The strategy, in 10 rules
1 · One readable shot
Subject → visible action → setting → light → camera → continuity. One primary action per clip; pack nothing more.
2 · The 70/30 heuristic
70% subject + core action, 30% camera + environment. Over-specifying geometry causes subject disintegration.
3 · One camera idea per shot
Bracket commands ([Push in] [Pan right] [Tracking shot] [Static shot]) are legacy V1 syntax; H3 wants natural language, max ~3 combined.
4 · Image-to-video beats text-to-video
A strong start frame anchors style and composition. Use our render stills as start frames whenever the lane supports it.
5 · Short clips, slow motion
4–10 s maximizes temporal stability; a “slow” modifier reduces frame-to-frame variance and drift.
6 · No negative_prompt field
MiniMax schemas don't document one. State constraints in the main prompt — “one continuous shot”, “the vessel stays centered”.
7 · On-screen text = gibberish risk
Gen models mangle text. Generate text-free frames; overlay Fraunces/Plex Mono titles in post. (Our UI films stay deterministic for this reason.)
8 · Describe exact text if text must appear
When unavoidable: “The title reads: …” verbatim, short words only, expect rerolls.
9 · Iterate one variable at a time
Branch from an approved take; change one thing per reroll — never rewrite the whole prompt.
10 · 2,000-char limit (API)
Studio prompts can be long but stay under ~2,000 characters; keep the shot definition in the first third.
Sixteen concept prompts
Every prompt is palette-locked to the Induction system (obsidian #0a0a0f · iris #6c7cff → magenta #c95cff → gold #c9a961 · live green #3dbe8b) and maps to a law from the front-end practices reference. Copy → paste into Studio → generate → overlay text in post.