AI Drama is the serialized-drama workbench in Video: one Series Workspace per series, one six-stage desk per episode, and generation that only runs after you confirm a price. This page is the operating manual; a shorter overview of the same workflow lives at AI Drama: Series Production in Six Stages, so existing links work.
Where you start
You enter from the Video module or the create entry inside any Project. The form is one short column: target Project and series name, both required; the format, a single-episode short or a multi-episode series; then a starting point, from an idea, an imported story or script, or an empty desk. Language, aspect ratio, episode duration, and visual direction sit folded under more settings as prefills, never gates.
Both formats build the same Series Workspace; a single-episode short is a series with one episode so far, so episode 2 later moves nothing. Creation is free and atomic: a failed save leaves no half-built workspace, errors appear next to their input, and typed content survives. An imported story is stored exactly as pasted, unparsed and uncharged.
A new short opens on episode 1, stage 1; a new series opens in story planning. Reopening sends multi-episode series to the overview and shorts straight to the stage waiting on you; deep links land on the exact episode, stage, or review item.
Navigation is fixed at six items: overview, story planning, asset hub, episode production, tasks and cost, publishing. The overview is a status hub rather than a seventh stage, listing progress, review backlog, blockers, running tasks, and a next step, with an "awaiting my review" queue holding the same candidates you would see in-stage.
The six-stage desk
| Stage | Confirming locks in |
|---|---|
| 1. Story and script | Adopted script plus the series snapshot it follows |
| 2. Shot planning | Executable shot sequence and requirements list |
| 3. Characters and wardrobe | Adopted character and wardrobe references |
| 4. Scenes and props | Adopted scene and prop references |
| 5. Shot production | Each shot's adopted video clip |
| 6. Sound and assembly | Reviewable, exportable, versioned episode |
The order is the recommended path, not a locked wizard; return upstream anytime, and only a missing objective input blocks a specific action. Confirming a stage, a task finishing, and adopting a result stay three independent states, and entering a stage generates and bills nothing. Anything you already own can be imported, bound, and confirmed. Nothing bills early.
Forward moves through one button, confirm and hand to the next stage, which opens a handoff card of outputs, bound assets, open questions, and missing inputs; missing ones link straight back to the editor. Handoffs are free, create no version, and never block on creative doubts.
Stages 1 and 2 are text work. One explicit request for a full script or a complete shot table is one text generation unit; ordinary back-and-forth bills on the shared assist pool. The shot table needs no storyboard images: create, duplicate, split, merge, delete, and reorder shots by hand, and leave a slot as a typed requirement until you bind an asset. Confirmation checks only the objective minimum, a scene order with content in stage 1, one valid shot in stage 2.
Stages 3 and 4 prepare this episode's view of series-shared characters, scenes, and props. No second asset library exists, and each defaults to the shot table's references. Every slot can reuse a specific Character and Scene version, an upload, or a new candidate from AI Drama, and each explicitly requested image is one 1-credit unit. A wardrobe holds named expression, pose, outfit, and accessory references with front, side, and back optional; extra angles form a reference set, never a new version. Day, night, weather, and lighting variants leave the master scene alone, version changes list affected shots while keeping old results, and pure text-to-video shots never force a character drawing first.
Stage 5 is per-shot by design. Each shot picks its own route: reuse a specific Video Studio result, upload footage, generate from text, or from one confirmed image; there is no whole-episode generate-everything switch. First and last frames can be prepared with bound references for continuity, but a video with an end frame still uses the first frame only, and the interface says so. Test shots, one to three recommended, are skippable; every new video is one 10-credit unit, confirmed with its full batch total first. Candidates land in review beside their references; adopting a clip neither confirms the stage nor starts sound work.
Stage 6 keeps five content types: main video, dialogue and narration, BGM, sound effects, and subtitles, shown as a shot filmstrip beside grouped sound lists. Assembly places adopted clips in shot order, reorders by drag or buttons, trims contiguously, and cuts hard; there is no clip splitting or transition library. Uploaded footage keeps its original audio at basic volume or mute. Dubbing lines bind a role, text, time range, and voice source: reused Voice Studio audio and uploads attach to specific lines, while synthesizable voices are platform built-ins or Cloned Voices from the current Project, held in a series casting table episodes inherit. BGM binds to the episode or a range, effects to a shot or a time point, from Music Gen and Sound Design results, uploads, or new generation. Captions start from confirmed dialogue, arrive by import, or come from recognition at 0.5 credits a request. The pre-export check blocks only unrenderable states; missing audio or subtitles only draw a note, and no automatic lip-sync exists.
Where the copilots appear
Nine specialists share the workspace; each page shows the one that belongs there: the producer-coordinator on the overview, the showrunner in series planning and stage 1, the visual director on the visual bible, the storyboarder in stage 2, the two designers in stages 3 and 4, the shot director in stage 5. Stage 6 offers two modes, sound director and editor, over one shared timeline that neither can overwrite.
Copilot output reaches production only by adoption. A workable suggestion arrives as a card with the concrete change built in and a short adopt button; adopting touches only the named fields, and chat never silently rewrites the editor. Answer "use this character, don't generate," and it builds a selection or binding card instead of a generation task.
One main conversation exists per series or episode, page, and copilot, kept across refreshes and devices; new ones start from current data, old messages carry a stale-advice notice after changes, and chat history stays private. Ordinary questions bill on the shared assist pool, and prompt work inside a paid generation is part of that unit.
