Text Flash Cuts — Style Prompt

Avant-garde TV anime in the manner of the early-2000s "text frame" school: full-screen heavy Mincho title cards cut in for two to nine frames, architecture reduced to symbolic geometry, a red-black-white palette, characters as silhouettes or extreme close-ups, and the signature over-the-shoulder head tilt. Painted characters come from image generation; everything else (cards, buildings, rain, cuts, inversions) is code. References (grammar only): the early-2000s TV-anime "text frame" school, with its flash title cards, symbolic architecture and over-the-shoulder head tilt. Borrow the editing and design grammar; never a work's characters, lines, card texts, logos or specific shots.

1. Essence, and what it is not

Three layers, and the collision between them is the look: - Type as picture. A word fills the frame, bleeds off an edge, or stands in a vertical column. Cards are dialogue, chapter titles and punctuation of the cut. - Symbolic architecture. Stairs that never end, window grids, lockers, a one-point street: flat polygons, no props, no sky, no texture. - Flat characters in three states only: black silhouette (far), flat cel colour (mid), fragment (an eye, a hand, a mouth corner) filling the frame.

It is not a lyric video (type is not synced to a song and never animates letter by letter), not a manga panel film (no speech balloons, no screentone), not a horror film (no gore, no blood; dread comes from cutting, silence and stillness), and not a slide deck (never centred text on a plain card followed by a page-turn).

2. Materials & rendering

• Characters: generated cel art, clean thin uniform black lines, completely flat fills with a single hard shadow tone, no gradients, no paper grain. Keyed from a chroma-green background into RGBA; any pure generated red is pulled to the film's red.
• Everything code-drawn is hard-edged polygons on Canvas 2D. No blur, no glow, no grain (mux grain 0). The only smear is the head-tilt ghosting (three 28 % copies of the head layer).
• Silhouettes are the same art filled with one colour (source-in), so a character can be black, white or red without new images.

3. Colour logic

Secondary values only for depth inside geometry: #d8d4cc, #26262a, #1b1b1f. Skin and hair in generated art are the only other colours. - Inversion is red-black swapping, never RGB inversion (RGB inversion makes cyan): grayscale → contrast(5) → invert → multiply red. White becomes black, black becomes red. Used as 1–6-frame accents on impacts, never as a look. - Each speaker owns a field colour (for example: protagonist = red rail, black type; second character = black rail, white type). Keep it fixed through the film so the audience reads who speaks without voices.

4. Type

• Cards and titles: Noto Serif SC 900 (heavy Mincho feel, OFL). Single glyphs 700–1100 px, allowed to bleed off the frame on flash cards; dialogue columns 120–170 px, never bleeding.
• Small vertical notes (episode numbers, speaker tags, counters): Noto Sans SC 900, 40–64 px.
• English: one line of Cormorant Garamond Italic 600, about a quarter of the Chinese size, grey on black or white, black on red.
• Vertical punctuation: 。and ，move to the top right of their cell; ？stays upright.
• Never centred on the frame. Type sits in a corner, on a rail at the edge, on an object (a letter), or bleeds.
• Two classes of text: flash cards (2–9 frames, deliberately below reading time, only ever repeat or fragment words that are also shown in full) and held lines (dialogue, a letter, the end card), each held for its full reading time and checked with readcheck.mjs.

5. Motion quality

• Editing is the animation: flash cards 2–9 frames at 30 fps; architecture shots 0.3–1.6 s; nothing static for more than a second. Lock cuts to a beat grid (150 BPM = 12 frames a beat works well at 30 fps; 8th = 6 frames, 16th = 3 frames).
• Card slam: frame 0 at 1.08×, frame 1 at 1.02×, then 1.0; jitter 10–14 px on the first two frames, 1–3 px after (whole-pixel random per frame, no motion blur).
• Rails slam in from the frame edge in 3 frames and are counted as on screen only when complete.
• Characters move by layers and steps, not tweens: a head layer rotated about the neck, a figure that jumps closer once per beat, a silhouette that bobs once per step.
• Accelerate by shortening shots (9 → 6 → 4 → 3 frames), not by moving faster.

