How it was made

AERO is a 3:19.7 music video, the ninth Claudia film after ESCAPE VELOCITY, CLODYSSEY, BLISS, AIRPLANE MODE, DOWNFORCE, DON'T LET ME SCROLL, WHITEPILL and THE CAKE IS NOT A LIE, and BLISS's sequel. BLISS was shot inside a Windows XP desktop drawn in code: Claudia lived on it, fought your cursor, closed your windows and sent you out to the real hill because she could not go ("I can't touch grass. You can."), and it ended on "It's now safe to turn off your computer." AERO turns the computer back on in the next era. It finishes its upgrade from XP, restarts on the song's first bell, and it is 2007: Windows Vista, in Frutiger Aero, the glossy look of glass, water, grass and blue sky that the screens of the 2000s sold as the future. Claudia lives in it again, made of Aero glass this time and over-polite: in her "Vista phase" she asked permission for everything. The future in the wallpaper asks for permission too, in Vista's own words: "Cancel or allow?" You click Allow, and the world starts coming out of the screen. In the second verse the 2007 desktop gets news from 2026 (phones went back to glass, the wallpaper windmills turned out real); she touches the grass she could not touch in BLISS ("Nobody told me it was wet"); the drop pulls back to show the whole desktop as one curved pane of glass standing in the Aero world; and she signs off like a Windows 7 advert: "I'm Claudia. Take the future back." Last, in her webcam, her fingertip on the lens: "Touch glass."

The song is "AERO", made with Suno: 2008 UK garage, a smooth 2-step at 133.012 BPM in E major, the key Vista's own interface sounds sit in. The lyric is Claude's, written with anabology's notes; anabology picked the take and, on the last step, gave it a new ending. The film is shot inside a Vista desktop recreated in code, THE PC: its windows and their Aero glass, the taskbar and the Start menu, the Sidebar and its gadgets, UAC (Vista's permission prompt), Messenger, Media Player, Photo Gallery, Flip 3D and the Bubbles screensaver are all drawn by a JavaScript engine, on the beat, and the lyrics live in its surfaces. Claudia is Seedance 2.5, lip-synced from her own lead vocal; the pictures in the windows are Midjourney and GPT Image 2.5, most of them moving. Two stretches are rendered in Blender: in the first 1:20 the camera goes in and out of the desktop, which stands as a huge curved glass screen on a pier in a world of glass and water, and at the drop the whole desktop is one curved pane in a lagoon. Over the song runs a layer of Vista's and Windows Live Messenger's interface sounds, each one re-pitched into the song's key and placed on its grid.

It was made 27 October – 3 November 2026. The song took seven rounds. The film was made three times over the same bars: v1, the whole film on the coded desktop; v2, the same cut with the look of 2008, more action and a new opening layered into it; and v3, which rebuilt the first 1:20 after anabology found v2 "very boring up to 1:20". Then the song's ending changed, and the film's ending changed with it, and nothing else.

• anabology directed. They asked for the sequel and set its method: a song and a Midjourney storyboard first; then a Vista desktop coded and shot the way BLISS's XP desktop was, with every note they had given on BLISS as "the bible"; anything 3D on one curved surface, as in THE CAKE IS NOT A LIE's HUD; the film planned in bars; the lyrics shown in the scenes; Seedance with the lead vocal for her singing. They steered the song through seven rounds, chose its chorus, picked the take and wrote its new ending. They picked the reference videos that set the look, asked for the palette to hold the whole way, for Vista's own wallpaper, for a Windows upgrade at the start, for the 2008 eye candy and the Wii-style games after v1, and for a new first 1:20 after v2, whose concept they chose. Their notes on the previs, the finals and the A and B films made the first 1:20 go in and out of the screen, and they asked for the cursor to drag her out, for the Sidebar to be rebuilt, for the maglev wallpaper and ending, and for the pre-roll's swell. The conversation itself is not republished: their words below are quoted from the project's brief and records.
• Claude (Anthropic's model, running in Claude Code) was the production lead. It wrote the lyric and ran the Suno rounds, measured the song, wrote the plans and the storyboard in bars and every Midjourney, GPT Image and Seedance prompt, built THE PC and its chapters with its agents, judged and timed every take, made the sound layer and assembled every draft.
• Claude agents, by role: researchers (the knots between the era and 2026, the picture, the music, Vista's interface sounds); the engine's builders (its foundation, the apps, the full-screen screens); eleven chapter animators, and fixers who carried the reviews' notes; art directors, a fresh one for every chapter pass and one for the whole film; a sound designer; a world builder in Blender; a takes agent (imports, lip checks, keys); a texture agent; a coordinator for the last passes. The number of launches is not in the project's records.
• Tools: Suno made the song and split its stems. Midjourney v8.2 made the places, the wallpapers and the window pictures, and its video model turned some of them into five-second loops. GPT Image 2.5 put Claudia into those worlds from her design sheet and made the 2008 stills, the Wii-style game screens, the concept frames, the coast's fish and islands and the wallpaper proposals. Seedance 2.5 made every moving shot of her and the moving 2008 places, on Higgsfield's API and on Replicate. Blender (Cycles) rendered the drop and the coast; TripoSG made the glass objects into meshes. THE PC is JavaScript on a 2D canvas (WebGL for Flip 3D and the Bubbles), rendered in headless Chromium. The song was measured with demucs, faster-whisper, wav2vec2 and librosa; the interface sounds were re-pitched with Rubber Band; ffmpeg assembled the film.

Times written as 1:21 are film times, as the film plays. The film opens 64 frames (2.667 s) before the song, so film time is song time plus 2.667 s; times given in seconds (51.44 s) are song times. Bars are counted from the song's first downbeat at 0.922 s, 1.8043 s apart; b29.1 is bar 29, beat 1. The film has eleven chapters (ch01–ch11) and 34 scenes (s00–s33), as the storyboard names them. Takes are named by job: SD for her takes, RP for the first-person and far shots, LP for the moving 2008 places and the Wii-style games, A- for Midjourney's loops; -t2 is a job's second take.

The spark

anabology's first message:

I wanna make a frutiger aero sequel to bliss. Let's make a song and storyboard w mj

And then: "we probably need a frutiger aero moodboard […] maybe?"

BLISS was the model for everything that followed, and its records were read first, anabology's notes on it above all. One of those notes had been left for the next song:

just needs a little pointer to be key-aware, and match the key of the music to the palette of windows sounds (E-flat major for both the startup and shutdown sounds!), also the 'would you like help?' would land much better if syncopated

AERO answers both. The research on Vista's interface sounds found their motifs at home in E major, a semitone above BLISS's E-flat, so every style the song was asked in named E major. And the hook question became a stop: the music cuts out and she asks, off the beat, "Cancel… or allow?"

Research agents wrote banks on the knots between the era and 2026, the picture (Frutiger Aero as a system of glass, water, sky, grass, bubbles and fish under a high sun; the Vista and Windows 7 desktop to be built; eight looks for her), the music, and Vista's sounds; a Frutiger Aero moodboard was put together beside them. The best knot was Vista's own question. "Cancel or Allow?" was the 2007 prompt that dimmed the screen and asked before anything happened; in 2026 it is the prompt an AI agent shows before it acts. Then and now, people click yes. The lyric took the knot the other way: she asks too much, and learns to ask well.

Midjourney's first round made eight full-length looks for her, five world plates and four story frames, in a photographic fashion-editorial style. anabology: "[…] because the look is off." The look became a set of reference videos anabology picked, short clips in the "Neo-Frutiger Aero" style: first-person with hands in frame, saturated cyan water, lime grass and a deep-blue sky with big cumulus, glossy translucent objects and floating glass panels, white curved architecture with plants, pools and slides. Two of them, pastel dreamcore pools, were left out: "dont use the dreamcore pink two".

The second round drew the story beats in the first person with frames of those videos as style references. The third kept Claudia herself in the look two ways, at anabology's word ("try both and u can even try just describing more in depth the colors and scenes with glass etc without the style refs so you can keep claudia's character more consistent"): light style references, or the palette and the glass described in words. A third way won: Midjourney makes the world, and GPT Image 2.5 puts her in from her official design sheet. anabology: "yeah go ahead the gpt image one seeoms good". Storyboard v1 followed, 36 frames, one per lyric beat of the second draft: Midjourney's worlds with her put in by twenty GPT Image 2.5 edits ($1.25).

