Part XVII — Hard Scenes: A Methods Library
Chapter 91. The Boardroom Confrontation: Six People Are Not Six Talking Faces#
In this chapter
The boardroom is a high-frequency scene in commercial microdrama and a combined stress test for AI generation: multi-person identity, fixed seating, eyelines, tabletop props, procedural information and shifting power all at once. The goal in episode six of Backlit Takeover is for Lin Xia to reverse the board using a proxy authorization — not for everybody to take a turn expressing an opinion.
91.1 Power relations and locking the space#
The scene power map.
On entry, Lin Wei controls the chair and the agenda, three directors support her, one is undecided, and Gu Zhou stays neutral. On exit, her supporters have not instantly defected, but procedural power has moved to Lin Xia. The scene's change is not that everyone looks more shocked; it is that the agenda, the evidence and voting eligibility have changed.
power_map:
entering:
agenda_owner: lin_wei
voting_proxy_holder: lin_wei
coalition: [dir_a, dir_b, dir_c]
swing: dir_d
exiting:
agenda_owner: chair_procedure
voting_proxy_holder: lin_xia
coalition_visible_crack: dir_b
next_question: why was the proxy witnessed by Gu Zhou
Lock the seats and identity slots first.
Draw an overhead floor plan and give every seat a stable ID. The head of the table, the door, the screen and the window direction do not change from shot to shot. Beyond an identity reference, each character needs a seat background, a wardrobe brightness value and a signature gesture: Lin Wei taps the table with her fingertips, Lin Xia keeps her left hand on the authorization, Gu Zhou folds his hands.
A multi-person wide shot is not required to render six faces sharply in close-up; it is responsible for geography and alignment. Identity detail belongs in single and two-person coverage. Demanding high-fidelity performance from all six people in a wide makes the model spread its resources and swap features.
91.2 Coverage, dialogue, and multi-person generation#
The coverage matrix.
| Coverage | Narrative function | People | Risk |
|---|---|---|---|
| Establishing wide | Seats, alignment, the chair | 6 | Identity swapping |
| Lin Wei medium close | Controlling the agenda | 1 | Background seats drifting |
| Lin Xia medium close | Producing procedural evidence | 1 | Hands, the authorization |
| Gu Zhou reaction | Witnessing without declaring | 1 | Over-surprise |
| Director two-shot | The crack in the coalition | 2 | Synchronized reactions |
| Authorization insert | A readable fact | 0 | Generated text errors |
| Vote hand group | Power changing | 3 | Hand count and direction |
Coverage is risk allocation. The main wide is used twice only: to establish at the top, and to re-establish after the vote. In between, the scene runs on stable reverses and evidence inserts.
Dialogue advances the agenda.
Lin Wei opens the item. Lin Xia challenges voting eligibility. Lin Wei attacks her standing. Lin Xia produces the authorization. The chair confirms procedure. Gu Zhou confirms he witnessed it. A crack appears in the director coalition. Every line changes the agenda, the evidence or the alignment; six people do not get equal speaking time.
The failed version gave each director a line stating their attitude, and ran two minutes with no power action. The final version lets three directors speak through a raised hand, a file pushed aside and an exchanged glance, and concentrates the lines on whoever actually changes the procedure.
Multi-person generation strategy.
Generate a seating baseline with no performance first, then lock the background for each camera position. In the wide, limit character movement to breathing and small eyeline shifts, and complete complex action in single coverage. Characters can be generated separately and composited, but table contact, chair backs and shadows must be unified.
Three or more in frame uses a cast layout: slot, identity, wardrobe, degree of visibility and action ownership. Background directors do not carry a full character prompt — only a stable silhouette and seat state — so they do not compete for identity attention.
91.3 Key action and the sound space#
The vote.
The vote is not six hands rising at once. The chair announces a fresh vote, cut to Lin Xia's steady gaze, then show two directors raising their hands separately, Lin Wei's hand resting on the table, and Gu Zhou not voting. The edit assembles the result, avoiding the generation of many simultaneous hands.
Hand inserts inherit cuff, seat direction and table surface. Each hand binds to a character ID, and the end state writes into the vote event. If the identity of a hand cannot be read from the wardrobe, the insert loses its narrative meaning.
Sound and space.
Room tone in the boardroom is restrained, and small table taps and paper become the sounds of power. Music empties out before the authorization lands on the table, then a low pulse enters after the chair confirms. Dialogue from different seats carries slight directional placement, but phone playback still puts clarity first.
An ensemble scene cannot accent every reaction. The real sound events are only the announcement of the item, the appearance of the authorization, the confirmation of procedure, and the first hand to defect.
91.4 Repair, acceptance, and a director's stress test#
Fault tree and repairs.
Viewers cannot tell who supports whom: seating and wardrobe do not differentiate enough. People appear to be reading in turn: dialogue has no agenda action. Faces change in the wide: too much identity detail was demanded. The vote is not credible: hands have no character attribution. Space jumps: background coverage was not derived from one parent space. The scene is loud but not tense: music and reactions have no hierarchy.
Acceptance, SOP and deliverables.
Muted playback shows alignment and the change in the vote. Sound alone follows the agenda. Every character's seat is stable. The authorization text is accurate. Hands are attributable by cuff. The vote event matches the exit state.
Process: write the power map. Lock the floor plan and seats. Design the coverage matrix. Write dialogue as agenda action. Generate single-character principal action. Composite limited wides. Produce evidence in post. Split the vote. Build space in the sound. Run five-layer QC.
Exercise: design a five-person inheritance vote, delivering power_map.yaml, seat_plan.png, cast_layouts/,
coverage_matrix.csv, vote_events.jsonl and scene_qc.md.
The director's stress drill.
First make a thumbnail animatic with no dialogue at all, expressing the power change only through seats, files, gaze and raised hands. If testers cannot tell who controls the agenda, space and action are insufficient, and adding explanatory lines will not rescue it. Then play an audio-only version and check whether the speaker, the item and the procedural turn can be followed; if all six voices and deliveries sound alike, go back to the voice bible.
Finally, cut the wide-shot generation budget in half and require the director to keep the alignment legible. This can be solved by reusing the establishing shot, adding medium shots that include seat backgrounds, and identifying votes by place cards and cuffs — not by reducing six people to three. The drill forces coverage to serve the narrative rather than "looking like a big production." The passing condition is that after the expensive wides are reduced, viewers can still draw the main seating accurately and say which action changed the outcome.
The release evidence pack preserves the final seating plan, the vote events, the authorization facts and the muted comprehension results. When lines change or a director is recast later, these invariants are compared first — not just whether the new shot cuts in.
A note on sources#
This chapter describes a staging method rather than corporate procedure. What transfers is treating an ensemble scene as a power map with locked geography, allocating identity risk across coverage, and proving legibility with the sound off.