AI production workflow · 2026
Script to shotsto episode
7 gates
Each stage produces something the next stage can verify.
Make a short drama as a controlled sequence of references, shots and decisions—not one giant prompt. This workflow turns a vertical episode brief into editable coverage with attempt logs and a defensible production cost.
Live shot packet
Build one controllable shot
Generated prompt pair
Adapt model-specific syntax after the story and continuity locks are stable.
Reference frame
Vertical 9:16 cinematic frame in a narrow late-night pawnshop lit by warm pendant lights. Mara, a poised woman in her early 30s wearing a scarlet wool coat. Composition: a medium two-shot with both characters clearly visible. Continuity anchors: Mara’s face, scarlet coat, silver locket, warm lighting, and counter geography. Natural facial proportions, readable hands and prop placement, grounded short-drama production design.
Motion pass
Mara slides a silver locket across the glass counter; the shopkeeper recognizes it and freezes. Camera movement: a restrained slow push-in while both eyelines remain stable. Hold Mara’s face, scarlet coat, silver locket, warm lighting, and counter geography throughout the take. One continuous take with restrained performance, natural body motion and a clear final story beat.
Use only references, faces, voices and source material you have permission to use.
The production chain
Every step ends with evidence.
Do not advance because a single frame looks impressive. Advance when the named deliverable is approved and recorded.
- 01
Story brief
Lock the delivery
Define episode length, 9:16 delivery, audience rating, language, audio route and the rights you hold before generating any asset.
Deliverable
One-page episode brief
- 02
Visual identity
Build the continuity bible
Create approved reference frames for each recurring character, hero prop and location. Record wardrobe, lighting, geography and forbidden variations.
Deliverable
Character + location board
- 03
Coverage
Break the scene into shots
Turn the script into shots with one story job each. Start with a two-shot, reaction and moving reveal before expanding coverage.
Deliverable
Numbered shot list
- 04
Pilot route
Choose the bottleneck tool
Choose by the hardest requirement—references, multi-shot action, native audio or finishing—not by the strongest public demo reel.
Deliverable
Primary + fallback route
- 05
Composition
Generate the frame first
Approve character identity, prop position, eyeline and screen direction in a still reference before paying to animate the shot.
Deliverable
Approved start frame
- 06
Motion
Animate one clear beat
Describe subject motion, camera motion and scene motion. Add one variable at a time and record every failed, discarded and accepted attempt.
Deliverable
Attempt log + selected clip
- 07
Finished sequence
Edit, audit and release
Cut the sequence before judging quality. Repair continuity, finish audio, check captions and rights, then calculate cost per usable second.
Deliverable
Master + production record
Minimum viable coverage
Test three shots before an episode.
The pilot pack exposes character, prop, eyeline, geography and camera-motion failures without paying for a complete scene.
Two-shot
Proves both characters and the prop can share the frame while the story action remains readable.
Count every attempt
Reaction close-up
Tests face, wardrobe, eyeline, lighting and whether the edit can cut into the emotional beat.
Count every attempt
Moving reveal
Tests prop continuity, camera motion, location geography and the second character across the reveal.
Count every attempt
Route by bottleneck
Assign tools a production job.
Runway
Reference-led direction
Start here when recurring subjects, controlled frames and generation close to finishing are the main requirement.
Kling AI
Multi-shot action + audio
Test this route when connected character action, longer coverage or native audio is the difficult part of the scene.
Luma
Frame-directed iteration
Use a frame-led route when key compositions, extensions, reframing or finishing passes drive the workflow.
Migration warning · Sora 2 API
OpenAI marks the Videos API as deprecated and scheduled to shut down on 24 September 2026. Do not make it the dependency for a new long-lived production pipeline.
Pilot gate
Greenlight the pipeline, not the prettiest clip.
Cut the three pilot shots together and review the sequence against the same release requirements the episode will face.
Open the budget calculator →Greenlight when
Stop and revise when
Identity survives all three edits
Face or wardrobe changes between coverage
Prop and screen direction remain readable
The edit reverses geography or loses the prop
Dialogue and audio route fit the schedule
Lip sync or voice repair dominates the shot
Retry-adjusted cost fits the episode cap
One accepted clip hides an unsustainable retry rate
Rights and source records are complete
A face, voice, image or music source is unclear
Rights checkpoint
A generated shot still needs a rights trail.
Keep permission records for faces, voices, reference images, scripts, music and branded material.
Record the model, host platform, plan, generation date and source assets for every final shot.
Verify live commercial-use and watermark terms before final generation—not after release.
Escalate unclear likeness, copyright or contract questions to qualified legal counsel.
FAQ
Before the first episode
Can AI make a complete short drama?
AI can generate many production assets and shots, but a reliable episode still needs human story decisions, continuity control, editing, audio finishing, rights review and quality assurance. Treat the generator as one department in a production pipeline rather than a one-click episode machine.
How do you keep AI characters consistent between shots?
Start with approved reference images, keep wardrobe and lighting anchors explicit, generate stable start frames before animation and change one variable at a time. Judge consistency only after the shots are edited together, because isolated frames hide eyeline and geography problems.
What should an AI video prompt include?
Separate the reference-frame brief from the motion pass. The frame brief establishes subject, scene, composition, lighting and continuity anchors. The motion pass describes the story action, camera movement, scene movement and the final beat of one continuous shot.
How many generations does one usable AI video shot need?
There is no honest universal number. Log attempts on a locked three-shot pilot, then use that project-specific retry multiplier in the production budget. Public demo reels and isolated successes do not provide a usable acceptance rate.
Should a new AI short drama pipeline use the Sora 2 API?
No new long-lived pipeline should depend on it. OpenAI's current video-generation documentation marks the Sora 2 Videos API as deprecated and scheduled to shut down on 24 September 2026. Existing users should plan migration rather than new production dependence.
Controlled test ledger
Record queued requests, failures, rejects and accepted usable footage in one evidence chain.
Choose an AI video tool
Compare seven creator and developer routes by production bottleneck.
Calculate the budget
Price retries, rejected generations and each usable finished second.
Runway vs Kling
Compare workflow, native audio, API cost and the same locked scene.