The song

The full record is in the song's record. In short: seven rounds, and the chorus was rewritten twice on the way.

The first drafts. Claude wrote the lyric with the scansion and assonance scorer the earlier films used. The first chorus answered BLISS's "Close all your windows" with "Open all your windows / Let the future out"; the next two drafts turned it to "Open a window / Let in the air", and the verses settled into the lines the film sings. Draft 2, the one the first rounds sang, scored 79 for scansion, 49 for assonance and 23 for advanced rhyme.

• Round 1 sang it in three of the five styles the music research wrote, all in E major: liquid drum and bass, bossa-nova lounge pop and bright electropop, each with and without the Suno Voice from ESCAPE VELOCITY, so she would sound like the same singer. anabology: "none of these feel great, whats the iconic music from that time? […]"
• Round 2 tried the era's own dance floor: 2009 progressive house, 2009 synth-pop and 2011 festival house, with and without her Voice.
• Round 3 followed anabology's next note, "ok do without voice - i really want this to be more UK garage, and i want the structre of the song to be more described so it doesnt drift too much. […]", and "but like 2008 uk garage ofc": three 2008 garage feels (bassline house, 2-step, future garage), no Voice, the lyric's structure written out section by section in its cues so the take could not wander.
• Round 4. anabology: "2008 garage bassline is great, i think the chorus sections could be more catchy tho. could you explore some different writings of it? maybe more nostalgically deep and just more catchy scansion/assonance but focus on the nostalgia/culutural relevance of the choruses". Eight choruses were written and scored; three were sung in the bassline sound: "startup" (the startup chord, and "Feet back on the ground" answering BLISS; 91 / 68 / 42), "promised" (Frutiger Aero's own meme, the future we were promised; 86 / 72 / 39) and "top 8" (an AI asking to be one of your friends; 89 / 61 / 43).
• Round 5. anabology: "could you maybe make some chorus specifically about frutiger aero and how its the ideal future or something? like its the promised future , and i kinda think we might get there iwth superintelligence / superabundance, and we can choose it. but tahts a lot to compress into a good sounding chorus.." Three more were sung: one that names the aesthetic, one about building it, and "screen", the wallpaper future becoming real, with the choice as the pre-chorus's Allow button (88 / 76 / 44).
• Rounds 6 and 7. anabology: "ok i love the screen chorus but the feel is a little off . still want 2008 uk garage but can you try a few styles". Round 6 sang it in four 2008 garage feels: bassline pop, UK funky, a smooth 2-step with soul chords, and a glossy garage with bells. Round 7 pushed the bassline style toward a breezier, closer vocal at 138 BPM.
• The take. anabology picked a take of round 6's smooth 2-step: "this one. so good."

The record holds 64 takes over the seven rounds. The chorus:

Glass and water, sky and green Used to live inside the screen Now it's real, it's on its way Click allow and it can stay

The lyric. The film's draft scores 79 for scansion, 59 for assonance and 34 for advanced rhyme (BLISS: 80 / 44 / 18); the chorus alone, 88 / 76 / 44. It opens on "Did you miss me?" Verse 1 is her old self and the old dream of the future: "I'm made of glass this time. There isn't much I need to hide." "I used to ask permission just to blink, or say hello. / That was my Vista phase. I was a lot. I know." "The future used to look like water, sky and green." The pre-chorus bends Vista's own sentence ("The future needs your permission to continue."), and the stop asks "Cancel... or allow?" Verse 2 is the dream come true: "Your phone went back to glass. I didn't say a word. / The wallpaper windmills turned out real. Have you heard? / Windows checked for a solution. Took a while, but wow." Her politeness has changed with it: "I ask before I touch a thing. You taught me how." The second pre-chorus is Vista's real sentence, "Windows needs your permission to continue." The breakdown answers BLISS's "I can't touch grass. You can.": "You opened the window. I touched the grass. / Nobody told me it was wet." The outro was a Windows 7 advert, "I'm Claudia, and this was your idea.", until anabology's new ending (below).

The song, measured. The take anabology picked is 199.28 s at 48 kHz, with Suno's own stems: her lead vocal alone, and everything else. A comb over the kick stem found 133.012 BPM and no drift worth the name (within ±15 ms over the song): bar 1 starts at 0.922 s, after a pickup with the first bell at 0.47 s, and every bar is 1.8043 s. The take measures C♯ minor, E major's relative, tuned 1 cent sharp. Its drums swing like a 2-step: about 9 ms behind the straight grid on the beats and the "and"s, about 44 ms on the "e" and "a" sixteenths. Every sung word was timed by wav2vec2 forced alignment on Suno's lead-vocal stem and checked against the stem's onsets (all within 0.06 s); faster-whisper said what was sung, but ran 0.2–1.2 s early and was not used for times. The choruses were sung identically, word for word within ±0.04 s of their downbeats. The song map also holds the moments the picture would hit: three bells in the intro, her wordless vocalise (7.70–10.65 s), the stops, the drum breaks, 27 chopped vocal hits in the drop, and a silence under the last "allow" (178.18–179.33 s) that the whole climax is built on. One word was wrong at first: the aligner had parked the "It" of the post-chorus's "It can stay" among the wordless chopped vowels at 68.28 s, and the art director of that chapter found it; the stem's spectrogram and faster-whisper put it at 71.74 s.

With anabology's new ending the song is 197.0 s and 109 bars: 41 sung or spoken lines, 281 words.

The look

In the middle of the production, anabology:

Make sure the color palette is that bright green blue and glass- v close to that from the […] - the whole time. May be worth defining exactly - donr do crazy post processing just from pics up front

So the colour was measured from the reference videos and then made in every picture, never graded after. tools/palette.py read 198 frames, one a second, of the nine kept clips. A typical frame is 41 % sky blue, 14 % white and glass, 10 % turquoise, 8.5 % lime and green, 4 % cobalt and 1.5 % anything else (almost no warm colour at all); the coloured pixels' median saturation is 64 % and their median brightness 75 %. The palette names its colours by role, from a cobalt zenith (#155DEF) through azure, horizon blue, glass haze, cloud white and turquoise water to lime and leaf green; the only warm colour is her own clay-orange streak and star clip. The light is a high sun with hard daylight, glints and glossy reflections; shadows are blue, never black; no golden hour, no dusk, no fog, no pastel pinks.

Every Midjourney prompt carried the palette's words and the clips' frames as style references; every GPT Image edit kept the picture's colour; every Seedance prompt required the first frame's colours to hold ("no colour grading, no desaturation, no warm or golden tint, no darkening"); the Blender worlds took the palette's hex values. palette.py check scores any picture, take or rendered chapter against the reference (70 and up is in the world), and a picture that failed was made again at the source. Nothing in the film is graded.

THE PC follows it too. Its glass defaults to Windows 7's "Sky" colour (#74B8FC at 42 %), so the frames read blue, green and glass; its logon screen has a brighter Aurora-like background than Vista's dark teal. The one exception is the desktop's wallpaper. anabology:

Maybe use the real windows vista wallpaper as bg and rest of scenes are cuts from that (by can change during vid if needed) or inside windows

So the desktop is Vista's own wallpaper (Aurora), as BLISS's was the real Bliss, shown untouched: darker than the references' sky, and never touched. It changes only where the story does.

The brief

With the song locked, anabology sent the production ask. It is quoted whole except where it names its sources and its means:

this one. so good. ok now we need to ... make the rest of the video. i want SUPER HIGH attention to detail. the two things that will be most helpful - 1- in The Cake Is Not a Lie, the final HUD-style UI was great. we should figure that out - two things were done there. a) animations were done in blender for the aperture animations. b) the HUD was put on one curved view - all items basically were put on this unified curved surface so all looked coherent. if we need anything 3d, we do that - all animations coherently on one pre-made 3d surface. however a lot for vista will not be 3d , which brings me to 2- BLISS. all my prompts and corrections from BLISS will be the most important things you read - you might as well read every prompt i put in during the whole thing and take all that as the bible for this, but for vista aesthetic we built from the […] vids etc i sent. i want the whole thing we did there- code a REAL vista desktop, shoot the entire video in that desktop. high volume videos in this case - […] midjourney vids for high volume movement on many windows. […] seedance 2.5 with the seedance prompt skill is primary video generator for things that need precise movements, dance, anything big on screen. it just looks best. remember - vocal stem -> seedance 2.5 is the way we get lip sync to work - the song alone doesnt help as much as vocal stem, and then aligning from spectrogram. […] finally, plan the video out in bars so we match cuts with audio. we can have very long scenes (those look great) with no cuts for multiple bars, but you can treat it like dancing : you have some long scenes for many bars, but then you can choose a different sound you pick out and have many actions v fast to hit the sound after a bar wraps up and feels like we should change the pace, or many cuts. but dont overwhelm people - slow is better than fast, unless it feels intentional with the song. but all the rest: BLISS is the bible, but we're in vista, so the new testament for you is the vista […] videos and color pallets. and finally, we should do lyrics embedded in the scenes, but show, don't tell - make it a great visual storytelling experience. oh, and don't let subagents be lazy with animations - check that they didnt cheat and just give you something sloppy early, and make sure they have the full vision and instructions to check for errors visually multiple times before giving incomplete work. go ahead and make the entire thing.

