comparison (pick models and skills), Direct. Battle: up to 4 setups. A setup = model + skill + environment (e.g. Claude Opus 5.5 / GPT-6 Astra, skills like hyperframes / frontend-design / web-design-guidelines, environments like Native Claude / Recorded session). Compare, blind: same prompt, outputs side by side as playable live previews, names hidden until you vote ("I prefer A / B"). Inspect: every output keeps its full run record — LLM, gen AI, skill, software, environment, services contacted, turn time, turn tokens, status. Rank: votes roll up into leaderboards of setups, not just models ("Setups, not just models." — same model, different setups). Explore: community Outputs, Tasks, Skills, Compare, Community, Leaderboards. Real model names visible in the product: Claude Opus 5.5, GPT-6 Astra. Show only names/numbers you actually see in your captures. End card: Evals mark + wordmark, "Stop guessing. Battle it.", URL evals.ag. 1. Reverse-engineer the reference first (frames are the truth) Study the local reference file frame by frame (ffmpeg frame dumps + frame-diff). Its design notes are in /Users/evan/workspace/video-generator/videos/evals-hype/ref2/SPEC.md and MOTION_FIX.md — read them, but design and copy must be new. Write SPEC.md with measured numbers, not impressions: every cut (frame number), shot order and durations, average shot length, cuts per 10s camera moves per shot (start/end position, scale, curve), easing measured per frame (fraction of remaining distance covered each frame) type-on speed (frames per char), caret behaviour, word-by-word reveal spacing, new-word accent flash duration card/UI choreography (fly-in paths, stacking, depth), light↔dark section switches, split-screen layouts sound map: where every hit/whoosh/click lands relative to the cut Then rebuild shot for shot at the same rhythm. Only brand, copy, design and assets change. 2. Scenario Write SCRIPT.md before building: a fresh narrative arc true to what Evals really does (hook: a new SOTA model every week — which one is actually best at your work? → type a task → Battle up to 4 setups → blind side-by-side, vote → full run record → setup leaderboards → "Stop guessing. Battle it." evals.ag). Don't reuse the previous films' copy. Short punchy lines (2–5 words per beat), English. Every beat gets a timestamp on the music grid. 3. Design Write DESIGN.md once and make every section obey it (no per-section drift): palette (one accent per frame), type pairing (one sans for headlines; weights/sizes in px; tracking −0.02em, line-height 1.0) card treatment: real UI captures, rounded 14–18px corners, soft layered shadow Never: outline/hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI mockups, invented numbers/rankings/features. 4. Assets (all real, all new) Capture the real Evals UI with Playwright at 2–3× DPR: home composer (empty / typed), Run-mode menu (Battle / Side-by-side / Direct), setups stepper (1–4), a live run page (4 agents generating → "Your options are ready"), Compare (blind A/B with "I prefer A/B"), output run record panel, Leaderboards (setup rows), Explore Outputs/Tasks/Skills. Run one real 4-setup Battle on a visual prompt (e.g. a playable browser game) so the four results are genuinely the same task. Record real Evals outputs as clips from their live previews (screencast, 1600×900, 30fps, 6–7s each; press start/arrow keys or orbit so they move). Use outputs not used in earlier Evals films. Typing scenes: put a text layer exactly over the real input field, matching its font, size and colour. New music track (detect BPM and first downbeat with a script) and a fresh SFX set. 5. Motion rules (this is the 감도 — non-negotiable) One camera = one pure function of time cam(t) built from keyframes with velocity-continuous easing; applied from the master timeline. Deterministic: no timers, no randomness, any frame renders alone. Two easing families only: snappy ≈25% of remaining distance per frame (90% in 8–9 frames) and floaty ≈12% (18–20 frames). No overshoot unless the reference has it. Exits accelerate into the cut; the next shot carries the motion through it (velocity-matched). Real whip-pans: outgoing and incoming shots travel together over 0.30–0.45s, with velocity-driven motion blur that peaks mid-move. Never blur switched on over a