Chapter 8 · Planning the production

Yasser Ashour

2 hours ago

29 min read

Part 1 · Before the camera

On a set, planning protects the shooting day. An AI production has no shooting day: it has hundreds of small jobs, each cheap, each starting from nothing, each made by a model that remembers none of the others. Planning is what turns them into one film. This chapter teaches the documents that do it.

In this chapter

Where a change is cheap and where it is expensive, and the four gates that stop a production
The four decisions made at the start of every project: the speaking-shot contract, the delivery specification, the rights class and the look
The obligations list, and the production plan built from the treatment
Three linked records: the shot table, the shot record and the run record
The continuity plan, and the order of work: masters, then anchors, then the shots
Names, folders and the workbook; the spending packet and the budget

Before you start. Chapters 3 to 7 produce what this chapter organises. Chapter 9 shows how the shot record, the prompt and the settings divide the work of a frame; Chapter 33 keeps the run record on the platform; Chapter 34 teaches how cost, terms and rights are checked on the day.

8.1 Where a change is cheap, and the four gates

A conventional commercial passes through many hands in a fixed order: strategy, concept, the client's approval, treatments, the pre-production meeting where agency, client and director agree the plan (the PPM), the shoot, and the edit, grade and sound. Every stage multiplies the cost of a change. A script change during strategy costs an afternoon; the same change in the grade can cost the budget.

An AI production changes the shape of this, not the principle. Ideas, pictures and refinements overlap, and the bottleneck moves from how much you can shoot to how fast you can decide. What the loops do not remove is the cost of a question left open. A frame made before the wardrobe is settled must be made again once it is, and so must every frame built from it. That is a planning failure, not a prompting one.

Figure 8.1 — The same work in two shapes. Generation loops freely between four kinds of gate; the decision log runs underneath, so a rejected idea is not proposed again later.

Work runs on until it meets a gate, and only four kinds of gate stop it. No consequential gate is silently skipped.

Gate	What stops the run	What you record
LOCK	a choice that materially changes the idea, the scene architecture, the look or the route	the real alternatives, a recommendation, and the choice, dated
SPEND	any paid or credit-consuming generation	the exact batch, the route, the price read on the day and the retry allowance (8.10)
RIGHTS	a real person, a cloned voice, a protected brand asset, licensed music or footage, or any other use that needs authority	the approval and its scope; without authority, that route stays closed (Chapter 34)
ACCEPT	a person's judgement of creative output, or of the final delivery	the verdict and who gave it; nothing accepts itself

Approval of a treatment, an asset or a plan is not approval to spend: a locked treatment is not permission to generate, and a signed frame is not permission to make forty clips.

Keep a decision log from the first day. Record every decision you lock (a wardrobe, a voice, a route, a runtime) with its reason and date, and every idea you reject, with the reason too. An assistant has no memory of yesterday's rejection, and neither has a crew member who joins in week two; without the log, a rejected idea comes back.

8.2 Decisions made at the start of every project

Four decisions shape every shot and cost the most to change late. The director makes them on the first day, writes them on one card and locks them at the first LOCK gate. Where the client's word is needed, the client agrees them in writing.

The speaking-shot contract

A shot in which a visible face says a line is made in two halves: a voice, and a picture of a mouth that fits it. The join decides whose audio the client hears, so the director states at the start which contract applies. There is no default; both are legitimate, and Chapter 22 teaches how each is built.

Exact take	New performance
What ships	the locked take, re-encoded	the audio the video model returns, once the director has judged it on the clip
The picture	a talking plate is made for the take, and Sync Lipsync 3 fits the mouth to it	a video model performs the line from the signed frame and the voice
The voice	a consenting actor's recording, or an approved ElevenLabs take	the same voice, given to the model as its reference; on an English project Seedance can generate the audio directly with the chosen voice
Choose it when	the exact reading is approved; the line recurs across versions; captions, mix and repairs must follow a waveform you own; dialect or legal accountability decide	a fresh rendition is acceptable; face and voice are wanted from one performance
It costs you	a plate and a lip-sync pass for each visible mouth	a second judgement of every word on the clip; captions follow the clip

Egyptian and Saudi dialogue always comes from a voice actor or from ElevenLabs, under either contract. A video model is never asked to invent a dialect from text; where it performs, it performs the voice you cast.

You may decide by shot class, for example the hero speaking shots on the exact take and background lines on a new performance, but write the classes down. Each shot's record carries the contract as one line: the contract, the take, what ships, which audio the captions follow, and any change with its reason. A shot never quietly changes contract, because that changes what the client hears. And design the exposure down first: a line spoken off-screen, in profile or from behind needs no mouth at all, and the cheapest lip-sync is the one you never run.