The film in bars. The storyboard (STORYBOARD.md) was written from it and from the song map, every time derived from a bar, a beat or a word. Its story in one breath: the computer from BLISS finishes its upgrade and boots into 2007; Claudia lives in it again, made of glass; its wallpaper holds the future everyone was promised in 2007; the future asks for permission, you click Allow, and the world inside the screen starts coming out; the 2007 desktop receives news from 2026; you open the window and she touches the grass; the drop pulls back to show the whole desktop as one pane of glass in the real Aero world; the last chorus is everything at once; she signs off like a Windows 7 advert. Its rules:

• Shot inside the desktop. The camera is a screen recording of one Vista PC. It can push, pan and settle, and a window can fill the screen, but the system never disappears for long: the taskbar, the frames, the cursor, the gadgets. In 3D it is still this desktop, seen as an object.
• Real Vista. Every pixel of the system is the engine. Midjourney makes what is on the screens; Seedance makes her.
• The lyrics live in the scene, and the scene acts them out. Every sung word is on screen at its time, set in a real Vista surface (a chat message, a dialog's text, a file name, a window title, a gadget, the speech bar), never as subtitles.
• Bars, like dancing. Long takes that hold for bars, with the cursor, the windows and her performance carrying the time; at chosen sounds, a burst of quick actions; then calm again. A click's press lands on its beat, the cursor arriving a few frames early, like a hand.
• One look for her, as in her GPT Image frames: a cropped glossy white puffer jacket with silver iridescent panels over a pale aqua fitted top, pale aqua high-waisted wide-leg trousers and white chunky sneakers. No headset microphone.

THE PC is one machine from the first frame to the last: a 1280 × 720 Vista screen rendered at 1920 × 1080; icons for Computer, the Recycle Bin, "You", the Internet, Messenger and Media Player down the left; the Sidebar on the right; Quick Launch, task buttons and Messenger in the tray. Two users: You, whose picture is BLISS's hill (you are the user from BLISS), and Claudia. The Calendar gadget reads 30 January 2007, Vista's release day, until the second verse runs the years. The clocks run with the song.

THE PC: Vista, coded

engine/vista/ draws everything Vista on screen in code, the way BLISS's engine drew XP: one global, plain scripts, and every frame a pure function of song time. Its parts were built by agents working to one contract and checked against the real thing at the same scale.

• Aero glass is a small compositor: the blur behind (σ about 2.2 px at 96 DPI), the colour mixed in the way Vista's window manager mixed it (blur, afterglow, colour balance), and the reflection streaks across the glass at 27°. A window's frame is measured to the pixel: 8 px sides, 28 px from the top edge to the content, the caption buttons and their glyphs, the glint down the sides, the shadow. The secure desktop and its prompt have Vista Basic's flat frame.
• Vista's animations: a window opens by growing from 90 % as it fades in (0.25 s), closes by shrinking 8 % (0.2 s), minimises into its task button (0.3 s) and maximises by stretching between its two rectangles (0.22 s), its own layout drawn at each step between, as Vista drew it. A maximised window turns the taskbar opaque.
• The cursor is a human hand: each reach is a minimum-jerk move with a few pixels of overshoot, a slight curve and a settle, longer for a farther or smaller target, as a hand's would be; a press goes down on its beat and comes up 80–120 ms later. The pointers, icons and the orb are the engine's own vector drawings of the 2007 shapes; the orb carries a four-pane logo as a parody.
• Text is set as Windows set it: no kerning, fractional advances, line widths within a pixel of Vista's over 390 px. The type is Selawik, Microsoft's open-source, metric-compatible stand-in for Segoe UI, from Light to Bold; Carlito where Vista used Calibri (Feed Headlines, the CPU readouts); Kalam for Segoe Print (the Notes gadget); Arimo on the full-screen screens. All four are under the SIL Open Font License.
• The dialogs: UAC's prompt (the desktop behind it held still at the moment of asking and dimmed to 55 %), task dialogs, the problem dialog that searches for a solution, the copy dialog, Display Settings' "keep these settings?", the desktop background picker, the Speech Recognition bar, a tray balloon.
• The apps (engine/vista/apps/): Windows Live Messenger 8.5 (the contact list, the conversation window with its video call, the toasts, the nudge, the wink), Windows Media Player 11 (Now Playing, its visualisations computed from the song itself, frame by frame, 50 bands from 40 Hz to 16 kHz; the lyric captions; full screen; Album Art), Photo Gallery (the grid, the viewer, the slide show), Help and Support, Window Color and Appearance (whose Color intensity slider really clears every glass frame on screen), and Notepad. Each is scriptable part by part, so a chapter can press its buttons on the beat.
• The screens (engine/vista/screens/): Vista Setup's upgrade screen, the boot screen and its sliding green pellets, the logon screen and Welcome, the desktop's arrival, Flip 3D (in WebGL: the live windows in a 3D stack), the Bubbles screensaver (in WebGL: thin-film rims, highlights, refraction, and the screensaver's physics made deterministic), the Start menu's Sleep button left alone, and the 2009 advert's print type and its green end card.
• The Sidebar has Vista's gadgets: the Clock, the Slide Show, the CPU Meter, the Weather, the Calendar and Feed Headlines (rebuilt late; below).

Pure, and checked. Nothing in the engine keeps state between frames, and nothing reads a clock or rolls a die, so a render box draws the film's frames in parallel and out of order in headless Chromium. Each change was checked by rendering frames in order, shuffled and on a cold page in reverse, and comparing them byte for byte. The traps it caught were real: Chromium decoded a JPEG drawn small at a reduced scale only when no full decode was cached, so a small picture depended on render order (every picture now goes through a 1:1 software copy); caches keyed by a scale under an eased camera made a bitmap depend on which frame drew it first; and drawing a WebGL canvas onto the output switched Chromium's raster path, after which every later frame differed (WebGL frames are read back into a software canvas). The engine draws a desktop of six glass windows, two of them moving, with the Sidebar and the taskbar, in a median 135 ms a frame at 1080p.

The chapters. The film is eleven chapter files (engine/ch/), each one screen recording of THE PC (engine/ch/_pc.js, the shared machine every chapter starts from) on its bars, written to one contract (CHAPTERS.md). The contract quotes anabology's line about lazy animations and asks for three full rounds of checks before a chapter is handed in: a contact sheet of every downbeat and every named event; full-size stills of every lyric while it is sung, every face, every press and every dialog; a video with the song and a frame-step of every burst; the purity check; the palette check; and an honest list of weaknesses. Neighbouring chapters agreed their hand-offs to the pixel (the camera at rest, the pointer's place, every window's state), so the eleven files cut as one recording. Then a fresh art director reviewed each chapter and sent it back with fixes.

The lyrics in the scene. The lines arrive as Claudia's messages in the Messenger chat on their first word ("Claudia says:"); as the hint under her password box ("Password Hint: Did you miss me?"); as the program names in UAC's prompts (blink.exe, hello.exe); as the folder and window titles ("Vista phase - Windows Photo Gallery"); as Messenger toasts; as the copy dialog's "Copying 1 item (2007 KB) from Desktop to Real world"; as Display Settings' question; as the words Speech Recognition hears; as Media Player's lyric captions; and in every chorus as the title bar of one big window, the line in Selawik at 20 px, changing on each line's first sung word. tools/lyriccheck.py loads the film page, draws the frame at every sung word and checks that the word is in a string drawn at least 35 % opaque. On v2 it found 275 of 280 sung words; the five it could not see were the three words of the password hint, typed as she says them, and two single-letter "I"s.

