Part III — Episodes, Scenes, Dialogue, and Cut Points
Chapter 16. The Episode Pack and Script QC#
In this chapter
16.1 Why a literary script cannot be handed straight to generation#
A literary script tells a human reader what happens. It usually does not preserve the state, assets, risks, timing and commercial intent that production requires. The shot designer has to reinterpret it, the generation station guesses again, and the editor is the first to discover that no room was left for the cut point or the audio.
The episode pack is an episode's production contract. It compiles story facts into data every station can reference while keeping the readable script alongside. Any shot, sound or ad creative can be traced back to its beat and its commercial reason.
16.2 The ten blocks of a pack#
A complete pack contains: identity and version; goal and entry/exit state; the opening hook; repayments and new debts; current character state; scenes and assets; timecoded beats; dialogue and narration; the sound and subtitle plan; and cut points, risks, compliance and approvals.
Storyboards and generation records extend it later, but they reference the same pack rather than starting an unrelated numbering scheme.
16.3 The E001 pack for Backlit Takeover#
episode_pack:
episode_id: E001
version: 0.7
status: needs_review
duration_target_seconds: 75
purpose: prove the core fantasy and the counterattack-by-record engine
state_in:
CH_LINYUN:
public_status: former employee declared to have left
emotion: contained alertness
knows: she holds the acquisition authority
does_not_know: the father's account is still receiving funds
board_belief: Lin Yun is here to make trouble over old family business
state_out:
CH_LINYUN:
public_status: acknowledged representative of the acquirer
emotion: a private shock following a procedural win
board_belief: she is entitled to audit, but the authority's source is suspect
opening_hook:
type: public_exclusion_plus_hidden_status
visual: the cut badge lands beside the acquisition agreement
line: Former employees have no right to attend.
target_emotion: anger plus anticipation of contrast
pays:
- why Lin Yun is entitled to return to Shenghua
borrows:
- who authorized her to represent the acquirer
- why a dead father is still receiving payments
commercial_question: who is behind the acquisition, and why is the father's account still active
characters:
- id: CH_LINYUN
wardrobe: WD_LINYUN_GREY_SUIT
right_hand_in: PROP_BADGE
left_hand_in: PROP_OLD_PEN
forbidden: [crying early, volunteering her full backstory]
- id: CH_LINWEI
wardrobe: WD_LINWEI_IVORY_SUIT
goal: have Lin Yun removed before the directors arrive
- id: CH_ZHOULAN
wardrobe: WD_ZHOULAN_DARK_GREEN
goal: define the investigation as family revenge
locations:
- LOC_BOARDROOM_DAY
props:
- PROP_BADGE_CUTTABLE
- PROP_ACQUISITION_AGREEMENT
- PROP_AUTHORIZATION
- PROP_OLD_PEN
- PROP_LEDGER_UI
beats:
- id: B01
time: 0-4s
function: visible_exclusion
state_change: employee identity publicly removed
action: security cuts the badge; fragments land beside the agreement
dialogue: null
audio: scissors foregrounded, room tone briefly drops
risk: hand, scissors and badge contact is medium-high; breaking into shots permitted
- id: B02
time: 4-14s
function: pressure
state_change: Lin Wei defines Lin Yun as a troublemaker
action: Lin Wei does not look at her; raises a hand to security
dialogue: Former employees have no right to attend.
audio: low-intensity cool music bed enters
- id: B03
time: 14-25s
function: strategy
state_change: Lin Yun induces the opponent to leave an explicit record
action: Lin Yun starts recording and lays the phone face up on the table
dialogue: The reason. Say it again.
audio: music drops, phone confirmation tone
- id: B04
time: 25-42s
function: proof_and_reversal
state_change: Lin Yun obtains standing to attend and to audit
action: the authorization is pushed to the center; the chair checks the signature
dialogue: Then the acquirer's representative will attend.
audio: a short low frequency after confirmation, not covering the line
- id: B05
time: 42-58s
function: antagonist_adaptation
state_change: Zhou Lan moves from denying standing to attacking provenance
action: Zhou Lan closes the authorization, stopping it circulating
dialogue: Who gave you this authority?
audio: music pulse resumes
- id: B06
time: 58-75s
function: new_personal_debt
state_change: the investigation connects to the father's death
action: the ledger screen shows the father's name and yesterday's date; Lin Yun stops turning the pen
dialogue: Freeze this account.