The delivery specification

The client's or broadcaster's delivery specification decides values you would otherwise guess, so obtain it in writing before the first frame. For each version it names the aspect ratio, length, frame rate and resolution; the codec and container; the audio layout and loudness; the captions and their languages; the safe zones; the file names. Two reach back into planning: the production is composed natively for the primary aspect ratio, and every other ratio is planned, not assumed (Chapter 30); and the delivered resolution, set against what the models return, decides whether any upscale is needed (Chapters 14 and 30).

Loudness is never set in a generator; it is measured on the finished mix against the named specification. Two families cover most short commercial work (checked 30 Sep 2026):

Specification	Where it applies	Loudness	Peak
EBU R 128, version 5 (Nov 2023), and its short-form supplement s1 (Aug 2020)	European broadcast; the definition includes advertisements, trailers and promos	−23.0 LUFS integrated; short-term not above −18.0 LUFS for short-form	not above −1 dBTP
ATSC A/85, edition of July 2026, Annex M	United States, short-form content	−24 LKFS, about ±2 dB tolerance	not above −2 dBTP, about ±0.5 dB of meter tolerance

For Egyptian and Gulf broadcasters, cinema advertising and the social platforms, ask the delivery owner for the specification in writing and record the answer with its date. ITU-R BS.1770 defines how loudness is measured; it is not a target.

The rights class

Classify what the job will contain before anything is uploaded, because a platform that has trained on an upload cannot un-train. The classes: neutral synthetic (invented characters, generic props); private or rights-cleared (your own photographs and locations); client-confidential (unreleased products, scripts); real person or biometric; voice; brand or trademark; licensed third party (music, stock, footage); public domain; prohibited or unverified. For each asset that enters the job, record the owner, the permitted purpose, the territory and term, any deletion or training restriction, and the evidence.

Three defaults follow. The real pack, logo and legal lines are composited in post and never generated. A living person appears only from consented photographs, under a release that covers AI use; a name never anchors a face. The voice in a final is a person who signed: a consenting native actor's recording, or a clone of that same person under a written release; a library voice is for screening. What the platform lets you upload, and what it may do with it, is decided by its terms as read on the day (Chapter 34).

The look

Chapter 6 teaches the craft of the look; here the director locks it on one page before the first frame, because every still inherits it and every clip protects it. The page holds the light world or worlds, with each source and the contrast ratio; the palette as proportions of named surfaces, with the brand colour protected; the finish as capture behaviour; the delivery texture (grain, halation), decided once and laid in post, never prompted; and each reference, with the one dimension borrowed. A look changed after frames exist means remaking them.

Template: the project decisions card (written for this book; not run)

Project decisions · [project] · [date] · stated by [director]; agreed with [client]
Speaking shots: exact take / new performance / by class [say which] · voices: [language and variety; actor or ElevenLabs] · captions follow: [take / clip]
Delivery: versions [aspect · length · frame rate · resolution] · loudness: [named specification, in the delivery owner's words] · captions [languages] · safe zones [ ] · files [codec and container, as named]
Rights class: [what the job contains: neutral synthetic / client-supplied / released talent / real person / voice / brand / licensed music or footage] · uploads allowed: [ ] · real-name likeness: [off / on, for whom, released] · voice source: [ ]
Look: [light worlds and sources · contrast ratio · palette proportions and protected brand colour · finish as capture behaviour · delivery texture, added in post · references, each with what it gives and what it does not]

8.3 Obligations before footage

The first planning document is not a shot list. It is a short list of what the audience must know by the end of the film, in the order it becomes meaningful. These are obligations: the events the cut owes the viewer, whatever else it does. Write them before any footage is chosen, because every later choice is judged against them.

Start from the event that changes the audience's understanding, never from the prettiest shots. For a dialogue scene the obligations may be the question, the listener's recognition, the answer and its consequence. For a product film they may be material, use, result and identity: what it is made of, how it is used, what it does, whose it is. The inputs are the accepted treatment and the canonical script (the locked text of every line; Chapter 21); the runtime, aspect, exact copy and brand time (how long the brand must hold on screen).

