Reference E · House rules
Reference
The fifteen articles the author ratified on 18 July 2026, printed exactly as ratified, with the amendment and the ruling of 30 September 2026 that carries it to Seedance 2.5. Then the rules this book teaches, stage by stage, each with the section that teaches it, and the author's decisions of 30 September 2026.
In this reference
The fifteen ratified articles and Amendment A-01, exactly as ratified, with the ruling that A-01 now covers Seedance 2.5
The book's rules by stage, one line each, with the section that teaches it
The author's decisions of 30 September 2026
How to read it. The articles are law. Nothing in this book overrides one, and only a dated ruling of the author changes one. Every other rule here is a line from the chapter named beside it; the chapter gives the reason and the way to apply it. Where a rule depends on something that can change (a version, a limit, a price), the chapter carries the date it was checked, and this page does not repeat it.
E.1 The fifteen articles, as ratified
The articles are printed exactly as ratified on 18 July 2026, so they speak in the terms of the July research they were written in: blocks, bookends, runs, the V7 prompt file. The plain-words table after them says what each asks of you at the desk, and where the book teaches it.
This Constitution records the current director-ratified production laws. It distinguishes those laws from external platform facts: a creative or operational policy can be settled here, while a provider capability conflict remains conservative, with its verification record in the companion evidence file. Each article states what it governs and supersedes. The current set is ratified as amended on 18 July 2026; later change requires an explicit dated amendment rather than silent drift. C-01· The opening slot. The still prompt opens with the Medium Declaration. The Subject & Action block must declare the asset's physical temporal state: a frozen instant for a moving shot keyframe, or a justified STATIC state for a sheet, product plate, held shot, or other non-moving reference. Evidence: the inspected production corpus contains 106 still prompts: 74 in Run 1 and 32 in Run 2. The same corpus contains 93 motion prompts: 61 in Run 1 and 32 in Run 2, for 199 prompts total. The 106 figure refers only to still prompts and is not a live-output success count. Early-token priority is a production synthesis, not a universal measured law. Supersedes: megaprompt Phase-8 order item 1; the V7 prompt file's still-order item 1. C-02· Bookends and parameters. Opening bookend = the Medium Declaration; closing bookend = the Finish Block plus the Exclusions line — a technical close, never an emotional one. Aspect ratio and output resolution are route-specific controls recorded in the shot sidecar and generation receipt. They remain outside prompt prose unless the exact native or provider grammar explicitly requires them there. Higgsfield, fal, native applications, and other providers are each ground truth only for their own recorded route; none is global truth. Flags: any validator that treats one wrapper schema as universal. C-03· Intent lives in metadata. story_intent· emotional_beat· goal_of_shot occupy the INTENTION field; the paste-prompt stays physical. The one sanctioned trace of intent inside the prompt is a physically phrased production context, carried by the Medium Declaration ("a hero frame for a premium athletic-wear commercial") — a form official guidance recommends to help the model infer the intended register. Supersedes: the stored-preference placement of context blocks inside prompts; production practice already worked this way. C-04· Density. There is no universal word-count floor. A prompt is complete when it satisfies the applicable workflow grammar, preserves every required anchor and state, and passes the block or field audit without padding or contradiction. The 32 inspected Run-2 still prompts measured 390–598 words with a median of 500; that is a useful target band for complex, identity-sensitive cinematic stills, not a universal platform law. Each task declares an evidence-informed working band appropriate to its complexity, and completeness outranks length. Supersedes: the ≥250 and ≥350 absolute floors and any validator that equates word count with compliance. C-05· Order. World first, photography second, technical close — the twelve blocks of §2.2. Identity and content sit early where attention is strongest; camera, light, atmosphere, finish, and exclusions follow an established world. Supersedes: the stored grammar's early-camera placement. C-06· Motion quantification — FINAL. Approximate single-magnitude quantification is permitted and encouraged for subtle moves: one distance OR one angle, prefixed "approximately," plus duration and a speed profile ("advancing approximately fifteen centimeters over five seconds, a settling forward motion"). Timed phase segments ("0–1 s: … / 1–4 s: …") are additionally supported for phased moves. Banned: stacked simultaneous metrics, coordinate-style precision, and contradictory spatial instructions — the documented warp trigger. Adjectival phrasing remains fully valid for bold moves. Evidence: the Run-2 corpus uses the instrument across 32 motion prompts; external guides demonstrate timed quantification and attribute warping to contradiction and motion intensity, not to numbers. Supersedes: the V7 blanket ban on travel distances. C-07· Motion from the start. The behavior is law: motion begins immediately, never "begins in stillness, then…". The literal string "from frame