audio: music pulled before the name appears, pen strikes the table, closing low hit
narration:
mode: first_person_retrospective
segments:
- id: N01
text: Zhou Lan never lies. She simply decides first, for everyone, whose word is not worth trusting.
use: optional; test a dialogue-only version against a narrated version
subtitles:
emphasis: [former employees, acquirer's representative, yesterday's payment]
avoid_regions: [character eyes, the authorization title, the name on the ledger]
cliffhanger:
type: proof_creates_personal_mystery
withheld: who is actually using the account
last_visible_fact: the father's name beside yesterday's date
next_episode_first_shot: the freeze request requires a live verification code
compliance:
- no real company or hospital identifiers
- documents and UI use fictional information
approvals_required:
- story_editor
- production_designer
- rights_compliance
This is not a final shot list. It preserves the purpose and risk of each action, leaving the shot director to decide sizes and cutting.
16.4 Versions and status#
Pack status includes at least: draft, needs_review, approved, in_storyboard, locked, superseded.
approved means only that the next station may begin. locked means any change requires impact
analysis.
Every version number comes with a change log: what changed, why, and which beats, assets, sound and cut points it affects. You cannot manage facts through a filename ending in "final 2."
16.5 The cold read: test comprehension, not politeness#
Give someone unfamiliar with the project only this episode, with no background explanation. Ask them who wants what; who is blocking it; at which moment the power changed; what they want to know at the end; and which people or objects they could not follow.
Do not ask whether it is good. Test the retelling first. If three people give three different core conflicts, comprehension has not passed. Record what they actually said rather than writing down the author's expectation as the result.
16.6 The table read: test language and relationships#
Have different people read the parts, timed, with natural interruption allowed. The table read looks at whether lines can be spoken, whether characters are distinguishable, whether information repeats, whether emotion has room, and whether the turns can be heard.
Do not explain while reading. Afterwards, mark stumbles, misreadings, overlap points and unnatural stresses. Scratch AI voice can supplement rhythm testing, but human reading exposes written-prose sentences more reliably.
16.7 The timed read: build a real time budget#
For each section record clean dialogue, pauses, action, reaction and transitions. If the total runs over, do not mechanically speed up the voice — decide which information, repetition or reaction can go.
Coming in short is not a licence to add empty scenery either. Check whether consequence, character choice, or reading time for the proof is missing. Being short is not the problem; an incomplete state change is.
16.8 The blind listen and the muted watch#
Listen to scratch audio with your eyes closed to test voices, relationships, causality and audio rhythm. Watch the storyboard muted to test action, proof objects and the shift in power. The two do not need identical information, but the core conflict must hold in both.
If it is incomprehensible with eyes closed, the character voices and forms of address may be insufficient. If it is incomprehensible muted, it may be leaning too hard on narration and explanation. Entrances are often muted, so key action has to be visually legible.
16.9 Asset review#
List every character, costume, location, prop, sound, music state and screen graphic. Each must already have an ID or be explicitly marked as new. Check first appearance, reuse value, rights status and production risk.
Casual additions that never enter the asset register cause face changes, location drift and runaway cost later. Writers may request assets; they must not let a generative model invent a critical asset on the spot.
16.10 Generation feasibility review#
The generation producer and the shot director check beat by beat: is there one primary action; what is the risk from contact, multiple characters, text, reflections, fast motion and lip sync; can it be broken into shots; can a still keyframe validate it first; is there an alternative.
High risk is not automatically deleted. Tag its commercial value, then decide the investment. The complex banquet shot at the first paywall justifies a larger budget; an ordinary transition does not.
16.11 Story QC scoring#
Scoring dimensions include: clarity of the first three seconds, goal and pressure, protagonist strategy, state change, preparation for the reversal, repayment of old debt, specificity of the new question, visual producibility, asset budget, and consistency with the promise.
Scores only locate weaknesses. Unclear rights, broken core causality, a dishonest cut point and severe budget overrun are blocking items and cannot be offset by high scores elsewhere.
16.12 Issues must be routable#
issue:
id: ISSUE_E001_014
severity: blocking
category: story_clarity
evidence: all 3 cold readers believed the father personally received money yesterday, rather than a name on an account
affected: [E001_B06, PROP_LEDGER_UI, E002_B01]
return_to: script_room
proposed_fix: add an account-holder field and have Shen Yan note a name is not proof of who operated it
owner: STORY_EDITOR_01
status: open
"This part is unclear" cannot be executed. An issue needs evidence, a scope of impact, a responsible station and an acceptance method.