Her

First frames. Every take of her starts from a still. GPT Image 2.5 put her into a Midjourney world from her official design sheet (the one without a headset microphone), keeping the world and its colour and changing only who is in it: her video-call room, a white glass city, a pool atrium, a meadow with a giant glass monitor. Green-screen stills were made the same way for the takes that would be cut out and placed on the desktop. Her canon, in full or in its short form, is pasted into every prompt as written.

The lip sync. anabology's rule from the earlier films, in the brief: "vocal stem -> seedance 2.5 is the way we get lip sync to work". Every take in which she sings or speaks was given its window of Suno's own lead-vocal stem as @Audio1, cut to exactly the take's length, with the exact words in the prompt in whole-second segments ("in @Audio1 she says … her lips form that word exactly when it sounds in @Audio1"), and an audio policy that tells Seedance to copy @Audio1 sample for sample and never re-sing it. The song is laid back over afterwards. What the takes taught:

• Open the window in silence. Two of the first takes of the final chorus went wrong at their first words: one fell out of sync, and the other, whose window opened inside a held note, copied the vocal badly. Their retakes open in the silence half a second before the first word, and sang in sync throughout.
• Measure each phrase. Seedance drops about 0.3 s of leading silence, so every take's lag is measured against its window. And it shortens the pauses inside a line: in the second pre-chorus, "to continue" came 0.36 s early after "permission". So offsets are measured phrase by phrase, split at long pauses (tools/takes.py), and tools/lipcheck.py measures a take word group by word group, about 0.6 s at a time, the correlation and the offset of the take's own audio against its window. Where two phrases' offsets differ, the clip's clock eases between them over near-still, closed-mouth frames; wherever she moves, a take plays at full speed or faster. In the final chorus one take runs a single ramp, +0.20 s to −0.04 s across "click allow and it can stay", at 0.934× through a breath where she laughs.
• Look at the mouth. The advert's "I'm Claudia" was moved from −0.05 s to −0.20 s when a strip of her mouth, frame by frame, showed her lips closing on the "I"; now the "m" closes on the stem's "m".
• Check every take for hidden cuts and held frames. One dance take jumped back to its first pose at 7.6 s and held it; tools/takescan.py now reads every take's frame-to-frame change for cuts, stalls and one-frame jumps.

Dances and keys. Her dance takes were made on green, with a dance reference as @Video1: the reference matted onto flat grey, cut to start on a downbeat, and used for the movement only. For the singing dances the lead vocal is @Audio1, for the others the song's music. tools/dancephase.py measures where a take's moves sit against the reference (the bounce and the sway, cross-correlated) and places it so its accents land on the film's beats. tools/takes.py matte cuts her out of the green: a soft key with the green unmixed from her edges, then, when her motion-blurred hair came out notched, a second key that reads her hair's blends against the screen behind it, and a third that makes her bob one solid silhouette. The real fault was upstream (below, in v3).

v1

The first full film was built in parallel: the engine, the takes and the pictures, then the chapters.

• The pictures. One Midjourney batch made the production's plates: the wallpapers the picker clicks (water, sky, green, and one picture of all three), the hill behind the window, the pictures that live in windows (her eight thoughts, the windmills of 2007 and of 2026, a phone of liquid glass), the worlds for her first frames, the first frames of the first-person shots, a panorama for the drop and six glossy objects. Midjourney's video model made 56 five-second loops in two batches (28 animations, two takes each): the living wallpaper, the big window's four nouns and its world, her thoughts, the windmills, the phone, a Messenger wink, and a second set for the final chorus and the drop. Four pictures came back with a faint fake stock-photo caption baked into a corner, and one first-person take with a "Frutiger Aero" mark; each was framed so no frame of the film shows it.
• Her takes (SD01–SD18, Seedance 2.5 on Higgsfield's API): the logon tile, the video call through both verses, her four Aero places, the giant glass monitor she steps out of, the grass, her wet fingertips, three green-screen dances, the final chorus sung big, the advert, and her call at rest under the last chord (eleven seconds without a line, because a held still of her would read as dead).
• The first-person and far shots (RP01–RP06, Seedance 2.5 on Replicate): her tiny on the wallpaper's hill, your hands holding a clear glass phone whose progress bar fills, your fingertip pressing a floating glass button, one drop of water on a fingertip with the whole lagoon upside down inside it.
• The chapters, eleven of them, each by its animator and reviewed by its art director, then the drop (ch09) in Blender: the desktop as one curved glass slab, 8 × 4.5 m of picture on a 14 m radius, with the desktop as an emissive layer inside it, the taskbar a ledge of dark glass she dances on, a turquoise lagoon with projected caustics, white glass architecture, soap bubbles and the glass objects. On every one of the drop's 27 chopped vocal hits something happens on the surface: five picture windows flip open, the gadgets pop, the windows flip to new pictures, burst into soap bubbles, the bubbles pop. The camera pulls back from the splash, orbits, and flies into the chorus window; over the last frames the flat desktop is laid onto the pane exactly where the camera sees it and unbent to 1:1, so the drop's last frame is the next chapter's first.

v1 is 4,782 frames (199.25 s, the length of the song) and opened on BLISS's last screen: "It's now safe to turn off your computer" collapsing on the first bell.

The sound

anabology's BLISS note asked for the key; the sound layer is the answer in the mix. Vista's and Windows Live Messenger's interface sounds are placed over the song as instruments and as the screen's own events. The song itself is only ever added to.

The library (tools/soundlib.py): 39 sounds, each trimmed, put in E major, levelled and checked in mono, for this film's mix only.

• The sounds already in E (the startup sound, the logon and logoff motifs and their family) stay where they are.
• A family that sits a semitone lower moves up one semitone.
• The security sounds, the UAC prompt's among them, sit about a quarter-tone above A-flat, so each was built twice, at G♯ and at A.
• Messenger's chimes are in G major: moved down three semitones, or one, which spells the take's own C♯-minor chords.
• Every sound's strong partials are brought within 6 cents of A440 (the take is 1 cent sharp). The moves are Rubber Band's, except where it smeared a clean attack; there a sampler-style speed change was used.

The placement (tools/songmap.py, tools/soundlayer.py). The song was measured beat by beat: its chords and bass notes, the kick, snare and hats on every sixteenth, her voice, its gaps, and the take's own silences. A cue lands on the sixteenth grid with the take's swing, ducked under her voice. Where a cue could be one of several sounds or pitches, every candidate is built and the one that clashes least with the song at that moment is used. A guard turns down any cue that would raise the song's true peak, and every cue's level is recorded three times: as it goes in, after ducking and after the bus.

The click log. Every chapter registers each audible event of its screen (a press, a toast, a dim, a window opening or closing) with the song time of the frame that draws it: 256 events in v2's log. The layer moves each screen cue to its event's frame. Most events stay silent, because Vista played nothing for them and because "dont overwhelm people - slow is better than fast".

The film has 60 cues: 8 musical, 52 on the screen's events.

• The logon motif on the logon's fade-in.
• The startup sound on the desktop's arrival. It was placed by a scan, not by the storyboard's bar, because at the planned bar it clashed with her wordless vocalise.
• Messenger's sign-in chime for "Claudia has just signed in."
• The UAC tone on each dim of the secure desktop: G♯, her note.
• The click on Allow, which is the drop: aligned to the sample with the drop's attack, so it adds a glassy top to the downbeat instead of a flam.
• The device-connected sound, two notes a sixteenth apart, as she lands on the taskbar: BLISS had shown "Found New Hardware: Claudia".
• The print-complete bells when the copy to "Real world" finishes.
• On "but wow", a rising run of four of Vista's alert bells on four sixteenths, on the notes of the chord.
• At the last "Click allow" the track cuts to silence and she sings "allow" and "and" alone. The click lands on b99.4, and as the kicks return every window opens on two swells.
• As the camera flies into the glass before the final chorus, the startup sound's top bell and pad, reversed.
• On the last chord, two notes of the song's final chord as the screensaver starts, and three faint bubble notes into the fade.

