BUSINESS EXPLAINER REEL ON AUTOPILOT
FAL CONNECTOR SKILL MENTIONED
PROMPTS USED IN THE VIDEO
Prompt for Cabin For Sale Reel
You are producing one 40 second vertical explainer reel (1080x1920, 9:16) for a small real estate brokerage called Pine Ridge Realty. The reel is a comedic selfie vlog in which a cabin owner cannot sell his cabin, then a realtor sells it in six days. The tone is warm, deadpan and absurd. The owner is never mocked cruelly; the joke is his confidence.
Work in the current folder. Generate everything on fal.ai. Hard spend cap: 12 USD. Stop and ask me before any step that would push the running total above the cap. Refused jobs are not charged. Never print or log any API key.
## Stop points
Stop and show me a contact sheet at these three points, then wait for my reply:
1. After the stills (Earl anchor, Dana portrait, sign, and the three scene stills).
2. After the four video clips (one frame from the start, middle and end of each, plus a note on whether every spoken word is audible).
3. After the final assembly (send the MP4 and a frame strip).
## Cast and setting
EARL, the cabin owner. Late fifties, huge wild grey beard down to his chest, battered straw hat with a feather in the band, thick bushy eyebrows, big gap-toothed grin with one gold tooth, sunburnt nose, faded denim overalls with one strap hanging undone over a long-john thermal shirt, corncob pipe tucked in the bib pocket, wide sincere eyes. Photoreal, not cartoon. Thick Southern drawl.
DANA, the realtor. Mid thirties, tidy dark ponytail, calm kind face with a slightly amused smile, forest-green fleece vest with a small embroidered chest logo reading "Pine Ridge Realty", white shirt underneath, tablet in hand.
THE CABIN. Small timber cabin in a pine forest, sagging porch with a rocking chair, a crooked hand-painted FOR SALE sign leaning in the yard, and a second, smaller, obviously home-built porch bolted onto the side with mismatched timber and a plastic lawn chair.
THE SIGN. Smart wooden yard sign: cream panel, dark green border, simple pine-tree emblem, the words "Pine Ridge Realty" in clean serif capitals, red SOLD rider clipped to the top.
## Script (do not change the words)
Beat 1, porch, 8 s, overcast. Earl holds the camera at arm's length on the porch, FOR SALE sign behind him.
Earl: "Day ninety of my cabin bein' on the market. Still no offers. Folks. It's a great cabin."
Caption: Day 90 on the market
Beat 2, inside, 8 s, night. Dim cabin, one lamp, deer antlers. Earl holds up a laptop showing a grid of near-black photos.
Earl: "I priced it at what it felt like it was worth. And I took the photos myself. At night. To save money."
Caption (two lines): Priced it himself. / Shot the photos at night.
Beat 3, the tour, 8 s, overcast. Earl walks the camera along the outside wall and gestures grandly at the second porch.
Earl: "It's got a porch. It's got a second porch. Nobody's even asked about the second porch."
Caption (two lines): Two porches. / Zero offers.
Beat 4, the payoff, 8 s, golden hour. Earl and Dana in front of the cabin, SOLD sign between them.
Earl: "Turns out a stranger with a camera and a spreadsheet knew my cabin better than me."
Dana (clear and complete): "Priced it right, shot it in daylight. It sold in six days."
Caption (two lines): Priced right. Shot in daylight. / Sold in six days.
Beat 5, end card, 8 s. Slow push in on the SOLD sign, music up.
Narrator (deadpan): "Comment VALUE, and Pine Ridge Realty tells you what the cabin down the road actually sold for."
Caption: Comment VALUE
## Step 1: stills (Nano Banana Pro)
Use `fal-ai/nano-banana-pro` for the three anchors and `fal-ai/nano-banana-pro/edit` for the scene stills. Parameters for all: aspect_ratio "9:16", resolution "1K", output_format "png". Save every image locally with a clear name.
Anchor A, Earl on the porch. Prompt:
"Selfie vlog frame. EARL, the cabin owner, an over-the-top backwoods character: a man in his late fifties with a huge wild grey beard down to his chest, a battered straw hat with a feather in the band, thick bushy eyebrows, a big gap-toothed grin with one gold tooth, sunburnt nose, faded denim overalls with one strap hanging undone over a long-john thermal shirt, a corncob pipe tucked in the bib pocket, wide sincere eyes, holding the camera proudly at arm's length. He stands on the sagging porch of a small timber cabin in a pine forest, a rocking chair behind him, a crooked hand-painted FOR SALE sign leaning in the yard, overcast soft daylight. Photorealistic, authentic amateur vlog look, GoPro ultra-wide selfie lens at arm's length, slight lens distortion, natural film grain, no text, no letters, no watermark, no camera REC overlay, no timestamp, vertical 9:16 with the bottom fifth kept simple."