16.13 SOP for compiling and approving a pack#
First, load facts from the series bible and the state ledger. Second, fill in the episode contract. Third, add character, scene and asset references. Fourth, compile beats, action, dialogue, narration and timing. Fifth, add sound, subtitles, cut points, risk and compliance. Sixth, run the cold read, table read, timed read, blind/muted checks, asset review and feasibility review in order. Seventh, route issues to the correct stations. Eighth, revise and record the changes. Ninth, obtain story, art production and compliance approval. Tenth, lock and hand to shot design.
16.14 Fault tree#
Symptom: the pack is long and still cannot be storyboarded. It is full of plot description and missing state changes, asset references and action before/after states. Recompile rather than adding more prose.
Symptom: every department is working from a different story. There are no shared IDs, versions or source of fact. Stop downstream production and establish which pack is currently approved.
Symptom: the table read passes and the cut still runs long. Only dialogue was counted; action, reaction and reading time for proof were missed. Do the five-layer timing.
Symptom: asset count suddenly balloons. The writers skipped asset review. Merge locations, reuse wardrobe, and require approval for critical new assets.
Symptom: QC raises many comments and nothing gets fixed. The comments have no severity, evidence or routing. Rewrite them as issue queue entries.
16.15 Checklist, exercises and deliverables#
Check that the pack has entry and exit states; that every beat has a function and a time; that characters and assets use stable IDs; that narration, dialogue, sound and subtitles are connected; that the cut point has a next-episode repayment; that risk and compliance are recorded; that all six reviews were completed; and that every blocking issue is closed.
Exercise one: compile one page of literary script into an episode pack. Exercise two: run a three-person cold read, recording only the retelling with no explanation. Exercise three: rewrite five vague comments as routable issues. Exercise four: after locking a pack, simulate deleting one prop and list the scope of impact.
Deliverables for this chapter: the approved episode pack, the cold read report, the table read and timing report, the independent audio and picture reviews, the asset requirement table, the generation risk table, the issue queue, and the approval and change records.
16.16 The pack manifest and immutable inputs#
An approved pack produces a manifest listing the version and hash of the script, beats, scenes, lines, state, assets, rights, shot requirements and acceptance criteria. Downstream tasks may reference only one complete pack ID; they may not assemble inputs from chat logs, old PDFs and personal spreadsheets.

Figure 16-1 Multiple creative and production inputs pass through validation and version locking to form an episode pack carrying a manifest and hashes. Shot design, sound, keyframes, continuity and budget tasks all reference the same pack, so no station reads a different version of the truth.
The chain on the right of the diagram represents references, not copies. Downstream may extract the minimum context it needs, but must retain the pack ID and the provenance of each field. When something is missing, raise an issue — do not invent a local "temporary fact."
Unlocking a pack produces a new version; older tasks are marked stale or continue running in an explicitly frozen batch. You cannot change one line in place and keep the same version number, because voice, subtitles and lip sync then become untraceable.
16.17 Gated script QC#
QC runs through four gates: facts and structure; performance and timing; assets and generation; and commerce and rights. A failure at one gate does not proceed to the more expensive next one — if a character's knowledge exceeds its permissions, fix the script before running a detailed table read. Each gate outputs evidence and issues, not just a total score.
Cold readers, the director, the asset owner, continuity and the producer work from different question sheets. Everyone commenting on everything at once produces duplicate notes and diffuse responsibility. The showrunner decides approval, conditional approval or return based on the blocking items.
16.18 Pack diffs and change impact#
New and old packs are compared automatically across facts, lines, state, asset requirements, shot requirements and commercial cut points. This is not a plain text diff: amounts, character knowledge and rights fields are high impact; a punctuation change may affect only subtitles; deleting a scene recomputes the episode's causality and budget.
The change report lists artifacts already generated, audio needing re-recording, shots needing redoing, assets that can be kept, and the regression scope. Approvers see the real cost before deciding — which stops "it's a two-word script change" from being mistaken for free.
A note on sources#
Chinese AI comic-drama practice consistently emphasizes the order running from script through assets, shot design, stills, video and edit. This chapter structures those handoffs so that human review becomes an explicit gate rather than a subjective rescue attempt after the video exists.