hard cut. Overlays are screen-locked (never inherit the camera whip). Text arrives with its container in one mask reveal. Overlapping action: the next element starts at ~60% of the previous tween. No start-stop chains. Nothing freezes: every hold carries a 1–3% slow drift (sine.inOut across the whole hold). Fast moves decelerate over ≥6 frames — no dead stops. No double cuts (two visual changes on adjacent frames). Busy footage gets only a slow push; energy comes from the cuts. Word reveals 2–4 frames apart; the newest word flashes the accent, then settles 3–12 frames later. Hero type-on 1 char / 2 frames; in-product typing 0.6–0.9 char/frame; caret solid (never blinks). Every cut on the beat ±1 frame; visual hit and SFX on the same frame. 6. Speed, edit, impact Speed by re-rendering, never frame-dropping: render at --fps 30/speedup and play back at 30fps (e.g. --fps 95/4 → 1.263×). Pick the speed-up so one beat = a whole number of frames. Quantize every segment to 8th notes; round segment start frames up. Impact FX in post (ffmpeg) on 4–6 big hits only (cut to dark, reveal word, flash, key word, logo): zoom punch +8.5%, decaying τ = 0.09s crop shake 16px at ~22Hz, decaying τ = 0.1s rgbashift 16px for 2 frames, then 7px for 2 frames 1–2 frames of white flash Lighter version on secondary cuts (3.5% punch, 6px shake, 1 frame of split). Find hit frames from frame-diff peaks and by looking. 7. Sound Music bed carries the film; no ducking. SFX bus ~3dB under the music overall, above it only on hits and in the music's own gaps. At each big hit, layer reverse swell + boom (high-passed at 60Hz) + short transient + glitch synced to the RGB split. Use a modest total (~40–45 sounds); never spread SFX evenly on every cut. Master at −14 LUFS, AAC 192k. 8. QA before you report (mandatory) python3 /Users/evan/workspace/video-generator/videos/evals-hype/scripts/motion.py <render> <bpm> <first_beat> must show: no dead stops, no frozen holds, every cut within ±1.5 frames of a beat, no adjacent-frame cuts. sh /Users/evan/workspace/video-generator/videos/evals-hype/scripts/strip.sh <render> <t> <out.png> (16 consecutive frames) at every transition, and look at each one. A good whip shows a smooth velocity ramp with both shots visible mid-move. Contact sheets alone hide motion bugs. Check text readability, safe margins, and that every number/name on screen exists in the real product. Deliver one 1920×1080 file renders/evals-demo3-v1.mp4 (no vertical version) plus a report: script, design summary, asset list, shot list with times, BPM and speed factor, hit frames with FX, SFX map, QA results.LLMGPT-6.1 SolGen AI—Skill/hyperframesSoftware—4.4M tokens·45m 11s00
1 turnECEvan ChaOct 6Make a ~60s product launch film for Evals (evals.ag) — the place to battle AI setups on your own task and vote blind. 16:9, 1920×1080, 30fps, music-locked, no narration — on-screen type tells the story. Build it in HyperFrames (HTML + one paused GSAP timeline) under videos/evals-<name>/. 0. Inputs Product: https://evals.ag (app: https://evals-app.vercel.app — use my logged-in Chrome session for anything behind sign-in). Motion reference: videos/evals-hype/renders/evals-ref2-short-fast2-impact.mp4 (approved). Match its feel exactly; change only scenario, design and assets. Core message: "Same prompt, different setups — battle them and vote blind. Stop guessing. Battle it." What Evals really does (only show these — no invented features, numbers, rankings or names) Ask: type what you want to make in the composer ("What would you like to make?"). Run modes: Battle (blind), Side-by-side comparison (pick models and skills), Direct. Battle: up to 4 setups. A setup = model + skill + environment (e.g. Claude Opus 5.5 / GPT-6 Astra, skills like hyperframes / frontend-design / web-design-guidelines, environments like Native Claude / Recorded session). Compare, blind: same prompt, outputs side by side as playable live previews, names hidden until you vote ("I prefer A / B"). Inspect: every output keeps its full run record — LLM, gen AI, skill, software, environment, services contacted, turn time, turn tokens, status. Rank: votes roll up into leaderboards of setups, not just models ("Setups, not just models." — same model, different setups). Explore: community