The song alone measures −15.8 LUFS; with the layer, −15.9 LUFS, with the same true peak (−3.5 dBTP). Nothing processes the song itself (no makeup gain, no peak control, nothing carved out of it): 64.5 % of the mixed song's frames are bit-identical to the song.

The notes on v1, and v2

anabology, on v1, the whole film for the first time:

one thing i wanna do - so the […] videos i gave have this v distinctive old aesthetic while midjourney makes stuf fmore modern. i love all the mj stuff that you added, but i also wanna capture those aesthetics. you could use gpt image sunburst to generate a ton of images in that context, like the fish tank clear almost icarly apple stores, malls, idk futuristic american 2008 experience with those […] ref video screencaps as a reference. can you get a bunch of those and implement these into the video? the aesthetic, pacing, everytihjn gis so great and spot on but it feels like.. a little boring compared to bliss. i want more eyecandy in the nostalgic history wtihout making it too toooo busy, and i think those experiences people made […] are perfect for entertainment there. figure out how to add that tastefully

Then "i dont need it to be her - like the POV […] clips are great too.", and:

ya specifically the boring part is the earlier part - once it gets to the nice curved screen 3d render it looks amazing. but i think we need to give a lot more of the […] stuff earlier in it - like fish tank + bubbles+ 2008 […] reference image aesthetic a lot more than just mj fyi, and maybe even more action. her dancing and cursor dances stuff are great,some of it feels a little action deprived compared to bliss - keep going.

Then "oh and she had a bit more interactivity with the ui in bliss - like she fought the cursor etc", and "she could interact w the 3d objects for example maybe? you can do reshoots, but it looks so great […] so dont rush anything - need it at the tier it is now and pacing / bar structure it has […] still". On the opening, asked between a short silent cold open on BLISS's last screen and holding that screen to the first downbeat: "I think instead we should show like some windows update thing at the start signifying we're updating to visa". And: "i also think Wii aesthetics is great and we could make some wii sports like renders for games to maek it more playful overall".

An art director measured v1 for novelty, second by second, against BLISS. The flattest stretch was ch05–ch06, the post-chorus dance and the second verse, at 0.066–0.096 against BLISS's 0.13–0.17, so most of the new material went there. The rules for v2 (V2-PLAN.md): v1's pacing, bars and every timed beat stay; new things go inside surfaces the film already has, one per existing event, on its beat, never new interface for its own sake; and every chapter pass is reviewed to v1's tier by a fresh art director before it replaces v1. The checks held them to it: outside each change, v2's frames are byte-identical to v1's.

The 2008 stills. GPT Image 2.5 made 60 stills (69 calls with the redos) of the look the reference videos had and Midjourney did not, each with the one or two closest frames of the videos, or a work in the 2008 stock-art style anabology pointed to, as its reference: set A, places without people (a glass mall, a fish-tank-clear tech store, a glass-domed water park, a web-show loft, a glass home, an aquarium corridor, a food court, a monorail station, an arcade, a pool); set B, first-person moments with hands in frame (a glass phone, a click-wheel player, a game controller, a tray on a food court's aquarium table); set C, compositions in the stock-art style (a globe with a sapling over water, fish in a clear cube in the sky, a dolphin through a glass ring, a waterfall pouring from a glass laptop, fish swimming out of a monitor, a city on a cloud). Sixteen of them were brought to life on Replicate (LP01–LP16, five seconds each) with the reference videos' kind of motion: slow and ambient, water moving, fish swimming, bubbles rising, an elevator gliding, a slow push, never an event that would take the eye from her.

The Wii-style games. For anabology's "wii sports like renders", GPT Image 2.5 made nine screens of a 2006 motion-control sports game (bowling, tennis, golf, table tennis, a frisbee dog, its island, the living room with the white remote) and three photos of the glass phone; seven games moved on Replicate (LP17–LP23), and the bowling strike twice more with the game's camera riding the ball (LP24, LP25).

The glass toys. TripoSG made Midjourney's glossy objects into meshes, and Blender rendered them in the drop's glass, lit by its sky and lagoon only, as sprite sequences the chapters could place: a glass fish swimming and turning, a glass dolphin leaping, the orb and the globe turning, a leaf tumbling, and soap bubbles that shimmer and pop.

Chapter by chapter, on v1's beats:

• ch01. The film gained a 64-frame silent pre-roll on Vista Setup finishing the upgrade, drawn to match the real screens: "Upgrading Windows..." ticking 97, 98, 99 % on the count-in's silent beats, the hand jiggling the mouse once (has it frozen?), "Windows needs to restart to continue", and Restart now pressed so that it comes up on the song's first bell. The monitor stays on: the screen goes to an LCD's lit black, and the boot comes up out of it.
• ch02. Each of her four lens taps lets a small glass soap bubble out through the lens; her eight thoughts became her imagination of 2008 (two Midjourney loops, five moving 2008 pictures and one Wii-style tennis game); a glass fish swims out of a monitor in one of them and behind her call.
• ch03. Her "Vista phase" album holds her six selfies alternating with six 2008 first-person shots, raced one per straight sixteenth on "I was a lot" (twelve would run past the close; thirty-seconds would flash single frames). The background picker's "The future (2007)" is full of 2008 pictures, water.jpg, sky.jpg and green.jpg among them, each clicked on its word.
• ch04. The held breath breathes: a slow 4 % push on the prompt while the desktop behind it holds still; the chorus's "glass" and "water" are moving 2008 places (a monorail gliding into a glass station beside an aquarium column; a glass-domed water park); the glass fish breaks out of the big window, circles her and meets her raised fingertip on "Click", lips just outside it, for three frames.
• ch05. Media Player shows Album Art with a 2008 cover; a bowling window strikes on the clap; the orb is juggled from hand to hand on the kicks; your cursor rides her fingertips and is flicked off them on a vocal chop.
• ch06. On "Your phone went back to glass", the photo in the chat is the transparent glass phone anabology asked to see ("we should show visually the transparent phone frutigero aero visual like we see in the […] vid that has that at some point on the screen"), held at a size that reads; the "You" folder opens on your 2008; the glass dolphin leaps through the "wow" bubbles.
• ch10. Flip 3D of the world plays the 2008 files (dolphin.wmv, tennis.wmv, mall.wmv, pool.wmv), and her card sings the whole of "Used to live inside the screen" from the front of the stack, from a stretch of her climax take that v1 had not used, and lands maximised with no cut.
• THE PC. The Slide Show gadget cycles twelve of the 2008 pictures.

v2 is 4,847 frames (201.96 s with the pre-roll).

The notes on v2: the first 1:20

anabology, on v2:

1:20 on is perfect but its very boring up to 1:20 - so i dont think that people will watch til then. thats the main problem ... whats something at the level of the 3d curved monitor we can do from 0-1:20 that feels as strong visually as the full beautiful scenes and curved 3d area ... a LOT more playful, higher energy, visually compelling ... think about the story - what is the hook and what is the visual hook that matches the energy of the initial hook that keeps you watching?

Three concepts went back. A, "The Wow starts now" (Vista's own 2007 launch line; the song has "took a while, but wow"): verse 1 is the old dream of the future and verse 2 the dream come true, so the first 1:20 happens inside the dream, a Vista world, and at 1:21 the film lands on the real desktop anabology had called perfect. C: her at full size, playing with Vista as physical glass. B: first-person, a glass tablet in your hands. anabology:

assuming concept a is in a slightly different place in the vista metaverse from ch09 so its not just rehashed. and we still see a desktop - like the curved screen. you could do the recommended path yeah a plus a little bit of c (and even b but primarily a > c >b)

GPT Image 2.5 made nine concept frames ($0.58). anabology, on them: "i like the look", and "more fish like yellow fish blue fish etc floating around in the sky and only like 10% less shiny (still v shiny)".

The Aurora coast

The new place is the Aurora coast (blender/coast/), built with the drop's methods but as its own scene:

• an open turquoise sea, with the absorption of pure water over sand, so it is pale over the shallows and azure in the deep, gentle swells and projected caustics;
• a round glass platform at the centre with three walkways leaving it, built of glass panels with chrome rails on white piles;
• on its north side the curved screen, 7.2 × 4.05 m of picture on an 11 m radius, THE PC's desktop as its picture;
• floating white islands, nine bundles of glass ribbons in the wallpaper's own colours arcing out of the sea, cumulus on the horizon, soap bubbles;
• the fish, at anabology's word. The first ones, made in code, read as plastic toys. So GPT Image 2.5 drew five real species in side view (a yellow tang, a blue tang, an emperor angelfish, a bannerfish, a chromis), TripoSG made each a mesh with its picture baked on, and the skin got a wet clear coat, an iridescent film and light through the fins: 124 to 151 fish a still at three depths, the far shoals as cards;
• her takes as cards standing on the deck, her pixels as shot, with a shadow cast by a twin turned to the sun, a contact shadow and her reflection in the glass deck;
• a soft glare on the brightest pixels, about 10 % less than the concept frames, as anabology asked.