one" stays banned in prompt text (field learning — the phrase itself misbehaves). C-08· The Run-2 baseline. Forensic likeness micro-anatomy, quantified motion, rig-physics anti-float, and multi-element mid-action status are the baseline standard of this guide. Run 1 is superseded practice, retained as evidence. C-09· The universality filter. Every rule is classified UNIVERSAL or PROJECT before adoption; project circumstances never masquerade as law. Mastorna's project set: dolly/crane-only camera systems, 1.85:1, B&W tonal clauses, restored-print finish. The universal patterns extracted from them stand as law (period-lock the vocabulary; aspect from the project master; a Finish Block always present, always project-defined). This article applies recursively to this guide itself — its first catch was Build 1's over-generalized text law, corrected by the director into Article C-11. C-10· Negatives and quality spam. ≤5 exclusions per prompt, surgical, positives exhausted first. No quality-tag spam ("4K, masterpiece, ultra-realistic, HDR") — counterproductive on a reasoning model; realism is earned through physical description; resolution is a parameter. C-11· THE TEXT TWO-LAYER LAW (corrects Build 1's LAW 9; source: the director, July 2026). Diegetic world text is production design: generated deliberately when the shot wants it, specified like any material — language, script, era, style, wear — with honest distance-based legibility and per-character QC; NB Pro documents multilingual text rendering as a capability, subject to per-character output QC. The editorial layer — titles, subtitles, supers, end-cards, watermarks, logo lockups as graphics — is never generated: composited in post from real assets. Product brand marks route through reference-driven generation from the real asset; compositing is the legal-grade fallback. Text survivability is assessed for every animated shot; Kling 3.0 documentation describes source-image signage preservation, including under camera movement; every returned clip still requires glyph-level review. The mandatory closing exclusion becomes "no overlaid text, subtitles, logos, or interface graphics" — it bans the overlay layer only; whether the world carries legible text is a per-shot production-design decision. Supersedes: Build-1 LAW 9; the V7 R12 blanket form; the legacy closing line. C-12· Style-token scoping. "Cinematic," "[director]-esque," and kin are banned in authored description (Law 2). The corpus-recorded t2v opener formulas are retained as bounded artifacts adopted as whole units — exempt as artifacts, never as license to write vaguely. C-13· Actor-anchor scoping. Real-name likeness anchoring ("modeled on …") is a project-level rights decision: natural for historical reconstruction, default OFF for branded and commercial work where likeness rights govern. The forensic micro-anatomy layer is the universal likeness engine either way; the Real-Person Protocol governs living persons above this guide. C-14· Motion-state language. Hedge-state camera descriptions ("barely perceptible," "largely still," "mostly motionless") are banned as defaults. A genuinely fixed camera is legal when declared as design: name the fixed frame and assign all motion to the world ("static tripod frame; all movement belongs to the scene: steam, flicker, breath"). Stillness is a decision, never an accident — in stills (STATIC metadata) and in motion alike. Refines: the megaprompt's banned-word list, from absolute to scoped. C-15 · References and verbal anchors are workflow-scoped. Full verbal description is mandatory when the call establishes or changes pixels: text-to-image, sheet and anchor generation, text-to-video, and image or video editing. In ordinary image-to-video, the approved source frame owns composition, appearance, and identity; the motion prompt names preservation requirements and adds only motion, timing, physics, endpoint, and surgical negatives. It does not fully re-describe the frame and fight the source. Exact model and provider reference-role syntax still governs. Resolution: the former universal re-description rule and conflict are superseded by this director-ratified scope, 18 July 2026. These fifteen articles are the constitution of the generation system, ratified as amended by the director on 18 July 2026. Later evidence may refine external platform facts; changing production law requires an explicit dated ruling. Every later part obeys the current wording above.
AMENDMENT A-01 to C-12 — ratified by the director, July 2026. Named-director technique tokens (e.g., "Guy Ritchie speed-ramping," "Snyder impact slow-motion") are permitted in Seedance 2.0 motion prompts only, strictly as technique tokens naming an editing or motion style. They remain banned as visual-style adjectives everywhere, banned in NB Pro stills, and unproven on Kling — treat as banned there pending evidence. Every use remains subject to the anti-cliché discipline. Recorded with it as precedents, rulings 1–3 of the same session confirmed existing law: natural-language briefing over command-syntax; C-10's ≤5 surgical negatives over exhaustive lists; C-02's params rule over aspect-in-prose.
Amendment A-01 and Seedance 2.5
Ruling of 30 September 2026. Amendment A-01 carries over from Seedance 2.0 to Seedance 2.5. Named-director technique tokens (a speed-ramp, an impact slow-motion) are allowed in Seedance 2.5 motion prompts only, strictly as the name of an editing or motion style, never as a visual style. They remain banned as visual-style adjectives everywhere, and they remain banned on Kling and in Nano Banana Pro prompts.
