# how_it_was_made
Storyboard first (verbatim dialogue + per-shot staging, deliberately one step before prompts), then 10 character sheets generated from each person's real profile picture, then 3 background plates (street, room, night street), then 19 scene stills that reference both the plate and the locked character sheet, then 17 video clips using each approved still as the start frame, then post: TTS voices, ambience beds, a built crowd chant, lower-thirds, captions, end cards.
~13,700 Magnific credits total. Stills on Nano Banana 2 Flash at 75 each. The 13 short replier clips on Kling 2.5 at 140 each; the 4 hero shots on Seedance 2.0 Pro at 1,680–2,240 each.
The decision that paid for the project: every clip was generated SILENT and voices were layered on in the edit. Cutout-cartoon mouths are two-state flaps, so real lip sync — the expensive capability — is unnecessary. All 17 clips on the premium model would have cost ~35,000 credits. The cheap tier came to 1,820. The art style paid for itself.
Finishing was done in Descript, with overlay placement driven by an agent against the transcript's speaker labels, and every visual/audio taste call made by hand.
// best_dead_end
The orb. One replier's profile picture is an abstract orange-and-white sphere, so the character is a floating orb that hovers at head height, says 'ship,' and is acknowledged by nobody. Three animation attempts, three different failures: attempt 1 bounced it down the street like a basketball with dust puffs; attempt 2 held position but slowly rotated 180°; attempt 3 grew a second body sphere and a coal smile and became a snowman.
I stopped burning credits and shipped attempt 2 trimmed to its clean 1.15-second window with a freeze extend. The model's total refusal to leave an abstract shape alone was, in the end, the correct comedic instinct to overrule: the joke was always stillness.
# prompting_approach
Three things did most of the work.
1. REFERENCE CHAINING, not re-description. Never describe a character twice. Avatar → locked character sheet → every scene still references that sheet → every clip uses its still as the start frame. Consistency is inherited rather than re-prompted.
2. A LOCKED BASE BLOCK per stage, prepended to every prompt. For stills: the construction-paper cutout style, flat colors, hard outlines, 16:9, no text anywhere. For motion: 'stiff, limited, slightly jerky cutout motion; mouths flap open/shut in simple shapes; NO smooth realistic motion, no 3D rotation, background completely static except where named.' Without the motion block, video models smooth everything into semi-3D and the cheapness that makes the style funny disappears.
3. PHASE GATES WITH COST SIMULATION. Nothing advanced until the previous phase was approved, and every batch was priced before it ran. Cheap tier first as a proof of the motion style, premium only after.
A rule that emerged the hard way: state what must NOT move. 'Only the gentleman moves — the bird and all background elements stay completely frozen' had to be explicit, because otherwise the model picks whatever element it finds most animatable.