6. Camera grammar

Cut on beats. Transitions are cuts, 1–6 frames of inversion, a one-point tunnel rush, or a red slit opening on black. Never dissolves, wipes or page turns, and never a fully black frame between shots (use the slit).

7. Sound palette

• Instruments: harpsichord ostinato in 16ths, contrabass and cello pizzicato, a woodblock as a clock, glockenspiel for a lullaby motif, violins tremolo at a whisper, bowed vibraphone, timpani, bass drum, snare, low brass (trombone, tuba, horn) for stabs and a ♭9 chord, tubular bells, tam-tam. Avoid piano and wall-to-wall string pads.
• Card hit (synthesised): high-passed noise transient + metallic ring at 1760/2637/4186 Hz + a 52 Hz drop; pitch rises card by card as tension rises.
• Foley: locker clank, paper, phone vibration (155 Hz gated square), footsteps that grow louder, a weather bed, a neck-bone "crack" (low clack + click + noise transient) for the head tilt, heartbeat.
• Silence is a cut: at least two hard silences (a gate on everything but one sound), the longest right before the head tilt.
• Mix: clock and hits dry and forward, music under; limit transients before loudness normalisation (−14 LUFS, true peak ≤ −1 dBTP).

8. Native moves

A menu: use the ones your story needs. Each is built in code in demo/main.js.

• Flash card (shotCard): a solid field, one giant glyph (bleed allowed), slam + jitter, 2–9 frames; cards alternate with 3–6-frame picture cuts in a "type–image–type" saw-tooth. Fits content like: a countdown; a list of demands; a word that interrupts.
• Speaker rail (rail): a full-height column at the left or right edge in the speaker's colour, vertical type plus one English line and a one-glyph speaker tag; the picture beside it keeps cutting while the line holds. Fits content like: a phone call; a confession; a two-hander argument.
• Chapter card (shotChapter): one enormous glyph, a vertical episode number, an English line, a colour bar cutting in; four layouts so no two chapters look alike. Fits content like: a series trailer; a manifesto in parts; a product in four promises.
• The head tilt (shotTurn): the character art split into head and body layers (head polygon + neck pivot + a black collar patch that hides the seam). Back-turned (head −35°) → 60 ms wind-up → 120 ms whip to +0.1 rad with overshoot → 0.1·e^(−9t)·cos(26t) settle. The whip happens in full silhouette with three ghost heads; the frame it settles cuts hard to colour with 2 frames of red-black inversion. Fits content like: a reveal; a challenge; a brand mascot's "look".
• Symbolic architecture: geometric stairs (treads scaled by r^k, red light climbing the risers), window tunnel (nested rectangles k = 1/(z·0.32+0.08)), one-point street with lamps scaled by depth, locker wall, window wall. Fits content like: a school; an office tower; a city at night.
• Stepping approach: a figure that is redrawn larger once per beat, never tweened. Fits content like: a deadline; a rival; an incoming storm.
• Fragment insert: eye, hand, mouth corner or object at frame-filling scale for 6–24 frames, slammed and jittered. A reflection drawn into the iris ties two characters together. Fits content like: a clue; a touch; a lie.
• Multi-panel faces: one portrait cropped nine ways in a 3×3 grid (full, eye, mouth, upside down, silhouette), appearing on 16ths, then all heads tilt together with staggered springs. Fits content like: surveillance; a split personality; many fans.
• Count-up acceleration: numerals as flash cards alternating with ever-shorter architecture shots. Fits content like: a launch; floors of a building; days until.
• Object on a solid field: a key, a cracked watch, a ticket, drawn flat in red, black and white. Fits content like: a product; a clue; a keepsake.