The articles in plain words
Article | In plain words | Taught in |
The opening slot | A still opens with the medium and the purpose, and declares its physical state: a frozen instant if it will start a clip, a stated static state if nothing moves. | 9.2 The twelve elements; 9.4 The frozen instant, and the frame inventory |
Bookends and parameters | The prompt opens with the medium and closes with the finish and the exclusions, a technical close and never an emotional one. Aspect and resolution are settings of the route, not prose. | 9.1 What this stage decides, and where each decision lives; 9.5 The finish, emotion, exclusions and length |
Intent lives in metadata | Why the shot exists goes in the shot record. The prompt carries what a camera can see, and at most one physically phrased line of production context. | 9.1 What this stage decides, and where each decision lives; 8.6 The shot record |
Density | No word-count floor. A prompt is finished when every element the shot needs is settled; a complex, identity-bearing still measured 390 to 598 words in finished work. | 2.8 How long; 9.5 The finish, emotion, exclusions and length |
Order | World first, photography second, technical close: the twelve elements of Chapter 9. | 9.2 The twelve elements |
Motion quantification | One approximate magnitude, a duration and a speed profile for a subtle move; never stacked metrics or contradictory instructions. | 14.5 Measured moves: sizing one, and measuring it; 16.3 The prompt, in this tool's dialect (Kling) |
Motion from the start | Something visible moves at once; the phrase "from frame one" stays out of prompts. | 14.4 The seven elements, in order |
The Run-2 baseline | Forensic likeness, quantified motion, rig-physics anti-float and multi-element mid-action status are the standard. | 10.4 The identity clause; 14.4 The seven elements, in order; 14.6 Rig physics: give the move a body |
The universality filter | Sort every rule into universal or project-only before adopting it. | 8.2 Decisions made at the start of every project |
Negatives and quality spam | At most five surgical exclusions; no quality tags. | 2.7 Exclusions and the closing line |
The text two-layer law | The world's text is production design and may be generated; the editorial layer is never generated and is composited from real assets. | 10.12 Text in the frame: two layers |
Style-token scoping | "Cinematic" and "in the manner of a director" are banned in description. Amendment A-01: a named director's technique token is allowed in Seedance motion prompts only. | 2.6 Borrowed words, and other people's prompts; 17.3 The prompt, in this tool's dialect (Seedance 2.5) |
Actor-anchor scoping | Naming a real actor as the likeness is a rights decision for the project, off by default in commercial work. | 10.8 Real people, and an actor's name as a rights decision |
Motion-state language | No hedged camera. A genuinely fixed camera is legal when declared as design. | 14.4 The seven elements, in order; 14.6 Rig physics: give the move a body |
References and verbal anchors are workflow-scoped | Describe fully when the call makes or changes pixels; in ordinary image to video, the frame owns appearance and the prompt only moves it. | 9.10 Keyframes for motion; 14.1 Who owns what |
E.2 The book's rules, by stage
One line each. The section named is the one that teaches the rule and gives its reason.
E.2.1 Working with a model
Chapters 1 and 2.
Rule | Taught in |
A model reads your brief once, literally, with no memory. Everything it needs is in the prompt or attached to the job. | 1.1 A camera that reads; 2.1 Every model is a different reader |
Each model is a different reader. Write for the one in front of you, in the dialect its tool chapter teaches. | 2.1 Every model is a different reader |
The inputs own what they carry; the words carry the rest. Never let the words and the inputs disagree. | 2.2 The inputs own what they carry; the words carry the rest |
Two grades of prompt. The finished grade is the standard; the test grade only proves a route and is never shown as the model of a finished frame or clip. | 2.3 What every prompt settles; 9.6 The test grade, and when it is right |
Write each instruction once, in the place that owns it: intention in the shot record, settings in the interface, what a camera would see in the prompt. | 2.3 What every prompt settles; 9.1 What this stage decides, and where each decision lives |
Write natural language, as a brief a camera operator could act on. Keep command syntax and keyword lists out. | 2.4 Choose the writing format by the job |
Put each number beside the result it should give. A number is a cue for the look, never physics and never alone. | 2.5 Numbers, each beside its result |
Translate borrowed words into what the camera does. "Cinematic" never describes a frame. | 2.6 Borrowed words, and other people's prompts |
Say what is there first. At most five exclusions on any prompt (three or fewer is the working advice on a clip), each aimed at a drift you expect; then the closing line "no overlaid text, subtitles, logos, or interface graphics". | 2.7 Exclusions and the closing line |
No floor and no ceiling on length. A prompt is finished when everything the shot needs is settled, and not padded. | 2.8 How long |
When a reworded attempt returns the same defect, stop rewording and change what you give the model. | 2.9 Where to stop prompting |
Only the director signs a frame, locks a take or the cut, and approves a spend. The operator writes, runs and records every job. | 1.4 The director's job now; 1.5 The crew, and who uses which tool |
The cheapest attempt is a still. Sign the frame before you pay for motion, and repair the frame rather than re-rolling the clip. | 1.1 A camera that reads |
AI may be in the work and is never used to fabricate results. Disclose its use as the client and the platform require. | 1.6 What the industry has settled |
E.2.2 Concept and treatment
Chapters 3 and 4.
