Introduction: From One Lucky Generation to a Production System#

1. The most dangerous misunderstanding#

AI video gave ordinary creators a very specific illusion for the first time: that if the shots come out beautiful enough, the work will follow. The illusion is stubborn because a five-second sample is genuinely thrilling. A character turns her head, a coat hem lifts in the wind, the light looks cinematic, the camera pushes slowly in. It already looks like film.

But a microdrama is not a collection of five-second samples. By the second shot the audience is already interrogating it. Is that the same person? Why did she change clothes? Where did the report in her hand go? She is looking to the right of frame — why is the antagonist also looking right? The door was behind the desk a moment ago; why is it by the window now? The narration has not finished, so why did the music just become a different cue?

A single shot only has to work locally. A microdrama requires many local results to sustain one continuous lie together: that these people exist in the same world, that things are happening to them in sequence, and that each of those things is pushing them toward an outcome.

So this book defines AI microdrama as state-driven audiovisual production. State means every fact that cannot be changed arbitrarily at a given moment: who the character is, what she is wearing, what she knows, what she wants, what is in her hand, where she is standing, which way she faces, what the light is doing, where the music sits emotionally, and which answer the audience is still waiting for.

Generative models will not maintain those facts for you. A chat log is not reliable production memory. Reliable production comes from assets, state, versions, approvals and handoffs.

2. Four kinds of quality that must be kept apart#

Teams tend to judge results with one vague word: the quality is good, or the quality is bad. That verdict cannot direct a rework. Commercial microdrama has at least four kinds of quality that are related but must never be collapsed into one.

The first is shot quality. Is the face normal, are the hands undeformed, is the motion natural, is the image legible. The second is continuity quality. Do character, wardrobe, space, action, light and sound carry across adjacent shots. The third is narrative quality. Does the audience understand the conflict, the desire, the causality, the suspense. The fourth is commercial quality. Will the target audience stay, follow, pay, watch an ad, or take the brand action.

A shot can be gorgeous and still wreck the story. A story can be perfectly coherent and generate no urge whatsoever to buy the next episode. An ad can win clicks and deliver exactly the wrong people into the series. Only when the four are logged separately does a team know whether to regenerate, recut, rewrite, or stop spending.

Take episode one of Backlit Takeover. Suppose the shot of Lin Yun raising the acquisition authorization is beautiful, but the text on the document is unreadable — that is a shot and compositing problem. In the next shot the document has jumped to her left hand — continuity. The audience cannot tell why a document changes the power in the room — narrative. The audience understands the reversal perfectly but does not want to continue — that is a problem of promise and commerce. Those four failures cannot be handed to the same "generate it again" button.

3. Production is an information chain#

A project begins as a vague intention: make a female workplace revenge microdrama. Market analysis turns intention into audience hypotheses and an emotional promise. Story architecture turns the promise into characters, secrets and season-long escalation. Episode writing turns escalation into beats and actions. Asset development turns characters and world into reusable references. Shot design turns action into shots. The generation station turns shots into candidate material. Sound and editing turn candidates into an experience. Distribution turns the experience into measurable data.

Figure 0-1 The full production loop of a commercial AI microdrama. Arrows are production dependencies. The return path at the bottom shows that distribution data may only write back into story and production decisions that are not yet locked — it can never overwrite an approved asset without a version.

Do not read the eight stages as eight departments that never speak. Each stage compiles the previous stage's abstract promise into an object the next stage can execute and accept. Story engineering does not output "a good story"; it outputs characters, beats and cut points. Assets and continuity do not output "some nice images"; they output identities, wardrobe, locations and props carrying versions and state. Data write-back does not let an algorithm edit the script; it produces new hypotheses, change requests and batch decisions.

Information can be lost at every handoff. The writer writes for the first time she is not afraid of her foster mother. If shot design draws an ordinary medium close-up, the emotional change is gone. Shot design places an old fountain pen on the desk; if the generation station does not know the pen is cross-episode evidence, it may swap in a ballpoint. If the editor does not know this moment is the power reversal, the reaction shot gets cut. If the campaign team sees only "workplace," the creative becomes generic career inspiration and attracts an audience that never wanted revenge.

The job of a production system is to make sure each station knows not only what to do, but why it is being done, what may not be changed, and what standard counts as finished.

4. The smallest unit of production is not a video — it is a shot record#

A file called final_3_really_final.mp4 sitting in a folder has no production meaning. The system cannot tell which episode it belongs to, which beat, which version of the character, what emotional function it carries — and it cannot work out whether the file is affected when the character's wardrobe changes.

Figure 0-2 The shot state packet. The image at the center is only a shot the edit can use when all seven classes of state — identity, wardrobe, props, space, action, light and sound — hold simultaneously.

Those seven do not change at the same speed. Identity is usually stable across a season. Wardrobe may change by scene or by story phase. Action changes within seconds. Sound often continues straight through a picture cut. The purpose of the shot record is to give all of those time scales one shared handoff point. If any module is missing, downstream must not fill the gap by guessing.