What always needs your confirmation
Four review states cover every candidate: pending review, adopted, needs changes, intentionally kept, written as text plus icon, never color alone. The system never adopts anything, and batch adoption processes only checked items, after showing which current versions get replaced.
Money confirms every time: before a paid generation runs you see scope, unit count, unit price, and fixed total, and a batch lists each shot's execution method. Cross-episode batches take one primary media type per confirmation. Publishing, un-publishing, download permission, and Remix authorization are separate explicit actions. Deletion asks for confirmation, is unrecoverable, and protects running objects; an active task blocks deleting a finished video and says why.
Redos, edits, and upstream changes
Upstream changes never rewrite finished work. Revise a character version or the series bible and the system flags dependent episodes, shots, or audio as needs review, says what changed, and leaves old results in place. Only your confirmation clears the flag; opening pages, copilot replies, and task completions never do. Keep the old version, move to the new one, or take part of the impact and stop.
Regenerating because you dislike a result is a new paid unit, a fresh 10-credit video in stage 5; technical retries inside a task are free, and a technical failure releases its reservation. Version upgrades show a real impact preview before you pick which episodes or shots move; finished and published episodes stay out by default. In stage 6 the rule is per line: script edits mark old audio text has changed, a new series default voice marks only ungenerated lines, and a single-line redo rebuilds only that unit.
When tasks fail
Accepted tasks keep running after you close the tab, drop the connection, or switch devices; on return the page reads state from the backend, never from the browser. The task center, episode badges, and stage markers show every state including partial failure, each with a way back to the object. Queues show a real position, like 2 of 3 running, never a made-up percentage or arrival time.
Batches settle per generation unit, so partial failure is a normal outcome: successes stay, and you retry, replace, or discard only the failures. The queue lets you move a waiting item to the front, pause and resume a batch, and cancel parts that have not started; running units cannot be reordered, and a plan downgrade never stops running work.
Drafts save continuously and switching episodes saves first. A failed generation keeps its placeholder and inputs while other tracks stay intact, a failed export retries free and creates no version, and the failure reason, settlement, and recovery entry survive a refresh.
Where results live and how deletion works
Every generated image, clip, voice line, track, or effect exists as one Result, stored in the Project's matching Results area, AI Drama holding a reference to the same file. Each carries structured source and usage, displayed like "AI Drama · Night City · Episode 3"; results born in Image Studio keep their source even after adoption here.
Deletion has two units. Removing a use binding leaves the Result alive; confirming deletion of the Result itself is final, with no recycle bin and no undo. A finished video can return to the episode assembly or be deleted, one confirmed version at a time. Exports version in one direction: the first creates the artifact and its first version, later exports append, and old versions stay reachable.
Publishing and the public boundary
Export comes first: it renders the current assembly into a deterministic MP4, calls no generation model, and costs no video credits. The minimum exportable episode holds one shot with an adopted, processable clip; dubbing, music, effects, and subtitles can be empty on purpose. Private exports and downloads carry no platform brand watermark. Export is just rendering.
Publish is a separate action that binds exactly one fixed artifact version; re-exporting never quietly updates what the public sees. The payload contains what you chose: that version, a cover, title, synopsis, series name, episode number, and public author identity, plus creation notes only if you enabled them. It never contains the Project, drafts, the full script, per-shot generation notes, fixed prompts, relationship secrets, foreshadowing, copilot chats, tasks, credits, or unselected assets. A per-episode public final prompts switch, off by default, publishes an editable summary of the confirmed premise, visual, mood, aspect, and sound direction, and a new cover costs one 1-credit image unit, adopted before it goes public.
The public series page is a curated aggregation: new episodes never join automatically, re-publishing never swaps a listing without you, and un-publishing retires the link and listing while the private version stays; with zero public episodes the page is inaccessible. Public copies and visitor downloads carry attribution and platform watermarks, and un-publishing cannot recall saved files. A safety check runs first; without an allowing conclusion the episode stays private.
What generation costs
| Output | Credits per unit |
|---|---|
| Image | 1 |
| Voice lines and sound effects | 2 |
| Music | 5 |
| Video | 10 |
| Auto subtitles, per request | 0.5 |
Every explicitly requested new image, whether a character reference, scene variant, starting frame, or publish-time cover, is one 1-credit unit. Each new video is one 10-credit unit, and reused or uploaded shots cost nothing to place. Dubbing lines that fit inside the single-unit limit merge into one 2-credit confirmation and split visibly with a fixed total when they exceed it. A full script draft or complete shot table from one request is one text generation unit at the standard text rate.
Everything else is free: uploads, reuse, binding, structured edits, stage confirmation, review, timeline editing, export, publishing, and hand-edited captions.
Plans cap concurrent running units across every feature and device: Free 1, Lite 2, Pro 3. Full slots never reject a task; accepted work queues with credits reserved and starts when a slot frees. A series can carry an optional credit budget set to remind or block new tasks; blocking stops only new tasks in that series, and cancelling a queued item releases its reservation.
On a phone
The phone is a full production entry, not a review screen: create series and episodes, edit structured content, talk with the current copilot, choose and bind assets, manage shots, submit and watch the queue, review candidates, do basic assembly edits and export, publish, start a Remix. Navigation, editing, copilot, and review switch full screen; shot lists become expandable cards, graphs and timelines offer list alternatives, assembly presents a clip list with sound groups, and wide tables scroll only inside their container.
The shortest first loop: import a short script, confirm it, plan two shots, bind one ready video, generate one 10-credit clip, export. Episode 2 then starts from that same workspace.