Making an AI microdrama in 2026 follows a ten-stage workflow that runs from season bible to finished 9:16 vertical master. The stages are not interchangeable and the order matters: skipping character lock before video generation is the most common and most expensive mistake new AI producers make, because it forces a full regeneration of every shot in which the character appears. The workflow below is the sequence that mature AI production pipelines in China and internationally have converged on, adapted for teams using a node-based canvas rather than disconnected tools.
Stage 1: Season bible and beat sheets
Every AI microdrama that scales to 60 or more episodes starts with a season document that precedes any generation. The bible specifies the central premise in one sentence, the two or three lead characters with their core want and wound, the genre (romance, revenge, supernatural, or the emerging hybrid categories), the act-level arc across the full episode count, and the placement of the major reveals relative to the episode wall, which conventionally falls after episodes 5 to 10. Without this document, AI generation produces individually competent shots with no narrative coherence across the season.
The beat sheet is the per-episode breakdown: one row per episode with hook beat type, friction escalation, spike moment, Zeigarnik cut type, and intensity register from 1 to 5. This document becomes the primary input to the pipeline. Everything that generates downstream is produced from or against it.
Stage 2: Character lock
Character lock is the most important stage and the most skipped by new producers. It involves generating a complete reference library for every lead character before any video is produced: frontal and three-quarter views, three to four expression variants covering the emotional range the story requires, and a costume reference for each recurring outfit. Seedance 2.0's @Character tag system and Kling 3.0's Elements consistency feature both require this reference material to function reliably. Without it, AI generation produces different versions of the same character across shots. Usable-take rates in character-locked pipelines in China now exceed 90 percent; in unlocked pipelines they fall significantly below that, which translates directly to generation cost and timeline.
Location references follow the same discipline: a locked environment asset per recurring set, generated at the correct 9:16 aspect ratio with the lighting conditions the story requires. Vertical composition for indoor sets means foreground-background depth staging with faces in the dominant frame position, not horizontal tableau.
Stage 3: Script-to-scene JSON conversion
The beat sheet converts to structured scene JSON at this stage: one object per shot with fields for shot type, characters present, location, action description, dialogue line, and the beat type it serves. This is the format that drives AI generation through a node canvas rather than prose prompts drafted per shot. On MinionArts Vertex, the scene JSON is the input to the generation nodes, so the beat sheet discipline is enforced structurally rather than individually.
Stage 4: Shot generation with model routing
Generation uses a routing decision per scene type rather than one model for the full season. As of mid 2026, the production routing that performs best across cost and quality: Kling 3.0 for action, confrontation, and motion-directed scenes where its 4K output and physics strength dominate; Seedance 2.0 for dialogue-heavy scenes where its phoneme-level lipsync and unified audio-video architecture produce the best natural dialogue synchronization; Veo 3.1 for hero moments, episode one hooks, and any shot that will be used as paid social creative. Value-tier models handle B-roll and establishing context where premium quality is not recoverable to the viewer.
All generation is image-to-video anchored on the locked reference frame. Text-to-video without a character anchor fails the identity consistency test at the first shot comparison across episodes. The practical generation rhythm for a 60-episode season: batch by location and character combination, generating all shots in a given scene context before moving to the next. This keeps reference context active and reduces prompt reconstruction overhead.
Stage 5: Lipsync and voice
Where the video model generates native audio for dialogue, the line goes into the generation prompt directly and no separate voice step is needed. Where it does not, the voice pipeline runs in parallel with generation: ElevenLabs with a locked voice ID per character produces the dialogue audio, then a lipsync pass maps the audio to the generated face. The voice ID lock is as important as the visual reference lock a character whose voice changes register or cadence between episodes breaks parasocial continuity even when their face is consistent.
For multilingual production, which is increasingly standard for global distribution, the voice step runs once per target language against the same locked voice IDs, producing a dubbed audio track that the lipsync node applies to the same generated video. One production, multiple language outputs.
Stage 6: Music and SFX
Music generation produces a small library of recurring cues at season level rather than per episode: a tension ramp, a romantic swell, a betrayal sting, a cliffhanger button. Reusing cues across the season builds the sonic identity that audiences associate with the series, which strengthens parasocial engagement rather than weakening it. SFX nodes add the format-specific sounds the genre requires: the slap, the door, the gasp, the ring tone, the footstep. These are not background texture; in the vertical format's silent-viewing context, burned-in SFX on the audio track are the primary emotional signal when sound is on.
Stage 7: Assembly, subtitles, and QC
Assembly sequences generated clips per episode, mixes audio layers, burns in subtitles at the lower third with appropriate clear space above for the face, and produces the 9:16 vertical master. Subtitles are not optional: ambient viewing means a significant proportion of viewing happens with device sound off, and the format's dialogue-driven emotional beats require text comprehension to land. The QC pass checks three things per episode: character face consistency against the reference library, Zeigarnik cut timing at the correct second, and subtitle accuracy and placement.
Stage 8: The repeatability layer on Vertex
On MinionArts Vertex, stages 3 through 7 exist as connected nodes on a single canvas. The season beat sheet feeds the scene JSON node, which routes to generation nodes by model, then to voice and lipsync nodes, then to the music library, then to the assembly and export node. The entire workflow exports as a JSON template. Interface Forms expose only the per-episode inputs so writers can trigger episodes without operating nodes. The Director Node iterates through the full beat sheet, running the pipeline for each episode in sequence with the producer reviewing output batches rather than building each episode from scratch. This is the architectural reason AI microdrama economics work: the first episode costs you the pipeline build; episodes 2 through 60 cost only generation credits and review time.
Stage 9: Distribution prep
Platform submission requirements vary. ReelShort and DramaBox have specific metadata, thumbnail, and episode naming conventions. Series-level assets include a thumbnail per episode (the hook frame at second one, optimized for the three-second scroll decision), a short synopsis per episode, and a series description under 150 words. Hook frame selection is a commercial decision as much as an aesthetic one: the frame that most clearly communicates unresolved tension in a static crop performs best as acquisition thumbnail and paid social image simultaneously.
Stage 10: Iteration from data
Platform analytics return episode-level retention curves, unlock conversion per episode, and drop-off timestamps. A 72 percent drop at episode 8 means the cliffhanger is not strong enough. A low conversion rate at episode 6 means the wall-adjacent twist is not landing. These data points drive script and prompt revisions for the next production or the current one if the platform allows episode replacement. This is the advantage AI production provides over live action: regenerating a 90-second episode at a revised beat is a generation job. Reshooting a live-action scene is a crew call. The iteration loop that defines the format's best producers closes in days rather than weeks.
