Chapter 11 — Final prompts, one scene generated and assembled
Status: DRAFT 1 (2026-09-21), DESIGN v6 §4, §3 W5, §6. Sources: docs/artifacts/ROUND18-RULES.md; shots/README.md; shots/4C/prompt_v12.md; production/phase3/ledger.csv rows 499–511; production/phase3/post-submission/CLAUDE-TAKEOVER-2026-09-15.md:10; production/phase3/submission-2026-09-14/edit/build_cut20.py:1.
11.0 Objective
Given a planning card, you can write the final prompt, generate one scene under fences, read each take, compare takes with the protocol, and cut the scene audio-leading with adjacent-shot playback before any second scene is generated.
11.1 What you are deciding
Two things: what the generator is told, and whether the first scene assembles before you buy the second. The prompt is written from the card, not invented at the composer: references in a fixed order, the motion beat as one continuous action with floor contact and furniture distance, the shot size described edge to edge, the eyeline, one line at most and the affect of the face while saying it. Then one scene, every card under its own fence, every take read five ways and chosen with the five-step protocol, and the scene cut with its lines placed before the next scene's first credit. "Cut in parallel with generation, not after" (PRODUCTION.md:699) is how three attractive shots stop becoming three rooms.
11.2 How we did it
Eight rules before any credit. The local pass (Chapter 1's dead GPU) judged 60 takes on five frames each and produced fail codes: staging 72, state 70, light 28, eyeline 24, extra person 22, identity 14 (ROUND18-RULES.md:3–7). Eight rules came out and were "applied to all 38 live cards before any Higgsfield credit is spent". Headwords, from the file:
- Cast scope equals frame. A character tag is loaded only for a person visibly in frame; inserts load the hand plate only.
- No not-yet states. No sentence describes a later or forbidden event, even negated; write the state that holds for the whole card and end the card before it changes.
- No anatomical sides. Locate by frame position and light ("the ear on the frame-right side of her head"), never "left ear".
- Faces are affects, not prohibitions. "listening, jaw relaxed, lips resting together, eyes on the hook", never "mouth stays closed".
- Floor and furniture. What both feet stand on and the distance in cm to the nearest furniture, for every standing character.
- Background bodies by design, not by count. "Nobody else" is not enough; the frame is built so no extra body can exist.
- Shot size edge to edge. "Close" is kept only next to what touches each frame edge.
- Blocking reference for choreography. A hand on a small prop gets a still showing the hand positions, as a reference, "never a first frame".
The prompt's shape. A fixed skeleton of blocks "in this order and never varied" (shots/README.md:16–23), ACTIVE REFERENCES second; each reference carries a scope clause ("This reference controls geometry, materials, surface finish and fine-material patina, light logic and atmosphere only", shots/4C/prompt_v12.md:29) and the block closes with a count line. Every tag appears exactly once. Card 4C's prompt v12 ran to about 2,100 words with 2 image and 1 audio reference of a 9 / 3 / 3 budget (:11). Prompts were versioned per card (v11–v19 by 2026-09-14); the version number is the retake record.
Take budget. ROUND18-RULES.md:23–25: "two takes hard cap; a retake only when the prompt changes shape, never to polish detail." Our retake share was 46 % anyway (Appendix A), which is why v6 makes the budget a per-card tier set in Chapter 6 rather than a law, and permits one same-prompt seed retake on turn shots only (DESIGN §6).
11.3 Case 11-A — the half-second
- What I saw: card 4C's approved line (take D) laid over its retake video. The mouth "opens 0.25 s, settles 3.5 s"; the line placed at +0.5 s "starts 0.5 s after first mouth opening" (ledger row 499). Two alternates were built at +0.15 s and 0 s "for ear review; none is a syllable pass". A dedicated lip-sync tool was prepared with the same audio and video, price not shown on its button, and left unsubmitted (row 504). The Director: "4C keep v1 (+0.5 s), no Sync3" (row 511); the build note reads "Director: v1 aligned better" (build_cut20.py:1).
- Why I think so: the video was generated with the line attached as a reference (Chapter 7's rule: it conditions, it does not time), so the mouth and the line were never the same performance. Placement by window got the mouth active inside the line; nothing could make it a syllable match.
- What I changed: DESIGN §6's sync rule: window alignment on a 24 fps strip is a candidate pass only; acceptance is by ear and eye on playback by the named owner; when a tool cannot deliver that, the three named paths (a lip-sync tool, a coverage rewrite, a shorter line) are the learner's to pick and record. The Director picked placement plus playback and declined the tool; the book does not call the tool incapable.
- Would this happen on your tool? Unknown until you generate one spoken card with its line and watch it. Do that on the scene's hardest line first, in Chapter 9, so that this chapter's choice is informed.