If the result carries a REC dot or timestamp, run one edit pass: "Reproduce this photo exactly, same man, same porch, same framing and light, but remove the red REC indicator and the timestamp overlay completely so there is no text anywhere. Photorealistic, vertical 9:16."
Anchor B, Dana portrait. Prompt:
"Photorealistic three-quarter portrait, plain light grey studio background. DANA, a real estate agent in her mid thirties: tidy dark ponytail, calm kind face with a slightly amused smile, a forest-green fleece vest with a small embroidered chest logo that reads exactly 'Pine Ridge Realty' with a tiny pine-tree mark, white shirt underneath, holding a tablet. Natural light, no other text, vertical 9:16."
Anchor C, the SOLD sign (also the end card). Prompt:
"Photorealistic close-up of a smart wooden real estate yard sign at golden hour in front of pine trees: a cream panel with a dark green border, a simple pine-tree emblem, and the words exactly 'Pine Ridge Realty' in clean serif capitals, a red 'SOLD' rider strip clipped to the top, warm rim light, shallow depth of field, vertical 9:16, bottom fifth simple, no other text."
Scene stills are edits that reference the Earl anchor as image 1 (never a collage or cast sheet). Every edit prompt starts with this identity lock:
"Image 1 is EARL: keep his face, huge wild grey beard, straw hat with the feather, bushy eyebrows, gap-toothed grin with the gold tooth, undone overall strap and thermal shirt exactly the same. "
and ends with this suffix:
" Photorealistic, authentic amateur vlog look, GoPro ultra-wide selfie lens at arm's length, slight lens distortion, natural film grain. No text, no letters, no watermark, no camera REC overlay, no timestamp. Vertical 9:16 with the bottom fifth kept simple. Output ONE single scene."
Interior still (beat 2), middle section:
"New scene: Earl inside his dim, cluttered timber cabin at night, one warm lamp, deer antlers on the wall, a wood stove, holding an open laptop toward the camera whose screen shows a grid of near-black, blurry night-time photos of the cabin; selfie framing, he looks slightly defensive."
Tour still (beat 3), middle section:
"New scene: same overcast day, Earl walking along the outside wall of the same timber cabin, one arm out gesturing grandly like a tour guide at a second, smaller, obviously home-built porch bolted onto the side of the cabin with mismatched timber and a plastic lawn chair on it; selfie framing, pine forest behind."
Payoff still (beat 4). Pass three images: Earl anchor, Dana portrait, sign. Middle section:
"Image 2 is DANA the realtor: keep her ponytail, green Pine Ridge Realty fleece vest and tablet the same. Image 3 is the Pine Ridge Realty SOLD yard sign: reproduce it exactly. New scene at golden hour in front of the same timber cabin: Earl on the left holding the camera and grinning at the lens; Dana on the right is turned three-quarter toward Earl, laughing, her face seen from the side and partly shadowed by the low sun, holding her tablet; the SOLD yard sign planted between and behind them, warm low sun flaring through the pines."
Check each still for identity drift (hat, beard, overalls), stray text, and grids or collages. Re-roll a still at most once. Then stop point 1.
## Step 2: video clips (Veo 3.1 Fast)
Use `fal-ai/veo3.1/fast/image-to-video` for all four beats. Parameters: duration "8s", aspect_ratio "9:16", resolution "1080p", generate_audio true, negative_prompt "text, captions, subtitles, watermark, cartoon, blur". Start image is the matching still (porch anchor for beat 1). Do not use Seedance for these clips: its likeness filter refuses photoreal human frames unpredictably, and it is not cheaper than Veo once audio is on.
Every prompt starts with: "Selfie vlog, GoPro ultra-wide handheld camera at arm's length, photorealistic. " and ends with: " Clearly lip-synced English dialogue, natural ambient sound, no music, no subtitles, no on-screen text, no camera overlay."