The director can work this through with an AI assistant, treatment and script attached, in three steps. The assistant lists the information that must survive, and the director approves the list. It checks the list against the deliverable and names any conflict with a real alternative: drop a beat that is not essential, obtain the missing coverage, or change the runtime; the director decides. Once footage is logged, it proposes two structures that differ in when the audience learns the key fact, and the director chooses one for the first assembly (Chapter 29). The last two steps are LOCK gates. Three rules keep the list honest. Never cut the argument silently to hit a length: if the obligations do not fit thirty seconds, say so and offer a real alternative; the runtime is the director's decision. Ask for alternatives that differ in meaning, never "fast versus slow"; two cuts that differ only in pace are one cut. Protect the approved dialogue: no structure may change a word of it. Make the runtime an obligation too.

Template: the obligations brief (written for this book; not run)

Read [the treatment] and [the script]. List the audience information that must survive the cut, in the order it becomes meaningful: [the reveal], [the reaction that proves it], [the product evidence], [the brand obligation]. The film is [runtime] at [aspect]; the exact copy is [lines]; brand time is [seconds]. If the footage cannot meet all of these, name the conflict and one real alternative. Then propose two structures from the logged footage that differ in when the audience learns [the key fact], and say what each changes emotionally. Keep the approved dialogue exactly. Do not choose a structure by average shot length or by music beats alone.

Check what comes back. Every obligation must be an event a first-time viewer could name. If the assistant returns a scene summary, ask "what must the audience know by the end of each beat?"; if it resolved a conflict by itself, hand the conflict back as a decision. The numbered obligations go at the top of the edit plan, and every shot is judged against them: a shot that serves no obligation needs a reason to exist.

8.4 From the treatment to one production plan

On a conventional job the PPM turns the treatment into decisions every department has agreed, and a commercial fails when they clash: a wardrobe that fights the set's palette, a location that cannot take the planned move. The departments still exist on an AI production; their decisions go into the anchors, plates and sentences that every prompt carries, and a model renders whatever you fail to decide.

Department	Decisions to lock	Where they go
Production design	rooms, layout, props and their wear	location plates and set descriptions (Chapters 7, 10)
Wardrobe and styling	garments, fabrics, era, colours against the set	the identity clause and every prompt that names clothing (Chapter 10)
Light	sources, time of day, contrast, colour	the look and the light sentences of every frame (Chapter 6)
Casting	who each character is; any real person	masters, releases (Chapters 10, 34)
Camera	shot sizes, positions, moves, the aspect	the shot table and the prompts (Chapters 5, 14)
Sound	who owns each sound	the shot table and the contract line (Chapter 21)
Rights	every real person, voice, brand asset, track and clip, with its scope	the RIGHTS gate (Chapter 34)
What stays real	logos, packshots, legal lines, titles	composited in post, never generated (Chapters 10, 30)

Close pre-production with a PPM summary: one page recording the agreed decisions, the trade-offs reviewed on the way, the project decisions card, the route and track each kind of shot will take (Chapter 13), and what will be real rather than generated. Approve it, date it, and put every locked decision into the decision log.

Decide once, and carry the decision into every prompt word for word. A wardrobe line, a room description or an identity clause written once in the plan and pasted unchanged into each prompt is a continuity tool. The same thing retyped from memory each time is a source of drift.

8.5 The shot table

Every shot is described three times, by three linked records, because the three answer different questions. The shot table is the plan: one row for every shot you intend to make, short enough to read on one screen. The shot record is the director's page for one shot (8.6). The run record is what happened: every attempt, kept the day it runs (8.7). The three share one key, the shot's ID. The table points to the shot record, the shot record to the run record; nothing is stored twice.

Figure 8.2 — Three linked records. The table is the index, the shot record holds the prompt and its fields, the run record holds the truth about each attempt. The shot record is never pasted into the prompt; only the positions its mid-action field names are written there too.

Write the table before the first prompt. A conventional shot list gives the shot code, subject and action, size and framing, camera and lens, sound, duration, notes and priority. The columns a live-action crew never needed are these:

Column	What goes in it
Obligation	the number from 8.3 the shot serves; a shot with none needs a reason
Track	keyframe-first or direct-to-motion (below)
State in, state out	what is true in the first frame and the last: the clip must start where the frame is and land where the next shot expects
Route line and fallback	the eight parts of Chapter 13 (invariant, operation, grade, model and mode, inputs in order, settings, reason, fallback), decided before the prompt; the fallback is also a column of its own, so the retry ladder ends in a named alternative
References	the anchors, plates and product photographs the shot needs (8.8)
Sound owner	a silent plate, the model's own sound, a locked take, or the edit (Chapter 21); for a visible mouth, the contract of 8.2
Interval and handles	the stretch the cut needs, plus a little either side (Chapter 29)
Rights status; status	cleared, needs approval, or not needed, and where the release is filed; where the shot stands (8.9)

Mark each shot's track, because it decides whether a keyframe exists and where identity is checked. Keyframe-first: the start frame is made from the anchors and approved before any motion is bought; three steps (anchors, keyframe, motion); the motion call is one lean call (the frame, the motion, the protection clauses); identity is checked in the still, before the motion spend. Direct-to-motion: no artificial keyframe for a shot made by text or by references alone; two steps; one call carrying the anchors, the description of the frame and the motion; identity is checked in the returned clip, after the spend.

