The production at a glance
Before you begin
One picture and one table. The picture shows the stages of an AI commercial, the four gates that stop the work and the three records that keep it honest. The table says which tool does which job and which chapter teaches it.
The production in one picture
Figure 0.1 — The production in one picture. The stages run left to right and each ends in a yes from the director; the four gates can stop any of them; the three records share one key, the shot's ID.
Read the top row left to right. Each stage ends in the director's yes: a locked plan, a signed frame, a clip accepted in the running cut, a locked take, a judged cue, a locked cut. That yes becomes the authority every later step is checked against. The operator writes, prices, runs and records the jobs; only the director signs (Chapter 1).
Work runs on until it meets a gate, and only four kinds of gate stop it (Chapter 8).
Lock. A choice that changes the idea, the scene architecture, the look or the route. Record the alternatives and the choice, dated.
Spend. Any paid generation. Record the exact batch, the quote on the day and the retry allowance. Approval of a plan is not approval to spend.
Rights. A real person, a voice, a brand, licensed music or client footage. Without authority, that route stays closed (Chapter 34).
Accept. A person judges the output, or the delivery. Nothing accepts itself.
Three records keep a production honest. The shot table is the plan: one row for every shot, written before the first prompt. The shot record is the page for one shot: the exact prompt and the fields around it, never pasted into the prompt. The run record is what happened: every attempt, kept the day the job runs. The table points to the shot record and the shot record to the run record; nothing is stored twice. The decision log beneath them holds every locked decision and every rejected idea with its reason, so that a rejected idea is not proposed again.
Four habits hold at every stage. Sign before you build: a frame is signed before it moves, a take is locked before a face speaks it. Name the model in every request. Change no route silently. Plan from the card, spend from the quote, settle from the ledger.
Which tool for which job
Start with the tool in the second column; use the third only for the case the row names. Versions are those current on 30 Sep 2026. The first section of each tool chapter (N.1) gives its current version, and says what is announced. Bold chapter numbers teach a tool (Chapters 20 and 28 gather several smaller ones); the number in brackets is the stage chapter that teaches the craft.
Job | Start with | Fall back to, or add | Chapter |
An anchor still, keyframe, plate or character sheet of a signed identity | Nano Banana Pro | GPT Image 2.5 for a bounded edit on a frame with no signed face, a transparent element, a text- or layout-led frame, or fast exploration | 11, 12 (9, 10) |
A product with its real label | Nano Banana Pro, the real product photograph attached: the real thing is the authority | the real artwork set in post | 11 (10, 30) |
A transparent cut-out; a typographic or layout-led frame | GPT Image 2.5 | the real artwork set in post | 12 (30) |
Logos, titles, supers, legal lines, end cards | post, from the real assets, always | none | 30, 32 |
A measured camera move from a signed keyframe | Kling 3.0 (Kling 4.0 announced 28 Sep 2026, full release due in October) | Seedance 2.5 | 16, 17 (14) |
New compositions of locked people, products and places | Seedance 2.5, references to video | a keyframe for each shot, then a start frame on Kling | 17, 16 (13) |
A shot that must land on a second frame | start and end frames on Kling 3.0 or Seedance 2.5 | none | 16, 17 (14) |
Changing or lengthening accepted footage | Seedance 2.5, edit or extend | regenerate the shot | 17 (13) |
A third video model: text, first and last frame, or references to video | Wan 3.0 | none | 18 |
A model still to be tested before it earns a role | Veo 3.1, MiniMax H3; Runway, Luma, LTX and newer platform models | none | 19, 20 |
The voice in a client final | a consenting native actor, recorded | a clone of that same person under a written release; ElevenLabs (Eleven v4, with Eleven v3 beside it) for screening and casting | 23 (21) |
Egyptian or Saudi dialogue | a voice actor or ElevenLabs, always; never a video model's own speech | none | 23 (21) |
A native performance in a character's timbre | ElevenLabs Voice Changer | the performer's own recording | 23 |
A speaking shot where the exact take must ship | a talking plate (Kling 3.0 or Seedance 2.5), then Sync Lipsync 3 with the locked take | none | 24, 16, 17 (22) |
A speaking shot where a new performance is acceptable | a video model performs the line: Seedance 2.5 (on an English-language project it can generate the audio directly with the chosen voice) | Wan 3.0 | 17, 18 (22) |
A music cue | Suno, briefed like a composer | a composer or licensed library when ownership or exact timing is needed; Eleven Music where its licence permits the use; Lyria | 26, 27, 28 (25) |
Effects, Foley, ambience | recorded, or a licensed library, first | a generator for single events and short beds: ElevenLabs sound effects, Firefly, Stable Audio 3 | 28, 23 (25) |
The cut, captions, conform and handover | Premiere Pro | none | 32 (29, 30, 31) |
Upscale or frame-rate boost | after the cut is locked, and only if delivery needs it: the upscalers Chapter 32 covers | deliver at native size and say so | 32 (30) |
Submitting jobs, recovering them, keeping the ledger | Higgsfield | none | 33 |
Read the table with the project's decisions beside it. The speaking-shot contract chooses between the two speaking-shot rows; the rights class decides the voice rows; the delivery specification decides whether an upscale or a version is needed at all (Chapter 30).
Where to go next
New to AI production: read Chapters 1 and 2, then the stage you are in. Know your tool: go to its chapter and read only that. Something has gone wrong: Reference D sorts failures by what you see or hear and points into the tool section that fixes each. The job finder lists the jobs a director actually has and the section that teaches each.
Next: The job finder
Comments