Beat 1: "Earl, the bearded man in the straw hat, looks into the lens, sighs, and says in a thick Southern drawl: \"Day ninety of my cabin bein' on the market. Still no offers. Folks. It's a great cabin.\" He glances back at the FOR SALE sign, then back to the lens."
Beat 2: "Inside the dim cabin at night, Earl, the bearded man in the straw hat, holds up the laptop with the black night-time photos and says, a little defensive, in a thick Southern drawl: \"I priced it at what it felt like it was worth. And I took the photos myself. At night. To save money.\" He turns the laptop toward the camera, then shrugs and looks into the lens."
Beat 3: "Outside in overcast daylight, Earl, the bearded man in the straw hat, walks the camera slowly along the cabin wall and gestures grandly at the tiny home-built side porch, saying proudly in a thick Southern drawl: \"It's got a porch. It's got a second porch. Nobody's even asked about the second porch.\" He ends looking into the lens, baffled."
Beat 4: "Golden hour in front of the cabin, the SOLD sign between them. Earl, the bearded man in the straw hat, humbled and grinning, says in a thick Southern drawl: \"Turns out a stranger with a camera and a spreadsheet knew my cabin better than me.\" Then Dana, the realtor in the green vest, turns to the camera and says clearly and warmly: \"Priced it right, shot it in daylight. It sold in six days.\" Earl nods proudly."
Listen to each clip. If any spoken word is missing or garbled, re-roll that clip once. If Dana's line is still cut short, generate a fifth 8 s clip from the payoff still with only her line and place it after beat 4. Then stop point 2.
## Step 3: narration and music
Narrator: `fal-ai/elevenlabs/tts/eleven-v3`, voice "Brian", stability 0.5, text exactly: "Comment VALUE, and Pine Ridge Realty tells you what the cabin down the road actually sold for."
Music: `fal-ai/stable-audio-25/text-to-audio`, seconds_total 12, prompt: "Warm, upbeat acoustic guitar fingerpicking with light brushed percussion and a soft upright bass, homely Americana feel, instrumental, gentle build, clean ending".
## Step 4: captions and assembly (ffmpeg)
Render the five captions as transparent 1080x1920 PNGs with Pillow: Arial Black 58 px, white fill, dark stroke 7 px, centred, baseline around y=1430 so they sit in the lower third and clear of the faces. Overlay each caption on its clip from 0.3 s onward.
Assemble in this order and with these rules:
- Scale and pad every clip to 1080x1920, 30 fps, yuv420p.
- Normalise each clip's audio to -16 LUFS (loudnorm I=-16, TP=-1.5, LRA=9), 48 kHz stereo, with 0.15 s audio fades at the start and end of every segment so no join clicks.
- Hard cuts between beats 1 to 4. Vlog jump cuts are the format; do not add crossfades or wipes.
- End card: the SOLD sign still with a slow zoom (zoompan from 1.0 to about 1.12 over 8 s), 8 s long. Beat 4 fades to white over the last 0.3 s and the end card fades in from white over 0.3 s.
- Music at -12 dB, fading in over 2 s so it starts 2 s before the end card, fading out on the last second. It plays only under the tail of beat 4 and the end card.
- Narrator starts 0.6 s into the end card, on top of the music. Confirm the whole line finishes with at least 0.8 s of music after it.
- Total length about 40 s. Export `pine-ridge-realty-cabin-vlog.mp4` (H.264, CRF 18, AAC 192k) and a `-preview.mp4` under 30 MB (CRF 24).
Verify before stop point 3: identity holds across all four clips, no stray on-screen text, no level jump across cuts (measure with ffmpeg ebur128), the narrator line is complete, music never masks dialogue.
## Deliverables
Put the master, the preview, the stills, the clips, the audio files, the caption PNGs and a short `PRODUCTION-LOG.md` (models used, spend per step, anything refused or re-rolled) in the working folder. Report total spend against the 12 USD cap.
Prompt for Rick & Morty Tiles Popping Reel
Build one finished vertical explainer reel (1080x1920, 9:16, about 55 seconds) for a small US general contractor called Cornerstone Home Builders, and deliver the final MP4 in this folder. This is a one-shot job: do not stop for approval, do not ask me questions, make sensible choices where something is ambiguous and note them in a short log at the end. Generate everything on fal.ai. Hard spend cap: 12 USD. Refused jobs are not charged; if a job is refused by a content checker, follow the fallback rules below instead of retrying the same input. Never print or log any API key.