Rule | Taught in |
Rewrite the brief before you answer it, and separate its facts from its claims. | 3.3 Reading the brief, then rewriting it |
Push observations until they become tensions, with an enemy that is not a competitor. | 3.4 From observation to tension |
Write the refusal list before any idea, and ban the five expected answers. | 3.5 The refusal list comes before any idea |
Compare ideas with reasons, not scores. Name a recommendation and the strongest alternative. | 3.7 Testing an idea |
Choose the idea before you render it. Treat feasibility as information, never as a filter. | 3.8 What changes when anything can be generated |
The concept page records what may not change. | 3.9 The concept page, and handing it on |
Write a treatment on two rails: every feeling gets a sentence that begins "On camera this means". | 4.2 From feeling to what the camera sees |
The scene breakdown is the heart of a treatment: one paragraph per scene, in the present tense, with visible events, dialogue exact in the film's language and the voice-over in its own paragraph. | 4.4 The scene breakdown |
Specify the first two seconds, and time every line by reading it aloud. | 4.5 The first seconds, and how long a line takes |
Lock the laws of the world and leave the life inside it to be discovered. A protected moment is not changed by a later version. | 4.1 What a treatment is for, and who reads it; 4.6 Departments, the lock, and what is left to discover |
Read a reference for its craft logic and borrow one thing from each. Caption and credit every image. | 4.7 Visual references, and how to read them |
In a deck, show before you explain. A frame study is a promise about a beat, and a client edition carries no prompts, model names or working codes. | 4.8 Presenting the film; 4.9 The fresh read, the records and handing on |
E.2.3 Shots, light and the world
Chapters 5 to 7.
Rule | Taught in |
Every shot does a job no other shot does. Name the job before the size, and name the generation risk and the plan before you spend anything. | 5.1 What a shot list decides, and who decides it; 5.2 Writing the breakdown |
Write a move as a verb, a mechanism, a destination and a hold. Write "The camera is locked." when it should not move, and never a hedge. | 5.4 The move |
Keep camera-body, lens-brand and people's names out of prompts. Height, distance and lens class are emotional decisions. | 5.7 Placement and lens |
Fix the geography in the first second, keep one geometry note per scene, and originate vertical frames for speech and travel. | 5.8 Geography and format |
Name every light: what it is, where, hard or soft, its Kelvin, what it does to a named surface. Name the fill by the surface that causes it. | 6.2 Sources: one cause for each effect |
Give one contrast ratio with its reason, and expose for the skin first. | 6.3 Contrast and exposure |
Keep two light worlds on skin, not on the walls. | 6.4 Colour temperature, and two light worlds |
Plan the palette as proportions and write it as named surfaces, never as the string "60/30/10". | 6.5 Palette, brand colour and the grade |
Write the finish in the prompt as capture behaviour, with grain kept fine. Lay delivery grain and halation once, last, in post. | 6.6 The finish, in two layers |
Read a reference for its light, colour and finish separately, and borrow one dimension from each. Make the look in the still, once, in words. | 6.7 Reading a reference for its light; 6.8 Building a look, and briefing it in words |
Give the world a history and put wear where hands go, by place, in three or four material families. | 7.1 Why a generated world looks built; 7.2 Surfaces |
Describe clothes by material, construction and wear. Skin is behaviour under light, judged at viewing size. | 7.3 People: clothes and skin |
A product touches the world through scale, contact, overlap and shared light. Give every object that matters a state, and keep a ledger. | 7.4 Objects and the product |
Edges either carry a meaningful interruption or disappear quietly. Keep a period true in every object and every department. | 7.5 Density and the edges of the frame; 7.7 Period truth |
Run the realism check before you sign a frame. | 7.9 The check before you sign a frame |
E.2.4 Planning the production
Rule | Taught in |
Four gates stop a production: LOCK, SPEND, RIGHTS and ACCEPT. Keep a decision log that includes what you rejected. | 8.1 Where a change is cheap, and the four gates |
On the first day decide four things and lock them on one card: the speaking-shot contract, the delivery specification, the rights class and the look. | 8.2 Decisions made at the start of every project |
Write the obligations first. Decide each department's choices once and carry them into every prompt word for word. | 8.3 Obligations before footage; 8.4 From the treatment to one production plan |
Keep three linked records: the shot table (the plan), the shot record (the prompt and its fields) and the run record (every attempt, kept the day it runs). | 8.5 The shot table; 8.6 The shot record; 8.7 The run record |
Make the pictures in order: neutral masters, then in-look and environment anchors, then the shots, re-attaching the first-generation masters each time. | 8.8 The continuity plan, and the order of work |
Use plain, stable names, one workbook and five tracker states. | 8.9 Names, folders and the workbook |
Open every paid batch with a spending packet. Write stop rules as conditions, and price the shot at its true cost. | 8.10 The spending packet and the budget; 34.3 Spending inside an approval |
E.2.5 Stills
Chapters 9 to 12.
