Teardown — He Dived Into Freezing Kelp Every Day for One Octopus: 10 cuts, film + clips + every prompt
10 cuts across 2 motion tiers, 80% on tier 2. A short-form episode, published, with every generated clip (10), every source still (10), and the exact still and motion prompt behind each cut.
One published episode opened up: the finished film, 10 generated clips, 10 source stills, and the prompt behind each of its 10 cuts.
What is in it
- film/ — the finished episode — the published cut, exactly as it went out
- clips/ — 10 generated clips — every shot before the edit, including the plain ones
- stills/ — 10 source frames — the image each clip was animated from
- PROMPTS.md — one block per cut across all 10: framing, tier, still prompt, motion prompt
- production.json — the same record machine-readable, one object per cut
Claims, and how to check them
- every one of the 10 cuts has its prompt included
check: count the blocks in PROMPTS.md, or the objects in production.json - the archive holds 10 clips and 10 stills
check: list the zip; these counts were read from it - the episode is published and watchable before you buy
check: the listing links the public cut
What it does not do
- narration and on-screen text are English; the prompts and all documentation are English
- clips are re-encoded for download, not master bitrate
- no guarantee of a result — model behaviour changes between versions
- not a template: this is one episode, not a reusable standard
You need already
- nothing to read it — markdown, JSON, MP4 and JPEG
- an i2v generator and your own credits, to re-run anything
clips: 10cuts: 10format: shortsnarration: Englishstills: 10tier_split: dynamic:8 / static:2
Every AI video tutorial shows you the prompt that worked. This shows you all 10, including the dull ones.
One published animal bond short — a short-form episode — opened up. The finished film, all 10 generated clips, the 10 stills they were animated from, and for each of the 10 cuts the exact still prompt and motion prompt that produced it, beside the engine tier it was ordered on.
Watch the published cut before you buy: https://youtu.be/aiXKn1Ioya0
WHY THIS GENRE IS WORTH TAKING APART
A human hand touching an animal is the hardest thing to generate, and this genre is built entirely on that moment.
Look at which cuts were promoted to the expensive engine. Contact shots are, every time — saving money there costs more in re-orders.
WHAT IS IN THE PACKAGE
• The finished film, as published
• 10 generated clips — the raw shots before the edit
• 10 source stills — the frames each clip was animated from
• PROMPTS.md — one block per cut: framing, tier, still prompt, motion prompt
• production.json — the same record, machine-readable
THE ROUTING, IN THIS EPISODE
Cut split: 2 on tier 1, 8 on tier 2. Tier is assigned before a cut is ordered, and it decides the engine. A static shot on the expensive engine is pure waste; a hard-motion shot on the cheap one costs more, because the re-order costs more than the correct order would have.
WHAT THE EPISODE ARGUES
An octopus has about 500 million neurons — two thirds of them spread through its eight arms — no bones and three hearts, yet it can recognize an individual human and choose to reach out and trust them.
THREE THINGS YOU CAN READ STRAIGHT OFF THE PROMPTS
• Identity is locked with a positive count, never a negation — a forbidden thing tends to get summoned.
• Framing is stated as what fills the frame, not as a shot-size name. Shot-size vocabulary is interpreted loosely by every engine we have measured.
• Subject, action and environment are separate lines in the motion prompt. Collapsing them into one sentence is where drift starts.
HONEST NOTICE
• Clips are re-encoded for download, not delivered at master bitrate. They are reference and working copies.
• You may use these files in your own commercial and personal work, including client work, and modify them freely. You may not resell or redistribute the package itself, in whole or in part, as a product — including uploading the clips to a stock library.
• This is a record of what one episode took, not a guarantee of a result. Generation models change behaviour between versions, sometimes reversing it, so treat the routing as a procedure you can re-measure rather than a fixed ranking.