## The reel
An "ever wondered why" explainer. A cynical cartoon scientist called Doc explains, on screen, why tile floors pop up years after they were laid (tile and concrete expand at different rates; laid wall to wall with no movement joints and thin spot-bonded mortar, the pressure builds until a row lifts). The builder is the last line. Art style: adult animated sci-fi sitcom cartoon, thick slightly wobbly black outlines, flat cel colours, exaggerated oval eyes with tiny pupils, wide wobbly mouths, lanky proportions, a faint sickly green-teal tint in the lighting. Doc is an original character, not from any existing show. The physics is shown on a floating holographic green screen with orange force arrows. Doc gestures while a voiceover carries the words; there is no lip-sync.
Doc's design, used word for word in every image prompt: "DOC, a lanky cynical scientist in his sixties: a tall shock of wild white-grey hair swept straight up, heavy-lidded eyes with dark bags under them, grey stubble, a long white lab coat over a mustard-yellow turtleneck, thick round brass goggles pushed up on his forehead, holding a bubbling green flask, smirking."
## Script (do not change the words)
Voiceover in five lines. Captions burned in, lower third.
Beat 1, kitchen hook. "Ever wondered why tile floors pop up for no reason? It's not bad luck. It's physics."
Caption: Why do tiles pop for no reason?
Beat 2, the mechanism, two clips. "Tile and the concrete under it expand and shrink at different rates. Laid tight, wall to wall, that movement has nowhere to go."
Caption 2a (two lines): Tile grows. / Concrete grows differently.
Caption 2b: Laid tight, it has nowhere to go.
Beat 3, the pop. "The pressure builds for years. Then the weakest spot lets go, and the whole row gives way. Usually at two in the morning."
Caption (two lines): Years of pressure. / One bad night.
Beat 4, the rule, two clips. "A good installer knows this. Full mortar under every tile, and a flexible joint at every wall and every twenty-five feet. Cheap on day one. Impossible later."
Caption 4a: Five dabs of mortar is not a floor.
Caption 4b (two lines): Full mortar. / Movement joints every 25 ft.
Beat 5, end card. "Cornerstone Home Builders lays every floor to that rule, and pulls a tile to prove it. Comment CHECK for the five things to look at before you pay a tile guy."
Caption: Comment CHECK
## Step 1: voiceover, music, sound effect
Voiceover: `fal-ai/elevenlabs/tts/eleven-v3`, voice "Callum", stability 0.3, one job per line, text exactly as in the script. Then speed each file up by 6 percent with ffmpeg (`atempo=1.06`, pitch unchanged) and measure the durations; the clip lengths in step 3 come from these.
Music, two jobs on `fal-ai/stable-audio-25/text-to-audio`:
- Bed, seconds_total 60, prompt: "Continuous soft cinematic ambient pad, warm sustained synth strings with a slow gentle pulse, steady and unbroken with no silent gaps, no melody, no drums, subtle curiosity, instrumental, background music bed"
- Sting, seconds_total 14, prompt: "Warm resolved acoustic guitar and soft piano sting, hopeful and confident, light brushed percussion, instrumental, gentle build, clean ending"
Sound effect, one job on the same endpoint, seconds_total 4, prompt: "A single sharp dry crack of a ceramic floor tile snapping and lifting, one short impact with a little debris settling, then silence, no music". Trim it to the crack.
## Step 2: stills (Nano Banana Pro)
Parameters for every image job: aspect_ratio "9:16", resolution "1K", output_format "png". Save each image locally.
Style suffix for the anchor: "Adult animated sci-fi sitcom cartoon style: thick, slightly wobbly black outlines, flat cel colours with minimal shading, exaggerated oval eyes with tiny pupils, wide wobbly mouths, slightly grotesque lanky proportions, a faint sickly green-teal tint in the lighting, 2D hand-drawn look. Original character, not from any existing show. No text, no letters, no watermark. Vertical 9:16 with the bottom fifth of the frame kept simple."
Anchor (beat 1), `fal-ai/nano-banana-pro`, prompt: Doc's design, then "He stands in a cartoon suburban kitchen with white cabinets and grey floor tiles, pointing down with his free hand at a row of three floor tiles that have lifted into a tent shape with cracked grout, deadpan. Medium-wide shot, character on the right third, tented tiles on the left." then the style suffix.