A project adds its own columns without calling them universal; Mastorna's table added an evidence class and a source basis for shots reconstructed from a book, which belong to a reconstruction and not to a commercial.

Durations are planned in seconds and delivered in frames. Clips come back in fixed lengths that rarely match the request to the frame. Plan the interval the cut needs inside the length that comes back, and probe every file (Chapters 14 and 29). Mastorna's shot 6.1 is planned at four seconds on screen while its motion request asks for six; the cut takes its four from inside the six.

Written for this book · food and drink · the three records of one thirty-second film, a bottled hibiscus drink in a Cairo summer · not run

Example 8.1 — one film through the three records

The obligations, two rows of the shot table, and the shot record of the hero shot:

1. The afternoon heat is felt before the drink is seen.
2. What the drink is: ruby hibiscus, poured over ice.
3. The first sip lands: the shoulders drop, the eyes close.
4. The pack is recognisable and the claim line legible (composited in post).
5. The brand holds for two seconds at the end. Runtime 30 seconds, 16:9, with a 9:16 cutdown.

ID · serves · duration	The shot	Track and route line	State in → state out	Sound owner · rights · status
S03 · 2 · 5 s	the hero pour: ruby liquid into ice, one hand, late sun	keyframe-first; Kling 3.0, start frame, 5 s, sound off; fallback: two shorter clips cut on the tilt	cap off, glass empty → glass a third full	the edit (pour Foley) · stand-in bottle, real pack in post · keyframe approved
S05 · 4, 5 · 3 s	the pack shot with the claim line	not generated: the client's photograph, composited	end card, nothing in motion	the edit (music tail) · client-supplied, agreed in writing · waiting for the photograph

S03 · scene 1, the rooftop · the hero pour · track: keyframe-first
INTENTION: story intent: the drink is cold, ruby and real · emotional beat: anticipation before relief · goal of the shot: the audience wants the first sip
PROMPT: [the exact paragraph, pasted from its file]
PARAMS: still: Nano Banana Pro, 16:9, references in order: the stand-in bottle plate, the rooftop anchor, the glass photograph · clip route line: invariant: the signed frame's ruby liquid and cap-off bottle · operation: start frame · grade: finished · model and mode: Kling 3.0, start frame · inputs in order: start_image: S03_kf_v02 · settings: 5 s, 16:9, sound off · reason: one moving liquid system · fallback: two shorter clips
REFERENCE STRATEGY: the plate gives the bottle's shape, the anchor the place and light, the photograph the real glass; the signed frame becomes the clip's start frame
CONTINUITY ANCHORS: cap off; ruby, not orange; condensation beads on the bottle's shoulder; late sun from camera left; ice to a third of the glass
SOURCE-LOCK CHECK: [x] the stand-in shape [x] ruby liquid [x] ice visible [x] no label text
ANTI-AI WATCH: liquid hanging like a solid rod; ice cubes fused or floating; invented label text; bubbles like beer foam
MID-ACTION STATUS: the pour a third of the way in: a ribbon of liquid mid-fall between the lip and the glass, two ice cubes mid-turn, the first bubbles at the surface, three drops mid-run down the bottle's shoulder. Also written in the prompt, as positions.
REVISION LEVERS: (1) same frame, a thinner ribbon touching the ice near the glass wall (2) hold the bottle, three clear cubes (3) same frame, the ruby deepened one step where the sun passes through the liquid
DECLARED BAND: about 450 to 550 words: a hero product still with one hand, one liquid system, two light sources and a stated ratio
Also kept here: obligation 2 · palette plan: warm sand stucco 60 %, sky haze 30 %, ruby 10 % · the motion it owes: the tilt continues from the frame

Why it works

The row says what the shot is for and how it is made; the record says what the model is told and what to watch for.
Every frozen system in the mid-action field is a position the prompt can carry; none is a mood.
The levers are written before the first attempt, so an 80 % frame gets an edit, not a fresh roll.