11.4 Do it with your tools
| Need | We used | Any tool | If yours lacks it |
|---|---|---|---|
| Ordered references with scope | up to 9 images, 3 videos, 3 audio per prompt, named by tag | any reference input | fewer references, more words; start frame carries identity |
| Prompt versioning | shots/<card>/prompt_vN.md, highest N is live |
any files | never edit a prompt in place; a new number per change |
| Placing a line on its shot | ffmpeg build with per-track delay and the line window measured (RMS) | build_cut.py's audio and audio_offset columns place one file per shot; any editor with two tracks does the same |
— |
| Audio leading across a cut (J or L cut) | not in our build script; done in the mix (Chapter 14) | any editor with two tracks | accept hard cuts on dialogue until the mix stage |
| Adjacent-shot playback | segment pairs exported and watched | any player | — |
11.5 Your turn
Time cap: two weekends. Spend: the scene's cards × the tier budget, from your sheet.
- Final prompts for scene 1's cards (C.12), from the cards, through the eight rules. Version them. Have the agent check every tag appears once and every rule's headword has a sentence behind it; have the second reader read one prompt aloud and say what they see.
- Generate the turn shot first, under its fence: price read, one click, verify, ledger. Read five ways. Impression first, then the numbers. Take-comparison sheet (C.5) once you have two.
- The rest of the scene, one card at a time, within budget. Rung 1 changes one line and counts against the card's budget; a turn shot may spend one of its three on a same-prompt seed retake; rungs 2–3 create a replacement card with its own tier budget, and the original's remaining budget is void; rung 4 is the budget spent; rung 5 is yours.
- Place every line over its shot by ear: measure the mouth window on a 24 fps strip, place the line inside it, then watch at full speed with sound and decide. Record the sync decision per shot as a case (C.6) or a director's-notes line (C.8).
- Assemble the scene with the conform tool (new cut folder) and watch every adjacent pair: eyelines meet, the room joins, the direction holds.
- ★ Second reader watches the scene and answers the C.14 questions for the scene.
- Blind pick on the turn shot. With your take-comparison sentence from step 2 already written and dated, play the turn shot's two takes to the second reader in an order they do not know, full speed with sound, and ask which one does what the card's goal line says (read them the goal line, not the prompt). Write their pick under yours.
- Prompt audit drill (no spend): mark the five seeded sentences below against the eight rules, then check the key.
The five sentences for drill 8
These are seeded for the drill, not from our prompts. For each, write the rule numbers it breaks, or "clean".
- "Her left ear catches the lamp; the earring is on the left."
- "He does not stand up yet; his mouth stays closed while she speaks."
- "Medium close: her head and shoulders fill the frame, hair touching the top edge, the chair back cut by the bottom edge; she is listening, jaw relaxed, eyes on his hands."
- "Nobody else is in the restaurant."
- "Crystal stands 40 cm from the table on the tiled floor; Rock is out of frame at her right."
11.6 Gate (W5 — First scene assembled)
Each line needs valid evidence per DESIGN §5; outcomes are pass / fail / unverified, and unverified blocks.
- Every card of the scene has a versioned final prompt with references in order and the eight rules applied (evidence: the files; the agent's check).
- Every generation has its fence, queued row, read row and five-way read (evidence: ledger and files).
- The turn shot was generated first (evidence: timestamps).
- Every card with two or more takes has a take-comparison sentence written before the measurements were read; a one-take card has a one-line keep/retake note against its goal line (evidence: the sheets and notes).
- Every line's sync decision is recorded with its path (placement and playback, tool, rewrite, shorter line) (evidence: the case or notes line; self-attested playback).
- The scene is assembled in its own cut folder with verification (evidence: the files).
- ★ Adjacent-shot playback with eyeline and direction check per cut, reviewed by the second reader (evidence: the form).
- Performance task: the turn shot's take-comparison sheet carries your sentence dated before the measurements were read, the numbers, and the second reader's blind pick; the pick and your choice agree, or the disagreement is resolved in one written sentence by the card's goal line and the take that serves the goal line is the one in the cut (evidence: the sheet, ★ the reader's pick in their hand, the cut's manifest naming that take).
- Performance task: your marked copy of the five seeded sentences matches the answer key on every sentence (evidence: the marked copy).
Pass example from HALFWAY: the 5C retake (ledger rows 571–572): fence verified, price read, one submit, a 2 fps original-versus-retake strip, then the Director's playback. Fail example: 4C at +0.5 s, a line placed inside a mouth that was never saying it, accepted by the Director as the least bad of three.
Answer key for drill 8
- Rule 3: "left ear", "on the left" are anatomical sides; locate by frame position and light.
- Rules 2 and 4: "does not stand up yet" describes a later event, negated; "mouth stays closed" is a prohibition where an affect belongs.
- Clean: edge to edge (rule 7), an affect (rule 4).
- Rule 6: a count, not a design; the frame must be built so no extra body can exist.
- Rules 1 and 3: Rock is out of frame, so his tag is not loaded and the sentence naming him goes; "at her right" is an anatomical side.