Rule | Taught in |
A finished still is twelve elements in a fixed order in one paragraph: world first, photography second, technical close. | 9.2 The twelve elements |
Write camera and light as a specification, each number beside its visible result. | 9.3 Numbers beside results |
Freeze a moving shot's keyframe 20 to 40 per cent into the action, written as positions. Write into the prompt every system the record freezes. | 9.4 The frozen instant, and the frame inventory |
Name the foreground, midground and background, and what touches or leaves each edge. | 9.4 The frozen instant, and the frame inventory |
The finish describes capture. At most five exclusions, then the closing line. Delivery texture is added in post. | 9.5 The finish, emotion, exclusions and length |
Aspect and resolution are settings of the job, never words in the prompt. | 9.1 What this stage decides, and where each decision lives; 11.6 Settings (Nano Banana Pro) |
Two attempts, then change the lever: an edit from the signed parent, a composite in post, a new picture, or another tool. | 9.9 Repairing a frame until you can sign it |
A keyframe owns everything that exists; the motion prompt owns everything that changes. A speaking frame carries the mouth state, and a contact state shows the gap. | 9.10 Keyframes for motion |
Sign on the file, at both sizes and against the prompt, and record the frame so the next one can be checked against it. | 9.12 Signing and handing on |
A reference carries everything in it. Give each one job, name what to take, and fence the rest with "take nothing of its". | 10.1 What a reference does |
Master, anchor and in-look anchor have one meaning each. A drifting frame is re-anchored to the first-generation masters, never edited in place. | 10.2 Master, anchor and in-look anchor |
The identity clause is six to ten fixed facts, asymmetries first, written once from the approved master and pasted unchanged. The master wins any disagreement. | 10.4 The identity clause |
An anchored prompt runs manifest, identity clause, change contract, frame. | 10.5 The manifest and the anchored prompt |
The three-panel sheet holds one face. Attach a sheet only where the route reads it. | 10.6 The master's grammar, sheets and single views |
A real actor's name is a written project decision, off for commercial work. A real person as themselves needs a release and consented photographs, and no name. | 10.8 Real people, and an actor's name as a rights decision |
Crowds are clusters with a task, faces small by distance. A place is built from a geography master and plates, and a reverse is never a mirror. | 10.9 Crowds and extras; 10.10 Places and plates |
A product on sale is the authority on itself. Attach the real thing, and composite its marks and legal lines in post. | 10.11 Products: the real thing is the authority |
World text is specified and read glyph by glyph. Editorial text is composited from real assets. | 10.12 Text in the frame: two layers |
An edit changes one thing, names what it may move, and repeats the preserve list. | 10.13 The one-change edit |
Nano Banana Pro is named on every request. It has no mask and no seed, so the approved frame is the seed, and pixels that must not move are composited in post. | 11.1 What it is (Nano Banana Pro); 11.2 How it reads what you give it (Nano Banana Pro); 11.6 Settings (Nano Banana Pro) |
GPT Image 2.5 takes a transparent element, a text- or layout-led frame, a bounded edit on a frame with no signed face, and fast exploration. Signed faces stay on Nano Banana Pro. | 12.1 What it is (GPT Image 2.5) |
E.2.6 Motion
Chapters 13 to 20.