Our demo opens on a one-point tunnel rush into flash cards that spell the first line, uses count-up acceleration as its build and the head tilt as its climax, and ends on fragment inserts and the end card. A new film opens and ends another way: a clock that stops, an object on a solid field before any figure is seen, the head tilt spent early as a turn rather than saved for the end, an ending on an architecture shot reversed, or a card cut before the last word lands.

9. Characters and image generation

An image generator is required. The characters are drawn with the image tool the user provides: their Codex CLI through core/image/gen_image.py (--check first), or another image API or app, where they make the images from your prompts and save them under your file names. With neither, a film in this style cannot start: say so in the brief (core/README.md, "Image generation"); never draw stand-in characters in code.

Materials first. Before generating anything, ask the user whether they have their own character designs, logos or product images; what they give comes first. If not, characters come from the image generator, and before generating, tell the user how many images and of what (for example: "one model sheet each for two characters, then six poses and three close-ups, eleven images").

Pipeline. 1. A model sheet per character: a half-body in the defining pose on flat chroma green (#00FF00). 2. Every further image gets that sheet as a reference image (--ref in gen_image.py) (and the other character's sheet when both appear), with the prompt saying "the same character as the reference image". A new character uses an existing one only as a style reference ("a different person; the reference is for drawing style only"). 3. Close-ups (eye, mouth, hands, letter) on their own; full-frame inserts may keep a plain background, everything that will be composited goes on chroma green. 4. No text in any image ("no text, no letters, no logo, no watermark"); all type is set in code.

Prompt template (fill the bracket):

Flat cel animation style of a modern avant-garde TV anime, matching the reference image's drawing style exactly: clean thin uniform black linework, completely flat colors with one hard-edged shadow tone only, no gradients, no painterly texture, no blush hatching. Palette limited to jet black, off-white, pale skin and crimson red. [subject, pose, framing, expression, costume]. Background: perfectly uniform flat pure chroma green #00FF00, no shadow on the background, no floor. no text, no letters, no logo, no watermark.

Consistency rules. Fix each character's silhouette-defining features (hair shape, one accessory such as a red clip, collar) and repeat them in every prompt; give each character one eye colour that survives as a cue in close-ups. Measure the neck pivot and head polygon on the image you rig; a new image means new measurements.

10. Pitfalls of the medium

• RGB inversion turns red cyan → red-black swap (§3).
• A large head rotation shows a gap at the neck → only rotate far while the figure is a silhouette; in colour stay within 0.1 rad.
• A black silhouette in a black doorway disappears → doorways are white.
• A half-body image standing in a long shot "breaks at the waist" → generate a full-body back view for long shots.
• Flash cards counted as dialogue fail the reading check, and cutting away from a held line resets its clock → hold lines on rails or objects while the picture beside them cuts.
• Centred cards on plain fields read as slides → always bleed, offset, rail or object.
• Generated red (#ff0000-like) breaks the palette → pull saturated reds to #d4162b when preparing assets.
• Steady drones and pads fill the silences → gate everything but one sound before the peak.
• Very sharp hits plus loudness normalisation overshoot the peak → limit the mix before mux.sh.
• A subset font misses a new glyph and Canvas falls back to a system serif → build the subset from the source text (tools/fonts.sh collects every non-ASCII character in main.js).

11. Variation space

You decide the story, the characters, the spaces, the chapters, the colours each speaker owns and the order of the moves. All far from our demo:

• Structures: a manifesto in five words (each word a chapter card, each chapter one building); an interrogation (two rails facing each other across a table drawn as a single red line, shots of evidence between questions); a product's countdown (numerals and objects on solid fields, the product revealed with the head-tilt timing applied to a lid or a door).
• Openings: a single word on red before anything is seen; a clock that stops (woodblock, then silence); a one-point corridor rushing toward a closed door.
• Endings: the title appearing inside the last close-up (in an iris, on a blade, on a screen); an architecture shot reversed (the tunnel pulling out to show it was a pattern on a coat); a card that is cut before the last word lands.

How our demo was made (story, shots, score, end card, build): DEMO.md. Read it after your treatment exists.