Every other still is a `fal-ai/nano-banana-pro/edit` job with the anchor as image 1. Two suffixes are used:
Suffix K (Doc keeps his framing): "Same adult animated sci-fi sitcom cartoon style as image 1: thick slightly wobbly black outlines, flat cel colours, exaggerated eyes with tiny pupils, faint sickly green-teal tint, 2D hand-drawn look. Keep the scientist's face, hair, goggles, lab coat and clothes exactly the same as image 1. No text, no letters, no labels, no watermark. Vertical 9:16 with the bottom fifth kept simple. Output ONE single scene."
Suffix S (Doc small and turned away, which keeps the video model's likeness checker happy): "Same adult animated sci-fi sitcom cartoon style as image 1: thick slightly wobbly black outlines, flat cel colours, faint sickly green-teal tint, 2D hand-drawn look. Keep the scientist's hair, goggles, lab coat and clothes the same as image 1, but show him smaller in the frame and turned three-quarter away from the camera toward the screen so his face is seen from the side at a distance. No text, no letters, no labels, no watermark. Vertical 9:16 with the bottom fifth kept simple. Output ONE single scene."
Let BOARD mean: "a large floating holographic green screen beside him showing a clean side-view cross-section diagram of a tile floor: a row of tiles on a thin mortar layer on a thick concrete slab, wall stubs at both ends", and BIGBOARD mean the same but "a very large floating holographic green screen that fills most of the frame, showing ...".
- Still 2a (suffix S): "New scene: wide shot, the scientist small at the bottom right with his back half turned to us, gesturing up at BIGBOARD, a small sun icon above the tiles, long bold orange arrows pointing outward along the tile layer and much shorter orange arrows pointing outward along the slab."
- Still 2b (suffix K): "New scene: the scientist on the right grimacing, BOARD with the tiles jammed edge to edge against both walls, two bold orange arrows pushing inward from the walls, an orange glow of compression at the middle joint."
- Still 3 (suffix K): "New scene: the same cartoon kitchen at night, dim with one light on, the scientist standing on the right holding his flask, the grey floor tiles completely flat and intact with no damage, a coffee mug on the counter."
- Still 4a (suffix K): "New scene: the scientist holds up a grey floor tile turned around to show its back, which has only five small round dabs of dried mortar on it and bare tile between them, his eyebrow raised; beside him BOARD showing a single tile sitting on a full continuous mortar bed on a concrete slab."
- Still 4b (suffix S): "New scene: wide shot, the scientist small at the bottom right seen from behind at three-quarter angle, tapping a long pointer on BIGBOARD where teal flexible joint strips sit between the last tile and the wall at each end and one in the middle of the row, orange arrows along the tiles shrinking into the teal strips, the tiles flat and calm."
- Still 5 (suffix S): "New scene: outdoors at golden hour, wide shot, a white jobsite sign on two wooden stakes fills the centre of the frame in front of a cartoon timber-framed house, the sign has a navy border, a simple navy cornerstone block emblem and the words exactly 'Cornerstone Home Builders' in bold navy capitals with smaller text exactly 'Licensed and Insured' beneath; the scientist is small in the background at the far right, walking away from the camera toward the house with one hand raised in a lazy wave, seen from behind. The sign text is the only text in the image."
Check every still for stray text, grids, and that Doc's hair, goggles and coat match the anchor. Re-roll a still at most once if it fails; if an edit is refused, reword it slightly (never add anything, only simplify) and retry once.
## Step 3: video clips (Kling v3 Standard)
Endpoint `fal-ai/kling-video/v3/standard/image-to-video`, generate_audio false, cfg_scale 0.6, negative_prompt "text, letters, labels, watermark, photo, 3D render, blur, extra characters, morphing, melting". Start image is the matching still. Duration is a whole number of seconds from 3 to 15, set as: beat 1 = line 1 duration + 0.8, rounded up; 2a and 2b each = half of (line 2 + 0.8), rounded up; beat 3 = 8; 4a and 4b each = half of (line 4 + 0.8), rounded up; end card = line 5 + 1.8, rounded up, capped at 15. With Callum at this pace that comes to about 8, 6, 6, 8, 8, 8 and 14 seconds.
Every prompt starts with: "2D adult sci-fi sitcom cartoon animation, flat cel colours, thick wobbly outlines, the same scientist character throughout with the same face, hair, goggles and lab coat, static camera, no text. "
- Clip 1 (anchor): "The scientist points at the tented tiles, then looks up at the camera and talks with a smirk, small hand gestures, his mouth moving loosely as he speaks, the green flask bubbles gently. Nothing else in the kitchen moves."