Rule | Taught in |
Write the invariant first, choose the operation that protects it, and only then write the prompt. | 13.1 What this stage decides, and who decides it; 13.2 The operations |
The route line has eight parts and goes into the shot table, the shot record and the run record. | 13.3 The route line |
Leave the start frame behind only when the frame is negotiable, and never mix a start frame with anchors in one shot. | 13.4 When to leave the start frame behind |
One shot, one clip. Several shots in one generation bake the edit in. | 13.5 One shot, one clip; several shots in one generation |
The frame owns appearance, the words own change, the settings own the rest. Protect by name what must not change, and compare the first frame of every clip with your keyframe. | 14.1 Who owns what |
Sign the start frame at the state the clip begins in, and leave room for the move. | 14.3 The start frame: state 0, and room for the move |
A finished clip prompt has seven elements in order, and its first sentence carries what the shot is about. | 14.4 The seven elements, in order |
Give every move a body and name one mechanism. Never write a hedge word for the camera: declare it locked and give the motion to the world. | 14.6 Rig physics: give the move a body; 14.4 The seven elements, in order |
Size a push by E = D ÷ (D − d), and measure the returned clip on a rigid feature. | 14.5 Measured moves: sizing one, and measuring it |
End every clip on a composition you can cut on, then a hold. | 14.7 The endpoint |
Pin both ends only when the landing is the point. Make the two frames relatives, describe the travel, and check the middle. | 14.8 Start and end frames |
A product gets one small move with a visible cause and a stop. | 14.10 Product moves: rotation, reveal and arrival |
Read the next shot's state from the last accepted frame, and carry the residue. | 14.14 Scale and continuity |
Shape, size and length are settings. The label is not the file: probe every clip and quote the probed raster. | 14.15 Format, length, resolution, handles and probing |
The same failure twice means you change the input, not the adjectives. | 14.16 Checking a moving shot, the ladder, and when to stop |
Believable performance is grounded. Translate every emotion word into physics, prefer the body to the face, and give each beat one visible event. | 15.2 What makes a performance believable; 15.3 From the scene's psychology to one visible event; 15.4 The task and the cues |
One speaking mouth per shot, and a listener with a job. Where the body faces and where the eyes go are separate decisions. | 15.5 Blocking: where bodies are and where eyes go; 15.6 Listeners, handovers and two-shots |
Set the sound switch explicitly on every clip, and write the sound the clip should carry. The edit owns the final sound. | 14.4 The seven elements, in order; 16.6 Settings (Kling); 17.3 The prompt, in this tool's dialect (Seedance 2.5); 25.1 What this stage decides, and who decides it |
Kling 3.0: a signed start frame, one action, one clip. Sound off for every clip whose sound you design and on for a talking plate; never Arabic dialogue from Kling. | 16.1 What it is (Kling); 16.3 The prompt, in this tool's dialect (Kling); 16.6 Settings (Kling) |
Seedance 2.5: on a reference job the words bind, place, lock and protect, and describe only what no file carries. An edit or an extension starts from a source of 8n+1 frames. | 17.2 How it reads what you give it (Seedance 2.5); 17.3 The prompt, in this tool's dialect (Seedance 2.5) |
Wan 3.0: a job carries boundary frames or references, never both. It is the economy route and is never used for Arabic. | 18.2 How it reads what you give it (Wan 3.0); 18.1 What it is (Wan 3.0) |
A tool becomes a route when its inputs, limits and terms have been read on the connection you will use, a scratch job passes the checks of its stage, and the director decides. | 20.1 How to read an entry |
E.2.7 Voice and speaking faces
Chapters 21 to 24.
Rule | Taught in |
The edit owns sound. Write each shot's sound contract before its prompt. | 21.1 What this stage decides, and who decides it; 21.2 The sound contract, decided first |
In a client final the voice is a person who signed, or a consenting clone of that person. A library voice is for casting and screening. | 21.6 Casting a voice, and who may speak in a final; 34.8 Voices: what may ship |
Keep two texts for every line: the canonical line, which never changes, and the engine text, which carries the bracket and any repair. | 21.4 Writing the line: two texts |
Cast the voice before you write a tag. Audition, score the words and the dialect first, and let a native ear choose. | 21.6 Casting a voice, and who may speak in a final |
Direct with one bracket of voice quality, attitude and address. Punctuation is the score. Take three a round and change one variable at a time. | 21.5 Directing a line |
Repair only the word that fails, by the ladder, and let the reviewer approve every mark. Whole-line vocalisation comes last. | 21.7 The Egyptian line workflow: repair, measure, lock |
Lock the take before any picture is paid for: WAV, 48 kHz, mono, about −20 LUFS with peaks at −3 dBTP, 200 ms of head and 400 ms of tail. | 21.7 The Egyptian line workflow: repair, measure, lock |
Saudi means Najdi or Hejazi. Cast from samples in that variety, with a reviewer from it, and never from a "Gulf" label. | 21.8 Saudi characters: Najdi and Hejazi |
Numbers, dates and brand names have a canonical field for the screen and a spoken field for the voice. | 21.9 Numbers, dates, brand names and English terms |
Keep the dialogue ledger for every line and the take ledger for every generation. | 21.11 Checks, the gate and the records |
A speaking shot has no default contract. The director decides at the start whether the exact take ships or a new performance is acceptable, and writes it as a contract line before any picture. | 22.1 What this stage decides, and who decides it; 22.2 Write the contract before any picture |
Design the sync out first: a fallback for every shot, a written ceiling for visible sync, and a break shot between visible moments. | 22.3 Design the sync out first |
Route A, the exact take: a locked take, a plate that talks quietly with no words, and a Sync Lipsync 3 pass with cut_off. Time the plate to the take. | 22.4 The keyframe and the talking plate; 22.5 Route A: the exact take; 24.6 Settings (Sync Lipsync 3) |
Route B, a new performance: a video model performs the line. Seedance 2.5 can speak English in the chosen voice; Egyptian and Saudi dialogue is a cast take given as the voice reference, judged word by word, and never Arabic from text. | 22.6 Route B: a new performance |
Captions follow the audio that shipped: the take on Route A, the clip's own sound on Route B. | 22.9 The records this stage writes, and the gates |
ElevenLabs: name the model on every take, audition v4 against v3 on each new Arabic voice, and never regenerate a locked line because a newer model exists. | 23.1 What it is (ElevenLabs); 23.3 The prompt, in this tool's dialect (ElevenLabs); 23.10 Version notes (ElevenLabs) |
Sync Lipsync 3 has no prompt. Set sync_mode to cut_off on every pass, and prepare the take with clean single-speaker speech, 48 kHz mono, 200 ms of head and 400 ms of tail. | 24.2 How it reads what you give it (Sync Lipsync 3); 24.6 Settings (Sync Lipsync 3) |
E.2.8 Music and sound
Chapters 25 to 28.