8.6 The shot record

Behind every row of the shot table sits a shot record: the director's working page for that shot. It holds everything that shapes the shot but must not be pasted into a prompt: the intention, the planning reasons, the palette plan, the character's inner life. The prompt carries only the visible result. Written before the prompt, the record also holds the decisions that stop a second attempt being a hopeful re-roll: the start state, the checks, and the levers you will pull if the first attempt fails.

Template: the shot record (written for this book; not run)

[SHOT ID] · [scene] · [what it is, in a few words] · track: [keyframe-first / direct-to-motion]
INTENTION: story intent · emotional beat · goal of the shot (never pasted into a prompt)
PROMPT: [the exact paragraph, pasted from its file]
PARAMS: aspect [the project master] · resolution [ ] · model and mode [ ] · references attached, by name and in order [ ] · for a shot that moves, the route line [Chapter 13: invariant · operation · grade · model and mode · inputs in order · settings · reason · fallback]
REFERENCE STRATEGY: [which anchors or sheets this call uses, and what the frame becomes an anchor for]
CONTINUITY ANCHORS: [three to five locked facts this frame must reproduce]
SOURCE-LOCK CHECK: [the must-show items, ticked]
ANTI-AI WATCH: [two to four specific risks for this frame]
MID-ACTION STATUS: [the frozen systems and how far into the action each is; or STATIC, with the reason. Also written into the prompt, as positions]
REVISION LEVERS: [two or three edits planned before the first attempt]
DECLARED BAND: [about N words, and why]
Also kept here, never pasted: obligation served · start state and the motion it owes · character core and performance note (Chapter 15) · planning numbers with their reasons · palette plan and declared exceptions (Chapter 6) · sound owner (Chapter 21)

Two fields reach the prompt in their own right. Mid-action status is written into it as positions, and the intention reaches it as its physical translation ("eyes down on the tea, mouth closed and relaxed", not "quiet awe"). Parameters and reference strategy reach it only through the sentences that open an anchored prompt (Chapter 10); a risk you expect may become one of at most five exclusions (Chapter 9). The one trace of intent inside a prompt is a physically phrased production context in the opening declaration ("a hero frame for a premium athletic-wear commercial").

Planning numbers are your working language: focal length, stop, colour temperature, contrast ratio, distances. At finished grade they go into the prompt with the visible result beside each number (Chapter 9); the record keeps the reasons, such as why 3:1 and why 3200 K over 5600 K. On Kling the prompt carries one approximate magnitude for each moving system, with a duration and a speed profile (Chapter 14).

The palette plan is a split such as 60/30/10: about 60 % of the frame in a dominant colour, 30 % in a secondary, 10 % in an accent. Plan it here as proportions; write it in the prompt as colours on named surfaces, never as the string "60/30/10" (Chapter 6). The character core is what stays true of a character across the film; only this scene's want, obstacle and trigger go into a prompt (Chapter 15).

Mid-action status is the one field the prompt must repeat, because the model never reads the record. A phase promised in the record and missing from the prompt is a phase the model will not draw. Write the frozen systems as positions ("rear foot just released, grit hanging in the air"), for a still and for the keyframe of a clip (Chapter 9).

Revision levers. An image that is 80 % right gets a named lever, not a regeneration: a different reference, an end frame, a shorter move, an edit instead of a fresh frame (Chapter 29 teaches when to stop). The declared band is the length the prompt aims for and why. There is no floor and no ceiling: a prompt is finished when every element the shot needs is settled. A complex hero still lands around 400 to 600 words, a simple insert 150 to 300, an edit 40 to 150.

Read the record against the prompt before you submit. A filled field is not proof that it is true. The commonest contradiction is a mid-action status that names a wave, a breath or a step the prompt never mentions; the second is a motion clause hiding in a still ("as it begins to pull slowly away"). Bring the prompt into line with the record, or the record into line with what you meant.

8.7 The run record

The run record is the log of every attempt: the request as sent, the settings, the inputs, the job number, the charge, the file, its hash and the verdict. Keep it the day each attempt runs, failures included, and never select first and reconstruct the batch later. A record of the winners is not a record of the production: it cannot say what a shot cost, which failure recurs, or whether a route has earned the right to be a default.