The meaningful minimum unit is a shot record with an identity:

shot_id: E001_S006
beat_id: E001_B03
story_function: Lin Yun takes back the floor for the first time
duration_target: 3.5
character_refs:
  - CH_LINYUN@identity_v03
wardrobe_refs:
  - WD_LINYUN_GREY_SUIT@v02
location_ref: LOC_BOARDROOM@day_v01
prop_refs:
  - PROP_ACQUISITION_AUTHORIZATION@v04
state_in:
  authorization_position: table
  emotion: restrained
state_out:
  authorization_position: right_hand_chest_height
  emotion: controlled_dominance
screen_direction: faces_right
audio_intent: room falls quiet before line
commercial_reason: lands the first identity reversal while withholding who the buyer is
status: keyframe_approved

This record connects story, assets, continuity, sound and commercial intent. The video file is only one realization of it.

5. Cheap decisions must come before expensive generation#

AI lowered the cost of a single generation, which makes it very easy to skip thinking. Rewriting a premise ten times in prose is cheap. Changing a character design after generation invalidates dozens of keyframes. Changing wardrobe after the video is finished triggers a full-episode rework. Discovering after launch that episode one promised the wrong thing turns every episode made so far into sunk cost.

So production advances through ten gates: premise and hook; beats and cut points; script producibility; asset lock; shot design; still keyframes; moving shots; edit and phone viewing; ad creatives; distribution data. Each gate asks the same question: is the current information reliable enough to justify spending more on the next stage?

A gate does not mean the work can never change. It means change carries a known price. A character identity can still be revised after lock — but the system must list the affected assets, shots and budget, and a named owner must approve it.

6. Three ledgers: facts, decisions, problems#

A working production system maintains at least three kinds of record.

The fact ledger holds what is currently true: Lin Yun's appearance, the time of day in E001, the position of the doors and windows in the boardroom. The decision ledger records why a particular story, shot or sound version was chosen, and who approved it. The problem ledger records what was found wrong, how severe it is, the evidence, the responsible station, and how it was resolved.

Without the fact ledger, the models drift. Without the decision ledger, the team relitigates settled questions. Without the problem ledger, the same error repeats across dozens of shots.

The problem ledger especially must never just say "character inconsistent." An actionable entry reads: "E001_S006 — Lin Yun's identity mole below the left eye is missing; face does not match CH_LINYUN@identity_v03; blocking severity; return to still keyframe, shot design does not need rewriting." The more specific the problem, the cheaper the rework.

7. Humans are not the final decorating pass#

"The AI produces it and a human takes one last look" is another dangerous fantasy. Human judgment belongs where the information value is highest: choosing the story promise, deciding whether a character is compelling, approving casting and style, confirming whether a performance is believable, judging whether an emotion actually landed, deciding how data should be interpreted.

Automation is good at checking format, references, dimensions, black frames, subtitle overflow, loudness, missing assets and some kinds of visual similarity. It cannot reliably decide whether a humiliation is enraging rather than tedious, whether a cry is restrained rather than ridiculous, or whether a cut point leaves the viewer eager rather than cheated.

A good agentic system does not remove the human. It frees the human from repetitive carrying, and forces a stop before the decisions that matter.

8. One complete miniature run-through#

Suppose the user offers a single sentence: make a female workplace revenge microdrama. The system should not generate a script. It first compiles the vague request into an opportunity brief: who watches, where, how it monetizes, what emotion they expect, what production cost it cannot carry. It then produces several premises and opening hooks, and a human picks Backlit Takeover.

The story station builds the season's secret and evidence ledger; the writing station fully drafts only the first three episodes. The asset station creates Lin Yun's golden reference, the old suit, the pen, the boardroom and the authorization document. Shot design breaks episode one into single-action shots and records Lin Yun's hands, props and eyeline state in each one.

Still keyframes face identity and space checks first. Only approved keyframes go into image-to-video. Results are marked usable, repairable, regenerate, or revise the shot design. Dialogue is generated by scene, music is laid by emotional section, ambience holds across shots. When the cut is assembled, picture continuity is reviewed with the sound off, then the sound is reviewed with the eyes closed, and finally subtitles and readable evidence are checked on a phone.

The finished episode is not immediately scaled to sixty. The distribution station cuts several hook creatives from the same episode to test which actually attracts the target audience: the humiliation of being fired, the identity of the acquirer, the evidence of an anomalous payment, or the missing mother. The data returns to the story station and decides which promise later episodes should strengthen.

The point of all this is not to add bureaucratic steps. It is to make sure every expensive act of labor is built on information that has already been verified.

9. Introduction checklist#

Before starting any project, answer these:

• Is the project optimizing for paid unlock, ad revenue, brand outcome, or platform acceptance?
• Which primary emotional satisfaction is the target audience here to get?
• What does episode one promise the audience, and what must eventually be paid off?
• Which assets will be reused across episodes, and when do they lock?
• Which states must be inherited across shots?
• Which judgments must be approved by a human?
• What is the stop condition at each stage?
• When a problem appears, who decides which layer it goes back to?

If these questions have no answers, generating faster usually just accumulates unusable material faster.

10. Exercises and deliverables#

Take a microdrama concept you intend to produce and write three plausible failures for each of the four kinds of quality. Then write one of its shots as a complete shot record, including at minimum story function, asset references, entry state, exit state, audio intent and commercial reason.

Deliverables for this chapter: the end-to-end process map, the four-quality table, your first shot record, and blank templates for the fact, decision and problem ledgers.

A note on sources#

The practical basis for asset-first working, shot-design review, manual inspection of stills, short-shot generation and edit export in this chapter comes from the in-depth Xiaohongshu corpus collected for this project and from directly readable Zhihu production articles. Those provide front-line process evidence. This chapter reorganizes that experience into a production system rather than treating any individual creator's claims about speed or earnings as general fact.