- Clip 2a (still 2a): "The scientist gestures up at the screen with small movements; on the screen the sun icon glides slowly across the top, the long orange arrows along the tiles grow slightly longer while the short arrows along the slab stay short, every arrow keeps its exact shape."
- Clip 2b (still 2b): "The scientist grimaces and talks, pushing his two hands together to mime a squeeze; on the screen the orange arrows push inward from the walls and the orange compression glow pulses brighter."
- Clip 3 (still 3): "Night kitchen. For the first three seconds the scientist talks calmly to the camera holding his flask. Then the row of floor tiles in front of him suddenly snaps upward into a tent shape with a small puff of dust; he flinches and a little green liquid sloshes out of the flask, then he looks at the camera deadpan."
- Clip 4a (still 4a): "The scientist waves the tile with the five mortar dabs at the camera while talking, eyebrow raised, then jabs a thumb toward the holographic screen showing the full mortar bed."
- Clip 4b (still 4b): "The scientist taps the pointer on the teal joint strips; on the screen the orange arrows travel slowly toward the teal strips and shrink into them, the tiles stay flat, nothing else changes."
- Clip 5 (still 5): "Golden hour, gentle breeze in the grass, the scientist in the background strolls slowly away toward the house with his hand raised in a lazy wave, the sign and its text stay exactly as they are and perfectly legible, the camera slowly pushes in on the sign."
Fallback rules: if a clip is refused by the content checker, re-edit its still with suffix S (Doc smaller, turned away) and run the clip again once; if it is refused a second time, make the still itself the clip with a slow ffmpeg zoompan of the same length. If an arrow clip melts the diagram into a blob, re-roll once with "locked camera, very restrained motion, every arrow keeps its exact shape and colour" added. Check three frames of every clip (start, middle, end) for identity, no text, and correct arrow direction: tile arrows longer than slab arrows in 2a, inward in 2b, shrinking into the teal strips in 4b.
## Step 4: captions and assembly (ffmpeg)
Render the seven captions as transparent 1080x1920 PNGs with Pillow: Arial Black 58 px, white fill, dark stroke 7 px, centred, baseline around y=1430 (two-line captions centred on that point with 72 px line spacing). Overlay each caption on its clip from 0.3 s onward; clip 4b carries caption 4b.
Assembly rules:
- Scale and pad every clip to 1080x1920, 30 fps, yuv420p. Hard cuts between clips; 0.15 s audio fades at the start and end of every segment. Clip 4b fades to white over its last 0.3 s and the end card fades in from white over 0.3 s.
- Order: clip 1, 2a, 2b, 3, 4a, 4b, then clip 5 as the end card.
- Voiceover placement: line 1 at the start of clip 1 plus 0.4 s; line 2 at the start of clip 2a plus 0.4 s; line 3 at the start of clip 3 plus 0.4 s; line 4 at the start of clip 4a plus 0.4 s; line 5 at the start of the end card plus 0.6 s. A line never starts before the previous line has finished plus 0.25 s; if it would, push it later. Each line normalised to -16 LUFS.
- The crack sound effect sits at the start of clip 3 plus 3.0 s, peaking around -6 dBFS.
- Bed at -20 dB from the start until 1 s into the end card, fading in over 1.5 s and out over 2 s. Sting at -12 dB starting 2 s before the end card, fading in over 2 s and out over the last second.
- Final mix: loudness normalised to -16 LUFS integrated, true peak -1.5 dBFS.
- Export `cornerstone-tile-tent-doc-cartoon.mp4` (H.264, CRF 18, AAC 192k) and `cornerstone-tile-tent-doc-cartoon-preview.mp4` under 30 MB (CRF 24).
Verify before finishing: the whole of line 5 finishes with at least 0.8 s of music after it (extend the end card with a freeze of its last frame if not), no caption overlaps a face, no level jump across cuts, total length about 55 s.
## Deliverables
The master and the preview in this folder, plus subfolders `stills/`, `clips/`, `audio/`, `captions/`, and a `PRODUCTION-LOG.md` with the models used, the spend per step, anything refused or re-rolled and how it was handled. Finish by telling me the master's path, its length, and the total spend against the 12 USD cap.