Template: the run record line (written for this book; not run)

[shot]_a[attempt] · asked [id · settings] · recorded [id · settings, as the record shows them] · billed "[ledger label]" · host and interface ·
inputs: [role → what it must not give → local file · checksum · upload or job id] · parent: [start frame or earlier candidate] ·
prompt v[n] (exact text kept) · job [id] · submitted [time] · quote [x] · charged [from the ledger] · elapsed [request to file] ·
file [name · checksum] · probe [w×h · fps · frames · seconds · audio] · accepted interval [in, out) at [fps] ·
defect class [none, or a class from the shot record's watch list] · disposition [selected | rejected | held] · why [one line] ·
verdict [accepted / rejected: what failed / pending] · tech read [anyone else's technical observation; never a verdict] ·
origin [generated | upscaled | painted over | composited | practical | client-supplied | stock]

Asked, recorded, billed are three labels that can differ, because the name on the platform is not proof of what ran (Chapter 33). Each input carries its role, what it must not give (a face only, not the room), its file and checksum, and the upload or job id. The parent is the frame the job started from. The verdict is the director's: accepted, rejected with what failed, or pending. A tech read is anyone else's technical observation, kept in its own field and never promoted to a verdict.

Three things make it the third of the linked records. The accepted interval: the source's in and out points, the first frame included and the last excluded, at the file's own frame rate, so that a cut can be rebuilt from it. The origin label of every asset: generated, upscaled (with the tool and the date), painted over, composited, practical, client-supplied or stock; editing does not launder an origin. The day it runs: a record started after the fact is reconstructed, not recorded.

Decide now where the record lives: a sheet of the workbook (8.9) or a file beside each output, but only one of the two, so that there is a single authority. Create it before the first job runs. Chapter 33 teaches how the platform's job history, ledger and probes fill it.

8.8 The continuity plan, and the order of work

A continuity kit is the set of pictures that keep a film consistent: masters, character sheets, state references, location plates, product photographs. It is tempting to build it first and in full, a six-view sheet for every character and a plate of every room. That wastes credits on pictures no shot needs. Build the kit from the shot table.

Start from the shot. For each shot, write its first visible instant, the move, and what the move will reveal. Then build only what is missing. A frontal reaction may need only the portrait; a head turn reveals a profile, so it needs one; a coat coming off needs both states; a product turned to camera needs its hidden side, or a decision that the hidden side may be invented.

Name each picture's job before you use it: identity (who this is), state (which version exists now: wet, torn, in the coat), start frame (what is true when the clip begins) or edit source (whose pixels are authoritative in a repair). Choose the authority for each element and approve every new view against it. For an invented character it is the approved master; for a consenting actor, the approved photographs, and a generated view of them is only a proposal; for a product you sell, its photographs, dimensions and artwork. Inspect each new view at delivery size, for recognition, and at full size, for eyes, teeth, closures, labels and contact. A copy of a copy hides drift; checking against the first-generation master does not.

Word	In this book
Anchor	any approved picture a later job is built on: a master, an in-look anchor, a location plate
Master	the first, neutral picture of a character, a place or a product, made directly for its job and never derived from another output; the authority on identity (for a place, on its geography). Every later job references it, never a frame made from it
In-look anchor	the character approved in the film's own light, made from the master with the master attached
Environment anchor	an anchor that fixes a place: the geography master (a wide that fixes the landmarks), coverage plates, light-state plates, material close-ups
State window	the span of the story in which a look state (wet, torn, in the coat) is true; it changes only at a stated beat, and each state has its own master

The order of work

First the neutral masters of every recurring person, so that they carry no look to leak: the master face, the three-quarter and profile chained from it, the body front and back, and a separate master for every state the story needs. Then each in-look anchor, and the environment anchors. Then the shots, in story order, the first hero keyframe joining the masters and never replacing them, and a stress set (smoke, hard light, the face small in the frame) made before the shot table locks. Chapter 10 teaches how each is made.

Figure 8.3 — The order of work. Copies of copies drift, so every call goes back to a first-generation master.

Every later call re-attaches the first-generation masters, and a chain of derived pictures is never used as a master. A drifting frame is re-anchored, not edited in place: attach the master again and restate the identity clause; if the frame keeps degrading, remake it from the masters. One-change edits of an approved frame stay legal (Chapter 10). Every few scenes, make a clean frame from the masters and compare it with the current output, because silent drift is caught by comparison, not by looking at neighbours. A film with a single look may take a shortcut: make the in-look anchor first and let it stand in for the master, fenced ("take nothing of its light or grade") whenever it is reused, as the Mastorna trailer did. On a commercial with changing light, the neutral master first is the safer course.

The reference manifest

The continuity plan becomes one sheet of the workbook: the reference manifest, the single list every later prompt and every check reads. It has a row for every recurring picture.