Rule | Taught in |
The edit owns all sound. A video model's own audio is scratch, and every sound in the film is a file with a row, a frame, an owner and a rights line. | 25.1 What this stage decides, and who decides it |
Write the sound spine first: motifs, five layers, and silence as a cue. | 25.2 The sound spine |
Brief a cue around a theme, with the refusals apart and no artist named. | 25.3 Whether music enters, and the composer brief |
Choose music-first or picture-first by what carries the story. Lock the cue before any picture depends on it, and place its first attack, not the file's start. | 25.4 Music-first or picture-first; 25.5 Choosing a cue and placing it |
Spot every event on the selected take: contact, continuation and release as separate rows, one owner each, hard sync points marked. | 25.6 Spotting: every event, its frame and its owner |
Describe an effect as a physical event heard from a place. Record it when the prop is at hand, license a library file when the sound is ordinary, and generate only what neither gives. | 25.7 An effect is a physical event heard from a place; 25.8 Record, library or generate |
Put the transient on the contact frame and one bed under every cut. Fix timing in the timeline, never by regenerating. | 25.10 Timing to the frame; 25.11 Beds, layers and room tone |
Loudness is measured over the whole mix against the delivery owner's specification (EBU R 128: −23 LUFS, −1 dBTP; ATSC A/85: −24 LKFS, −2 dBTP) and never set in a generator. | 25.14 Loudness to the delivery specification; 30.6 Sound in the finish, and loudness |
Hand over separable layers, raw and placed, with the cue sheets and the rights line. | 25.16 The handoff to the mix |
Suno: Custom mode, one directed paragraph, no artist, composer, film or brand name anywhere Suno reads, refusals in Exclude Styles, and the ending written as behaviour. | 26.2 How it reads what you give it (Suno); 26.3 The prompt, in this tool's dialect (Suno) |
Suno: work on a paid plan from the first generation, download the candidate you choose, and tell the client in writing that copyright is not warranted. | 26.9 Limits, prices and rights (Suno) |
Eleven Music: name the model in every request. A self-serve plan excludes film, TV and radio. | 27.3 The prompt, in this tool's dialect (Eleven Music); 27.9 Limits, prices and rights (Eleven Music) |
E.2.9 Edit, finish and delivery
Chapters 29 to 32.
Rule | Taught in |
The assistant logs, proposes and does the arithmetic. The director decides the cut; a timeline that loads, a check that passes or a contact sheet never passes a cut. | 29.1 What the edit decides, and who decides it |
Keep three numbers apart: the seconds asked, the frames in the file and the interval used. Log intervals, not files, with every defect at the frame where it starts. | 29.2 Before the first cut; 29.4 Logging a take by its usable interval |
Assembly A is in causal order on hard cuts, with the audio that ships at its own timing, one room bed, slugs for what is missing and a slate with its version. | 29.6 Assembly A: the first causal cut |
One proposition per alternate. Move the picture boundary, never the dialogue, and never mirror a frame. | 29.7 Alternates, and J and L cuts; 29.8 Continuity and contact joins |
Never speed up speech. Trim lead-ins, repeats and finished tails first, and put a transient on the frame of contact. | 29.9 Rhythm and sound-led timing |
Check the file, then the picture, then the sound. Only an eye and an ear pass a shot. | 29.10 Judging the work in the running cut |
Change a select before you ask for a replacement. Diagnose before you spend, climb one rung at a time, and lock only when nothing is open. | 29.11 Back to generation; 29.12 Diagnosing, the ladders, and when to stop; 29.14 Locking the picture |
Lock the picture first. Then conform, repair, upscale only if the delivery needs it, colour, and texture last. | 30.1 The finishing order |
Probe, then conform on purpose. Never stretch a raster to hide a mismatch, and never trust a label. | 30.2 Probe and conform |
An upscale is a comparison with the native version at the same grade, and the delivery size decides. | 30.3 Repair, then decide whether to upscale |
Generated media is not camera log. Interpret the source, match to the masters, and apply the look as a separate step. | 30.4 Colour |
Delivery grain and halation are added last, restrained, at final resolution, and never asked of a generator. | 30.5 Texture: grain and halation, last |
Captions come from approved copy, timed to the audio that shipped, checked in the real player, and never reversed by hand. | 30.7 Captions, and Arabic text |
Plan every aspect ratio upstream, and give every version its own manifest line and check. | 30.8 Aspect ratios and versions |
Build graphics that may change natively. A generated graphic is pixels, and exact copy is composited. | 30.9 Motion graphics: native or generated |
Deliver the native project, the collected media, the stems, the title sources and a reference render. Add an interchange file only for a destination you have tested. | 31.1 What a handover is for; 31.3 Delivery masters and packages; 31.4 The routes into Premiere |
Two slates and a version: "DRAFT — NOT FOR APPROVAL" and "ROUGH CUT vN — FOR EDIT DECISION: [the question]". Every export carries its version. | 31.2 The folder, the names and the slates |
Reopen, probe and hash every output. A package that fails is rejected, not patched. | 31.3 Delivery masters and packages; 31.5 Building and verifying a handoff |
E.2.10 Running the production
Chapters 33 and 34.