Outputs, Tasks, Skills, Compare, Community, Leaderboards. Real model names visible in the product: Claude Opus 5.5, GPT-6 Astra. Show only names/numbers you actually see in your captures. End card: Evals mark + wordmark, "Stop guessing. Battle it.", URL evals.ag. 1. Reverse-engineer the reference first (frames are the truth) Download the reference and study it frame by frame. Write SPEC.md with measured numbers, not impressions: every cut (frame number), shot order and durations, average shot length, cuts per 10s camera moves per shot (start/end position, scale, curve), easing measured per frame (fraction of remaining distance covered each frame) type-on speed (frames per char), caret behaviour, word-by-word reveal spacing, new-word accent flash duration card/UI choreography (fly-in paths, stacking, depth), light↔dark section switches, split-screen layouts sound map: where every hit/whoosh/click lands relative to the cut Then rebuild shot for shot at the same rhythm. Only brand, copy, design and assets change. 2. Scenario Write SCRIPT.md before building: a fresh narrative arc true to what Evals really does (hook: a new SOTA model every week — which one is actually best at your work? → type a task → Battle up to 4 setups → blind side-by-side, vote → full run record → setup leaderboards → "Stop guessing. Battle it." evals.ag). Don't reuse the previous films' copy. Short punchy lines (2–5 words per beat), English. Every beat gets a timestamp on the music grid. 3. Design Write DESIGN.md once and make every section obey it (no per-section drift): palette (one accent per frame), type pairing (one sans for headlines; weights/sizes in px; tracking −0.02em, line-height 1.0) card treatment: real UI captures, rounded 14–18px corners, soft layered shadow Never: outline/hairline boxes, rules, underlines, highlight boxes behind words, cheap gradients, emoji, fake UI mockups, invented numbers/rankings/features. 4. Assets (all real, all new) Capture the real Evals UI with Playwright at 2–3× DPR: home composer (empty / typed), Run-mode menu (Battle / Side-by-side / Direct), setups stepper (1–4), a live run page (4 agents generating → "Your options are ready"), Compare (blind A/B with "I prefer A/B"), output run record panel, Leaderboards (setup rows), Explore Outputs/Tasks/Skills. Run one real 4-setup Battle on a visual prompt (e.g. a playable browser game) so the four results are genuinely the same task. Record real Evals outputs as clips from their live previews (screencast, 1600×900, 30fps, 6–7s each; press start/arrow keys or orbit so they move). Use outputs not used in earlier Evals films. Typing scenes: put a text layer exactly over the real input field, matching its font, size and colour. New music track (detect BPM and first downbeat with a script) and a fresh SFX set. 5. Motion rules (this is the 감도 — non-negotiable) One camera = one pure function of time cam(t) built from keyframes with velocity-continuous easing; applied from the master timeline. Deterministic: no timers, no randomness, any frame renders alone. Two easing families only: snappy ≈25% of remaining distance per frame (90% in 8–9 frames) and floaty ≈12% (18–20 frames). No overshoot unless the reference has it. Exits accelerate into the cut; the next shot carries the motion through it (velocity-matched). Real whip-pans: outgoing and incoming shots travel together over 0.30–0.45s, with velocity-driven motion blur that peaks mid-move. Never blur switched on over a hard cut. Overlays are screen-locked (never inherit the camera whip). Text arrives with its container in one mask reveal. Overlapping action: the next element starts at ~60% of the previous tween. No start-stop chains. Nothing freezes: every hold carries a 1–3% slow drift (sine.inOut across the whole hold). Fast moves decelerate over ≥6 frames — no dead stops. No double cuts (two visual changes on adjacent frames). Busy footage gets only a slow push; energy comes from the cuts. Word reveals 2–4 frames apart; the newest word flashes the accent, then settles 3–12 frames later. Hero type-on 1 char / 2 frames; in-product typing 0.6–0.9 char/frame; caret solid (never blinks). Every cut on the beat ±1 frame; visual hit and SFX on the same frame. 