Field	What it holds
Identifier; version and status	a stable production name, never a nickname; draft, approved, superseded or prohibited (a superseded picture never enters a new job)
Role; subject label	the picture's one job; the exact words for the person or thing on every route, so that a name or a pronoun does not drift
Route mapping	which upload position, ordinal or Element, on which model; a token that belongs to one model is never copied to another
Source and consent; validity window	who or what made it, and on whose consent; the scenes, wardrobe, damage and light states for which it is true
Locked wording; approved by	the approved identity clause and wardrobe line, pasted and never retyped; who approved it, against which master

The upload list is written from the manifest, and the prompt follows the list ("Image1=CHAR-A-S1; Image2=CHAR-A-S4-WET; Image3=LOC-POOL-L2; Video1=CAM-DIVE-REF"). If the upload order changes, the manifest changes before the prompt does. The locked wording is the identity clause: six to ten fixed facts, the asymmetries first (the heavier left eyelid, the nose's slight deviation), the name, the current wardrobe state and one detail anchor, written once from the approved master. When the words and the picture disagree, the master wins and the words are corrected.

What never changes inside a scene: the face facts, the body proportions, the hairline, the wardrobe within its window, prop geometry and wear, the location's landmarks. What changes at a stated beat: a look state, a light state, prop damage. Write each change into the manifest with the beat where it takes effect, build it as its own state master, and use it in every prompt from that beat on. A deliberate change is a directing note, not drift ("the bowl changes position between these cuts as the second clue; all other counter objects keep their places"). Keep an object ledger, the running record of what has happened to each prop, so that damage carries forward. And when something upstream changes, mark what depends on it: if scratch voice lines are replaced by locked takes, everything timed to the scratch takes is stale.

8.9 Names, folders and the workbook

Fix stable, plain names before ingest. Give every file its final name before it is imported into the edit (ingest), and after ingest move or rename it only with a relink plan (relinking is pointing the edit back at files that have moved). Use plain letters, digits, hyphens and underscores only: no spaces, no accented letters, no Arabic script in a file name (the Arabic text lives inside the records). A workable pattern joins project, shot, what it is and version: HIB30_S03_kf_v02.png is the second keyframe of shot S03; HIB30_S03_clip_v01.mp4 its first clip; HIB30_A02_master_v01.png the second anchor; HIB30_L01_take_LOCK_v01.wav the locked take of the first line. The version rises only for a real change of content.

Index the folders with a file manifest: a list of every file with its job and state, so that anyone can find the approved version of a shot without opening every folder. Your folders, names and manifest are the ones that count, whatever filing the platform offers.

Track each shot through five states: generated (a file exists), curated (chosen from its batch), approved (the director signed it), placed in the cut (its interval is in the sequence) and accepted (it works in the film beside its neighbours). A file count is not completion: forty clips in a folder can mean eight shots done or none.

The workbook. Everything above lives in one spreadsheet, a sheet for each kind of information: the shot table, the shot records, the decision log, the reference manifest, rights (the approval, its scope and date for each real person, voice, brand asset, track or clip), the dialogue and take ledgers (Chapter 21), music cues and sound effects with their rights lines (Chapter 25), quality checks, the run record, and spend (8.10). Three rules keep it useful. Link by codes, and keep every long text once: a shot row refers to a decision or an anchor by its code. Never shorten a critical detail to fit a cell: the full prompt, the reason for a decision, the licence terms. Record every shot, prompt and asset: a shot that is not in the workbook is a shot nobody is tracking.

The check sheet

Write the quality-check sheet at planning time and fill it in shot by shot. It is the last check before a shot's assets lock: four blocks, and any failure stops the lock.

Block	What it asks
Prompt truth	are the elements complete and in order; do camera and light carry their results; are the exclusions few and earned
Identity truth	does the call carry the locked wording or the approved frame; are the masters present and named identically; is the result judged against the masters, never the previous output
World truth	is the geography consistent with the plates; is every noun legal for the period; is the world's text honest to its distance; are the must-show items ticked
Delivery truth	is the mid-action status true and the endpoint a cuttable composition; are the watch items cleared frame by frame; are parameters and spend logged; was a hero keyframe approved before its motion was made

The last check is the one no machine runs, and it outranks the rest: the shot answers its intention line. The returned file gets its own checks after the job (Chapter 29); the director's eye and ear pass the shot.

8.10 The spending packet and the budget

Spend is the second gate. One record opens every paid batch: the spending packet, a page short enough to read in a minute and sign. Draft it from the shot table before the first job; Chapter 34 carries its approval line and the price check on the day.