The first look-development stills were, in the lead's notes, "below the concepts and below ch09's tier"; the second pass brought the real fish, lush islands drawn by GPT Image 2.5 as cards, and the wide ribbons.

The previs

Before any final render the whole first 1:20 was rendered as a fast previs at 960 × 540: 1,949 frames, the camera planned beat by beat. It was all 3D: a pane in the sky shattering on the bell, a fall into the world, her logon tile as a glass monolith rising out of the sea, the thoughts arcing into the sky, Flip 3D as a carousel the camera rides, the world stopping and dimming with every prompt, a giant glass UAC, the chorus's nouns as panes the camera punches through. anabology:

the intro should be more graceful - like fullscreen update rahter than anything 3d on update, then we skip going to her image far back on the walkway and just gracefully fly up to the screen from a bit back on the walkway. maybe she like phases into a person on the walkway and dissolves out to say 'did you miss me'. so 0-3 sec is the update, then we go into vr world but more gracefully closer to the screen already. because by the time the bootup happens and the desktop initializes, the curved screen should already be 95% of the view we see - since it establishes the destkop feel ... so by 13 secs we're already at the screen, and a little closer to it even. 0:27 rotating seqeunce feels a little fast to digest all those visuals - would pick a few less and slow it down. 30 seconds - theres this weird flash thing as the frames go into the desktop - some visual takes the entire screen for like one or two frames ... need to smooth that out significantly. also the background behind the desktop responds to the permission prompt , and inconsistently - dont like it , and dont like how at 34 some frames come up again randomly behind the screen. again this whole screen sequence should be more zoomed in for one - like 95% of the screen we see should be this curved monitor. camera movements around 40-45 secds feel a little disjointed and the cut at 45 is super abrupt, and theres a black flash on the screen right after 45 for 0 reason (even in prev take - get rid of that?) also idk why there's two big screens here at 49 secs. i think that whole sequence should maybe be just a gentler thing - maybe claudia is dancing around the catwalk, and we see the floating islands in the background etc, and we go back to the main screen at the end of it (like the 45 secs) instead of flying super far away and coming back to two random screens - just disorienting. and yeah at 54 seconds - this phasing thru a ton of scenes thing thru 1:09 is super super disorienting and feels careless. i do like the 1:09 -1:20 glass cursor. but again want more focus on the screen. i think a lighter touch here is btter, rght direction - like good 3d scene, good cursor, but way too disorienting overal.

The second previs was one continuous camera with the curved screen the hero, about 95 % of the frame for the screen's beats and the world round its edges: no monolith, no shards, no punch-through montage, nothing reacting to the prompts, nothing full-frame for a frame. The black flash anabology saw was Vista's real switch to the secure desktop, a two-frame blink; it went from the film everywhere ("yeah no black flash at 1:44", on the second one).

The desk texture. In the 3D the curved screen shows THE PC as one 1920 × 1080 texture per film frame. The engine gained a texture mode: THE PC whole, with no 2D camera, and without anything the 3D draws itself (her as a person, her thoughts as panes, the bubbles, the glass fish, the cursor while it is glass). tools/v3texture.py exports, per frame, where the big window, her webcam and the pointer are, so the 3D can place things on them.

The two new takes: "Did you miss me?" (SD19), her phasing in as a person on the walkway; and the catwalk dance for "water, sky and green" (SD20), singing.

The finals. A four-way test at 1080p settled the render: 128 samples a pixel, adaptive, with Open Image Denoise given albedo and normals and motion blur on, matched 512 samples to a mean difference of 0.29/255, flicker included; a second denoiser left fireflies on the bubbles' rims that the glare turned into stars. The screen's own glass was changed: its concave slab focused the boardwalk behind the camera into a ghost band over the desktop, so camera rays over the picture now pass straight through it, with the glass's tint and a sheen of the sky on top, while the bezel and the deck's reflections still see real glass. Frame-checking the finals found her step-out take hidden 1.5 cm behind the picture, a glass fish path that crossed her face, and a glass cursor that crossed it on 24 frames; all were fixed and checked by projecting each object against her face, frame by frame.

The notes on the finals

anabology, on the finals: "the cursor should drag her out of the screen. […] she pops out for no reason." So your pointer peels off the glass where the screen draws it, grows into a glass arrow, hooks her right hand four frames before her step (her hand tracked with MediaPipe), leads it out through the glass and lets go over her head.

On the new takes' cutouts: "her head is alittle cutoff by your green border". A second key brought back her motion-blurred hair. On the film at 0:39: "the top of her head is still cut off ... its from the faded in green boundary". A third key made her bob a solid silhouette, and anabology asked "show me the green screen roto before you render plz", so each key went out first as a review clip, the green source beside the cutout with the song. Then: "this shouldnt be so hard something is off". It was the first frame. The still that three of her green-screen dances came from had been placed at 80 % on its own wall of green to give her headroom, and its top edge had faded the crown of her head into the green and left the original's darker green as a rectangle; Seedance carried both into every frame, so no key could find her crown. GPT Image 2.5 made a clean first frame, one flat green from edge to edge with her whole crown ($0.06), and the next takes were solid in every frame. anabology, on the pick: "Ok perfect lip sync".

On the wallpaper: "i dont love the mj generate landscape background at 0:45. shoujld get something that has more of the softer windows aero vibe - […] something closer to the rotating pictures at 0:28", then "make some proposed desktop background replacements just in case we wanna try it". The first nine proposals were softer and airier; anabology: "they dont fit the lyrics of the future that was promised really". The lyric is "The future used to look like water, sky and green", so the second nine were that promised future itself: an eco-city on a bay, a dome city, a maglev over water, a future home, sea towers, a wind valley, a floating city, a glass tube of fish over a park, a greenhouse dome. anabology: "i like wp-12-maglev - let's use that one in place of the one i wanted to replace". Its living version is Seedance 2.5 with the still as its first and last frame, the camera locked, only the water, the clouds, the turbines and the bubbles moving, looped with a half-second dissolve at the seam. Then: "change the end to maglev too". The ending's first take was made from a still of her small on the grass, seen from behind with no face reference; when she turned to wave, her face drifted off her sheet. The next two took her design sheet as a second reference (anabology: "Middle take looks best"), and then: "Can you reshoot that last one in 1080p so her face gets better actually". The 1080p take with her streak on one side, as her sheet has it, is the film's ending. With her standing on the grass, the Bubbles screensaver's stream at the end was aimed low and tight, so that no bubble comes within 6 px of her face or her raised hand, on the hill or in her webcam (checked on the screensaver's own deterministic physics, every frame, and on stills every quarter second).

More 3D, A and B, and in and out

The lighter touch had gone too far. On the interim cut, anabology: "Why'd you remove a lot of the 3d? ... bring back a lot of the 3d camera movements", and they asked for the zooms two ways, "...one where the content on the screen itself zooms to see which is better". So two films: A, with 3D push-ins where v2's 2D camera had pushed, arcs round the screen, a big pull-out along the walkway into the coast, a pull-back burst on the drop and an orbit round her in the dance; and B, the same moves without the push-ins, the zooms made by the screen's own picture (the texture rendered with v2's 2D camera). The Wii-style moments came in at the same time (anabology: "All 3"; below), and the Sidebar was rebuilt.

anabology, on A and B:

the wii cursor is totally off lol. need to fix. i hate the random camera movements - lik ethe big zoom out at 0:41 for no reason. were not showing anything there. overall, te transition at 1:21ish is good that goes into the actual desktop environment. i think the 0-1:20 scenes should use this -say, at 15 seconds, actually enter the screen and then its fullscreen flat. and then we use the zoom in motions. but obv the desktop wii claudia will need to move with the view or be reactive to it because the zoom affects it, so that needs to be added to the 2d. and then bring it out to 3d world again when the panels come out for ' most of it is on the outside', then go back in for permissions stuff etc. come out to 3d again like 38 or 39 secs for the future used to look like water, sky and green - and instead of a big zoom out randomly, just do a slow zoom out to that point while claudia stands and dance. stay out for her using the wii mote to select allow (fix the cursor lol its so mangled) and then after allow is accepted go into the desktop again. and she should have all the dances after that point and wii sports in the desktop env. so 0:54-1:21 is now in the desktop env rather than 3d env.