Rule | Taught in |
Use no model, tool or preset that the platform made itself, and no tool whose engine is not named. | 33.2 How it reads what you give it (Higgsfield) |
Log the platform's job label beside the model you asked for. | 33.5 Examples at the standard (Higgsfield); 33.7 Checks before you sign (Higgsfield) |
Plan from the dated table, spend from today's quote, settle from the ledger. | 34.1 Three figures for one job; 34.2 Checking the cost on the day |
Spend is a gate. Every paid batch has a signed packet with a cap, a retry allowance, stop conditions and an end date. An assistant never tops up or accepts a billing prompt. | 34.3 Spending inside an approval |
Everything you upload can train the platform's models, and only an Enterprise plan stops it. Upload what is public, yours, or released with the client's written agreement; everything else is a stand-in. | 34.5 Higgsfield's terms, for a client job; 34.6 Client material and stand-ins |
A living person appears only from consented photographs, under a release that covers AI use, with no name and no tolerance for drift. | 34.7 Real people, and the references you attach |
Every voice, cue and effect in a final has a rights line. A licence you have not read stays out. | 34.8 Voices: what may ship; 34.9 Music and sound effects; 34.10 The rights line, and what may ship |
Seven points go into every client contract. | 34.11 What to put in the client contract |
E.3 The author's decisions of 30 September 2026
Given by the author on 30 September 2026. They govern this edition of the book.
E.3.1 The speaking-shot contract is decided per project
There is no default. At the start of each project the director states whether the exact take must ship or a new performance is acceptable, and writes it as a contract line before any picture is made. Exact take: the locked take ships, a talking plate is made for it, and Sync Lipsync 3 fits the mouth to it. New performance: a video model performs the line; on an English-language project Seedance can generate the audio directly with the chosen voice. Egyptian and Saudi dialogue always comes from a voice actor or from ElevenLabs, under either contract. The book teaches both routes equally.
Taught in: 8.2 Decisions made at the start of every project; 22.1 What this stage decides, and who decides it; 22.2 Write the contract before any picture; 22.5 Route A: the exact take; 22.6 Route B: a new performance.
E.3.2 Rough-cut labels
Two labels and a version. "DRAFT — NOT FOR APPROVAL" for work in progress; "ROUGH CUT vN — FOR EDIT DECISION: [the question]" when a decision is wanted. Every export carries its version number.
Taught in: 31.2 The folder, the names and the slates; 29.6 Assembly A: the first causal cut.
E.3.3 Arabic in a frame is always exact
Arabic text generated inside a frame must match the approved string letter for letter, everywhere. A final ى where ي was asked fails, and is fixed in post or run again. Brand names, product names and legal lines are composited in post.
Taught in: 10.12 Text in the frame: two layers; 30.7 Captions, and Arabic text.
E.3.4 Latest versions only
Tool knowledge is always the latest version and carries the date it was checked; craft knowledge may cite any year when it teaches something. Two versions stay beside the newer one: Kling 3.0 stays complete beside Kling 4.0 as announced (28 Sep 2026, full release due in October), because 4.0 is not yet available to use; and ElevenLabs v3 details stay beside v4 (released 28 and 29 Sep 2026), because v4 is unproven on Egyptian and Saudi Arabic.
Taught in: 16.1 What it is (Kling); 16.10 Version notes (Kling); 23.1 What it is (ElevenLabs); 23.10 Version notes (ElevenLabs).
Next: Reference F · Glossary
Comments