Research exemplar · cost discipline · the rule for every paid batch · written in the July 2026 research, never run

Example 8.2 — the packet that opens every paid batch

Before a paid or credit-consuming batch record: purpose; exact route; rights class; conditions; outputs per condition; retry allowance; current live debit; maximum spend; expected evidence; stop conditions; and approver. Approval of research, architecture, assets, or local authoring is not approval to spend.

The eleven fields are all recorded before the spend; an empty field is a decision not yet made, and the batch waits.
The last sentence keeps every other approval from standing in for this one: an assistant told "go ahead with the plan" has been told nothing about credits.

Template: the spending packet (written for this book; not run; the eleven fields of the packet, with an end date added)

Spending packet · [project] · batch [name]
Purpose [the shots, and what the batch must show] · Exact route [model id, mode, host, resolution, length]
Rights class [what every input is; the approval and its scope]
Conditions [what differs between calls] · Outputs per condition [n] · Retry allowance [n, and which ladder steps it spends]
Current live debit [the price read or quoted on the day, with the date] · Maximum spend [the cap] · Ends [date, no later than the next reset]
Expected evidence [what the result must show to count: probe, contact sheet, the director's eye and ear] · Stop conditions [events that end the batch]
Approver [name] · approved [date, scope, in the director's words]

Stop rules are conditions, not counts. A number of attempts is only a ceiling on spending. A stop condition is an event: the same defect comes back after the third step of the retry ladder (then the keyframe is the problem, not the clip); the price on the day is higher than the estimate; the route needs a right that has not been granted. The ladder is in Chapter 29.

Price the shot at its true cost. A generation's own credits are the smallest part of a shot that needs repair. The true cost adds the retries, upload and storage, the operator's time, review, repair, the editorial, effects, audio, text, grade and versioning the shot will need, and the cost of a shot that fails. A cheaper generation that needs more repairs can be the more expensive route. Estimate the terms you can measure, and record the rest (review time, repair time) in the run record, so that the next estimate has data.

Build the estimate from the shot table, row by row: the route, its planning price, and the attempts you will allow. Two attempts per shot is a working figure; a shot that fails twice for the same reason has an upstream cause, and a third attempt without a change is spending, not directing. Add the stills, the lip-sync passes for every visible mouth on the exact-take contract, and any test you need before a client job. Then read the plan's reset date, because unused credits may not carry over and a budget that spans a reset is planned as two. Each tool section's N.9 gives that model's planning price, dated, and Chapter 34 teaches how to check the cost on the day; the quote on the day decides.

Template: the budget estimate (written for this book; not run)

Shot [code] · route [model, mode, resolution, length] · planning price [credits, dated] × attempts allowed [n] = [ ]
Stills: [number] frames × [price, dated] × attempts allowed [n] = [ ]
Lip-sync passes: [number] × [last charged price, dated] = [ ]
Tests before the job: [what each settles] = [ ]
Repair, edit and finishing time: [credits or hours] = [ ]
Total: [ ] · credits available before the next reset: [ ] · approval to ask for: [cap, scope, end date]

Before the first paid job

The project decisions card and the decision log exist; the treatment and the plan are locked and dated.
The obligations list is approved; every shot serves one, or says why not.
Every shot has a row and a shot record with its mid-action status, declared band and levers.
The continuity kit is planned from the shots; each picture has one job and a manifest row.
Rights are cleared for every real person, voice, brand asset, track and clip the batch uses.
Names, folders, the file manifest and the run record's home are set up.
The spending packet is written, with a cap, stop conditions and the price read on the day.

Open questions

Neutral master or in-look anchor for a film whose look may change. No maker and no comparison settles it. The course here is the neutral master first, then the in-look anchor made from it.
Loudness for Egyptian and Gulf broadcasters, cinema advertising and the social platforms. No publisher's figure has been read; ask the delivery owner in writing (8.2).

What to remember

Generating is cheap; an undecided question is not. Every frame made before a decision must be made again after it.
Four gates stop a production: LOCK, SPEND, RIGHTS and ACCEPT. Keep a decision log, including what you rejected.
On the first day the director decides four things: the speaking-shot contract (exact take or new performance, no default), the delivery specification, the rights class and the look.
Write the obligations first, then decide each department's choices once and carry them into every prompt word for word.
Keep three linked records: the shot table (the plan, with each shot's track), the shot record (never pasted whole; mid-action status is the one field the prompt repeats) and the run record (every attempt, kept the day it runs).
Make the pictures in order: neutral masters, in-look and environment anchors, then the shots, re-attaching the first-generation masters every time.
Plain, stable names; one workbook; five tracker states. The shot must answer its intention line.
Open every paid batch with a spending packet; write stop rules as conditions; price the shot at its true cost.