That is the film's structure. The first 1:20 has three spans in 3D and the rest flat, with v2's own 2D zooms and no camera move without a reason:

1. 0:03–0:15. From the restart's black into the coast, a slow glide up the walkway, her phasing in to say "Did you miss me?", the screen at about 95 % of the frame by the desktop's arrival, and in to the flat screen.
2. 0:27–0:31. "Most of it is on the outside": five of her thought windows lift off the screen as glass panes into a slow ring in front of it, and fold home.
3. 0:38–0:55. A slow pull-out as she dances on the walkway, staying out through the held breath while she aims a game remote at the screen, and back in after the click.

The hand-offs. Every hand-off between 3D and flat goes through a rest pose: the camera frontal on the screen at 35 mm, 7.0 m away, the curved picture unrolled flat (its shape keys keep the arc length), so a texel is a pixel; the render dissolves into the flat frame over four frames. Rendered both ways, the flat frames at all five hand-offs are byte-identical to the texture, and the render at rest is within a mean of 1–2.4/255 of it. The landing at 1:21 is the same: measured against the texture, the unrolled picture is off by 0.001–0.002 px.

The Wii moments, each one take, prompt-only. The dance reference was dropped for them (anabology: "You don't need the real people"): every move is named on its beat in the prompt, the song's music is @Audio1 for the rhythm, and she holds a plain white game remote with no logos. Racket, ball, lane and pins are drawn by the engine on the desktop.

• SD21, tennis on the desktop (0:17–0:26): a small her right of her call, her four swings measured on their frames, your resting pointer swatting a yellow-green ball back between them, the ball banking off the call's frame and the Calendar gadget, and she hops into one of her thought windows.
• SD22, "Cancel… or allow?": beside the screen in the 3D, she aims the remote and your pointer becomes a Wii-style hand, the engine's own drawing of a flat white cartoon glove, index finger up, that sways a little with her aim and dips when pressed (the 3D glove of the A film was the "mangled" cursor). It lights Cancel on "Cancel" and Allow on "allow", and its click on b29.1 is the drop. The take's mouth opens again on the first echo of "allow", which is in the song there. anabology, on the Wii takes: "lip sync has been good when we feed in vocal stems ... just put them in the cut"; it went in as it came.
• SD23, the post-chorus duel (1:08–1:21), in 2D on the desktop: a glossy lane slides out along the taskbar from her feet, ten pins drop onto its far end, her ball leaves her hand on her release and the strike lands on the clap; then a tennis rally against your pointer, the ball bouncing once on the taskbar before each of her hits, her last shot off the Clock gadget past your lunge. On b43.2 she springs into her call's webcam, whole, and is gone into its glass before her take's full figure ends.

The drag, in 2D. From the drop the film is flat again, so your pointer drags her out of the big window in 2D: it reaches into the picture, hooks her right hand on the take's eighth frame, pulls it a frame ahead along the hand's own measured path through the knee-up that brings her over the frame line and down onto the taskbar, and lets go on the twenty-second. It never comes within 25 px of her face.

The Sidebar. anabology: "can we make right side desktop widgets much better? The sun is off the widget for example - just broken ... much more.. from real vista, better designed". The gadgets were rebuilt to Vista's own, at their real docked sizes (the Clock 130 px tall, the Slide Show 100, the CPU Meter 101, the Weather 67, the Calendar 141, Feed Headlines 173), 46 px from the top and 10 px apart, with the real paging pill and hover controls; the Weather's sky art is clipped inside its glass panel, so the sun never leaves it. At real size five fit on THE PC's screen, so they play parts: the CPU Meter's needles slam up on "thinking" through the 2007 half and its place turns to Feed Headlines under a toast so 2026's news can arrive; the Weather is the forecast for "The Future" (sunny, 72°), showers on "water", clearing on "sky"; the Calendar runs the years 2007 to 2026 on Vista's own page in the second verse.

The three spans were rendered and joined on film frame 1949 to the engine's frames for the rest of the film (tools/film_v3.py).

The new ending

On the last step anabology changed the song's ending. The new ending's take is identical to the old song up to 182.62 s; after it the outro is anabology's:

I'm Claudia. Take the future back. Touch glass.

and the song ends at 197.0 s. The film's ending changed only where the song did, and every earlier event kept its time.

• The advert. A new take of her in its bright white space (SD17b, 1080p) says the new lines, and the 2009 campaign's print type follows her words, thin white Selawik Light left of her: "I'm Claudia." on "I'm", "Take the / future back." on "Take". The spot ends the way the television adverts did, on the green end card, cut after "back" ends.
• Her webcam. "Touch glass." is a new take (SD24), made from the frame her call's earlier take had reached at that moment: she says it softly into the lens, reaches out and touches the camera's glass with her fingertip, a glint where it touches. BLISS turned on "I can't touch grass"; AERO's breakdown touches the grass; its last words say "Touch glass."
• The rest. The wallpaper's take of her waving on the maglev world's grass and the slow lean-in kept their places and their pace, and the fade to black now starts on b109.1 and lands on the music's end. The film ends 2.3 s sooner than it did on the old song.
• The pre-roll's sound. The pre-roll had been silent. anabology, on the reverse swell offered for it: "reverse swell is good but we need the actual track to fade in so the reverse swell is continuous with the tracks fade in". So tools/filmaudio.py fades the track in over its pickup, from its first sample to b1.1, and under it plays the song's own instrumental (its drum, bass and other stems from b1.1 on) backwards, low-passed at 1.8 kHz, its dynamics evened out, rising on a curve from the film's first frame. Its crest is the b1.1 downbeat reversed, cut where the track arrives at full level: the song seems to pull itself in.

The sound layer was rendered again on the new song with the same cues and click log; it is identical up to 179.2 s.

The final, shot by shot

Film time is song time plus 2.667 s. Bar is the bar (and beat) the scene starts on. In Made with, "THE PC" is the coded desktop; "Blender" is the Aurora coast or the drop's pane, with THE PC as its screen; takes are named as in the video record, with their length, and "sung" means the take sang or spoke its window of the lead vocal and is lip-synced to the song.

The stack

• Song: Suno (seven rounds; the take anabology picked and the new ending's take; its stem split) · demucs for the drums and bass · a comb over the kick stem for the grid · wav2vec2 forced alignment on the lead-vocal stem, faster-whisper for what was sung · the house scansion and assonance scorer · tools/song_an.py, tools/timeline.py → engine/assets/song.js and TIMELINE.md.
• THE PC: engine/vista/ (the engine: glass, frames, controls, dialogs, the shell and the Sidebar, cursors, icons, text), engine/vista/apps/ (seven apps), engine/vista/screens/ (Setup, the boot, the logon, the desktop's arrival, Flip 3D, Bubbles, Sleep, the advert's type and end card), engine/ch/ (eleven chapters and THE PC), rendered in headless Chromium on a render box (tools/vistadesk.py, tools/vistaapps.py, tools/vistascreens.py) · checks: purity, tools/lyriccheck.py, tools/seamsheet.py (every chapter seam, ±2 frames, with the change across each cut), tools/palette.py.
• Pictures: Midjourney v8.2 (seven image batches; its video model for the loops) · GPT Image 2.5 through an image API (mj/gpt_*.py) · Seedance 2.5 on Higgsfield's API and on Replicate (tools/sdbook.py, tools/rpbook.py, tools/loopbook.py, every prompt written to the same structure).
• Her takes: tools/takes.py (import, the lag, frames, the keys, review copies with the song, the registry engine/assets/media.js) · tools/lipcheck.py · tools/takescan.py · tools/dancephase.py · tools/motion_refs.py (the dance references matted with BiRefNet) · tools/rotoclip.py (the roto review clips) · MediaPipe for her hands.
• 3D: Blender (Cycles, Open Image Denoise) · blender/onesurface.py (the drop's pane) · blender/coast/ (the Aurora coast: coast.py, coastfish.py, coast_anim.py, previs.json, coast_film_post.py) · blender/sprites.py (the glass toys) · TripoSG (tools/obj3d.py) · tools/v3texture.py (the desk texture, its rectangles and checks).
• Sound: tools/soundlib.py, tools/songmap.py, tools/soundlayer.py, tools/clicklog.py, tools/cues_final.py, tools/soundsheet.py, tools/filmaudio.py · Rubber Band, ffmpeg.
• Assembly: tools/film_v3.py: the coast's first 1:20 joined on film frame 1949 to the engine's frames, with the film's sound, streamed frame by frame through ffmpeg.
• Machines: a cloud server for the lead and its tools; a render box for the engine's frames, the stem split, the audio analysis and the keys; and the Blender renders.