6. Speed, edit, impact Speed by re-rendering, never frame-dropping: render at --fps 30/speedup and play back at 30fps (e.g. --fps 95/4 → 1.263×). Pick the speed-up so one beat = a whole number of frames. Quantize every segment to 8th notes; round segment start frames up. Impact FX in post (ffmpeg) on 4–6 big hits only (cut to dark, reveal word, flash, key word, logo): zoom punch +8.5%, decaying τ = 0.09s crop shake 16px at ~22Hz, decaying τ = 0.1s rgbashift 16px for 2 frames, then 7px for 2 frames 1–2 frames of white flash Lighter version on secondary cuts (3.5% punch, 6px shake, 1 frame of split). Find hit frames from frame-diff peaks and by looking. 7. Sound Music bed carries the film; no ducking. SFX bus ~3dB under the music overall, above it only on hits and in the music's own gaps. At each big hit, layer reverse swell + boom (high-passed at 60Hz) + short transient + glitch synced to the RGB split. Use a modest total (~40–45 sounds); never spread SFX evenly on every cut. Master at −14 LUFS, AAC 192k. 8. QA before you report (mandatory) motion.py <render> <bpm> <first_beat> must show: no dead stops, no frozen holds, every cut within ±1.5 frames of a beat, no adjacent-frame cuts. strip.sh (16 consecutive frames) at every transition, and look at each one. A good whip shows a smooth velocity ramp with both shots visible mid-move. Contact sheets alone hide motion bugs. Check text readability, safe margins, and that every number/name on screen exists in the real product. Deliver renders/evals-<name>-v1.mp4 plus a report: script, design summary, asset list, shot list with times, BPM and speed factor, hit frames with FX, SFX map, QA results.LLMGPT-6.1 SolGen AI—Skill/hyperframesSoftware—221.7K tokens·7m 9s00
Output B, 1 turnECEvan ChaOct 6Make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464 You must reference this site https://evals.ag/. You must use font sanserif. Motion must be speedy. Never make ai slop.LLMClaude Opus 5.5Gen AI—Skill—Software—4M tokens·15m 34s00
Output A, 1 turnECEvan ChaOct 6Make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464 You must reference this site https://evals.ag/. You must use font sanserif. Motion must be speedy. Never make ai slop.LLMClaude Opus 5.5Gen AI—Skill—Software—5.6M tokens·17m 30s00
1 turnECEvan ChaOct 6make a viral video in tiktok trending!LLMGPT-6.1 SolGen AISeedance 2.5Skill/ai-video-generationSoftwarefal.ai680.3K tokens·8m 49s00
Output A, 4 turnsECEvan ChaOct 6make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464?s=20 what you make it is https://evals.ag/ siteLLMClaude Opus 5.5Gen AI—Skill/hyperframesSoftware—27.8M tokens·1h 3m 36s00
Output B, 1 turnECEvan ChaOct 6Build a single-page landing site for a coffee subscription startup with a hero, pricing table and signup form.LLMGPT-6.1 SolGen AIGPT Image 2Skill/design-taste-frontendSoftwareChatGPT Images922K tokens·8m 10s00
Output A, 1 turnECEvan ChaOct 6Build a single-page landing site for a coffee subscription startup with a hero, pricing table and signup form.LLMClaude Opus 5.5Gen AI—Skill/design-taste-frontendSoftware—2.8M tokens·7m 20s00
Output B, 2 turnsECEvan ChaOct 6make a demo intro video reference this video https://x.com/rajsinghfirst/status/2107161801322549725?s=20 demo site is https://evals.ag ! this siteLLMGPT-6 AstraGen AIEleven Sound Effects v2+1Skill/hyperframesSoftwarefal.ai4.4M tokens·29m 22s00
Output B, 2 turnsECEvan ChaOct 6make a viral video reference this thing https://x.com/motion_conquest/status/2106786068712558848?s=20 what you need to make it is https://evals.ag/ site! demo videLLMClaude Opus 5.5Gen AISeedance 2.5Skill—Softwarefal.ai15.1M tokens·31m 28s00
Output A, 1 turnECEvan ChaOct 6make a demo intro video reference this video https://x.com/rajsinghfirst/status/2107161801322549725?s=20 demo site is https://evals.ag ! this siteLLMClaude Opus 5.5Gen AI—Skill—Software—5.3M tokens·19m 46s00
Output B, 1 turnECEvan ChaOct 6make a viral demo video reference https://x.com/titouangillet_/status/2107016017259921464?s=20 what you make it is https://evals.ag/ siteLLMGPT-6.1 SolGen AI—Skill/hyperframesSoftware—1.7M tokens·8m 51s00
Output A, 1 turnECEvan ChaOct 6make a viral video reference this thing https://x.com/motion_conquest/status/2106786068712558848?s=20 what you need to make it is https://evals.ag/ site! demo videLLMGPT-6.1 SolGen AI—Skill/ai-video-generationSoftware—316.5K tokens·6m 51s00