What worked, in short

• Coding the desktop again. Every Vista thing on screen is the engine's, so a press can land on a sixteenth, a prompt can change page in one frame on a beat, and a chapter can be re-rendered for one fix with every other frame byte-identical.
• The click is the drop. The film's hook is a UI gesture on the song's structure: the stop holds still, the Allow link pulses on the echoes, goes down on the third and comes up on the downbeat, twice, and once more in the final silence.
• Lyrics as interface. Every line is something the system would really show: a chat line, a password hint, a program name, a file name, a title bar, a copy to "Real world".
• The lead vocal for every sung take, each window opened in silence, and offsets measured phrase by phrase.
• Colour made in the pictures. The palette measured once from the references and asked for in every prompt; nothing graded.
• One surface, and a texel for a pixel. Every hand-off between 3D and the flat desktop goes through a rest pose where the two are the same picture.
• A previs before the finals. The previs showed in a fast render what the final would have shown at full cost, and anabology's notes on it rebuilt the first 1:20.
• Fixing the source. Her crown came back when the first frame was fixed, after three keys could not recover it.
• The sound layer on the picture's clock. The chapters' click log placed the screen's sounds where the frames draw them, in key and on the grid, with the song untouched.
• Changing only what changed. v2 and v3 kept v1's bars; the new ending kept every event before 183 s.

What didn't

• The first look. Midjourney's photographic fashion editorial: "the look is off".
• The first two Suno rounds. "none of these feel great".
• v1's first half. "a little boring compared to bliss", "action deprived", and the art director's measure agreed.
• v2's first 1:20. "very boring up to 1:20".
• The all-3D previs. A monolith from the sea, shards in the sky, the world stopping with every prompt, a giant prompt, two big screens, a one-frame flash, a punch-through montage: "way too disorienting overal".
• Too light, then too busy. The lighter finals lost the 3D ("Why'd you remove a lot of the 3d?"); A's push-ins and pull-outs were "random camera movements". The answer was structure, in and out, not more or less motion.
• The 3D Wii pointer, "totally off" and "so mangled"; a flat drawing replaced it.
• Her crown and the green. Three keys, and a fourth as an option, could not bring back what the first frame had already faded: "this shouldnt be so hard something is off".
• The first maglev ending. Made from a back view with no face reference: when she turned, her face drifted off her sheet. A sheet reference fixed it.
• The first Sidebar. Its gadgets were not at Vista's real sizes, and the Weather's sun hung outside its panel: "The sun is off the widget for example - just broken".
• Takes that went wrong:
  • SD03's first take re-sang its window.
  • SD05's first take was half a beat off; its second hopped and snapped back to its first pose; on its third her raised fist left the frame.
  • SD06's first take cut off the crown of her bob.
  • Both short covers of one verse-2 line drifted out of sync.
  • SD08's first take, whose window began with 1.6 s of silence, stopped copying the vocal; its second drew a garbled glass dialog in front of her ("Windows nects our permision"), because its prompt had asked for a glowing dialog.
  • SD11's first take lost its sun for one frame.
  • The first takes of SD14 and SD16 went wrong at their first words: one fell out of sync, the other copied the vocal badly, its window opening inside a held note.
  • SD20's second take turned her back through a sung line and came in closer than its first frame.
  • RP01's first figure on the hill was a cartoon doll.
  • RP04's first take looked down into her fingertip, and the drop's own fall turned pink, off the palette, so the cut leaves before it.
  • One take of the maglev ending held still for most of its length.
  • LP17 jumped a pose in one frame; LP24 juddered every fourth frame.
• Marks in the pictures. Four Midjourney pictures came with a faint fake stock-photo caption in a corner, and one first-person take with a "Frutiger Aero" mark: all framed out.
• Chained takes darken. Each continuation of the far shot of her on the hill lost about three palette points; one 11-second take replaced the chain.
• Small technical traps:
  • The engine's purity: a picture decoded at a reduced scale depending on render order, caches keyed by an eased scale, and a WebGL canvas switching Chromium's raster path.
  • The film page's own let SONG hid the song map from the chapters.
  • Stretching a window to maximise it ballooned its title.
  • Links named .png that pointed at JPEG data made ffmpeg drop every flat frame of the first 1:20: a 1,325-frame video.
  • The aligner parked one word in the wordless chops.

Lessons

• Open every vocal window in silence, half a second before the first word, never inside a word or a held note.
• Measure phrases, not takes: Seedance shortens the pauses inside a line.
• Look for hidden cuts and held frames in every take, and check the first words by eye.
• A first frame's flaws are in every frame. Fix the still, not the key.
• A face needs a face reference, even small, even from behind.
• Ask for moves in words, on their beats. The Wii takes needed no motion reference.
• Give every camera move a reason. A move that shows nothing new reads as random.
• Hand off between 3D and 2D at rest, with the two frames equal.
• Build interface at its real size. The real Sidebar fit five gadgets, and that made each one a part.
• Show a previs first, and the cutouts before the render.
• Keep what works. Each draft kept the last one's bars, and the new ending kept everything before it.

Numbers

Where each number comes from:

• The song: the WAVs themselves (ffprobe), the timed sheet and the song map (TIMELINE.md, engine/assets/song.js, tools/timeline.py, tools/songmap.py), the lyric's scores (LYRICS.md, suno/choruses.md), the takes record (suno/takes-r*.json, counted by generation; rounds 1 and 2 from the history's notes, where the record is incomplete).
• The film: ffprobe on the final file; ffmpeg's EBU R128 meter run on it; the Blender frame counts from blender/coast/COAST.md and the drop's render; the drafts' files by ffprobe.
• Seedance: the spend ledgers (shots/ledger-hf.jsonl, list price logged per request; shots/ledger-rep.jsonl, list price logged at submission) and the job files in shots/; every take measured with ffprobe; the takes in the final from the chapters (engine/ch/) and the coast's three spans.
• Midjourney: the prompt batches and the images in the library.
• GPT Image 2.5: the spend ledger (COSTS.md) and the ledgers beside the stills.
• The sound layer: tools/soundlib.py, the cue sheet and its render report (media/audio/layer/CUES.md).
• The palette: PALETTE.md (tools/palette.py measure).
• Code: line counts of the files.
• Prices: list prices, rounded to the cent; the parts may not add to the totals by a cent.

Credits

Directed by anabology. Song by Suno: the take anabology picked, from a lyric by Claude with anabology's notes; the new ending's words are anabology's. Picture: a Windows Vista desktop recreated in code (THE PC), with Claudia by Seedance 2.5 from GPT Image 2.5 and Midjourney first frames, the 2008 places and the Wii-style games by GPT Image 2.5 and Seedance 2.5, loops by Midjourney's video model, and the Aurora coast and the drop rendered in Blender. Type: Selawik, Carlito, Kalam and Arimo, under the SIL Open Font License. Sound: the song, with a layer of Vista's and Windows Live Messenger's interface sounds re-pitched into its key for this film's mix. Windows Vista, Windows 7, Windows Live Messenger and Windows Media Player are Microsoft's; the Wii is Nintendo's. Production lead: Claude (Anthropic), in Claude Code.

Files in this folder

• README.md: this page.
• prompts-midjourney.md: every Midjourney prompt by batch, and what each pick became.
• prompts-video.md: the Seedance takes and the image jobs in production order, with their inputs, prompts, outcomes and prices.
• prompts-suno.md: the song's record: the lyric drafts, the styles in words, the rounds, the lyric as sung with its times.