My last attempt at a vertical drama died in a group chat: three freelancers, fourteen days, $430 in deposits, and one finished 90-second episode nobody watched twice. The storyboard artist vanished at shot nine. The voice actor billed per line and delivered half of them. I spent a weekend stitching the remains in a free editor and shipped the result with someone else's watermark on it. Episode 2 never happened. This month I reran the same kind of episode alone in x-rush's AI Drama workbench, kept a receipt for every step, and the whole loop, script to publish button, is below with real numbers. Every credit figure quotes the billing sheet as of October 12, 2026; if the sheet moves, the math moves with it.
The workspace comes first
AI Drama lives inside the Video module or any project's create entry, and the form is one short column: project and series name, format, starting point. I named mine Night Market, picked multi-episode, and chose import, pasting in the corpse of the old script. It stored the paste exactly as written, unparsed and uncharged, and creation itself cost nothing — a failed save leaves no half-built workspace behind. A single-episode short is the same object with one episode in it, so nothing migrates when episode 2 eventually shows up. Inside, navigation is fixed at six places: overview, story planning, asset hub, episode production, tasks and cost, publishing. The overview is a status hub, not a seventh stage; the desk itself lives in episode production. (The full operating manual lives in the AI Drama workflow docs. I'm giving you my path through it, not the spec.)
Words first: stages one and two
Each episode runs through six stages, and stage 1, story and script, opens with the showrunner copilot. My request, verbatim: "Episode 3, night rain. The stall keeper's secret comes out. End on the lantern going dark. Keep it under twelve scenes." One explicit request for a full episode script is one text generation unit, 0.5 credits; the twenty minutes of back-and-forth after it bill on the shared assist pool, which I tally at the end. The confirm check is deliberately thin: a scene order with actual content. Nobody grades my dialogue.
Stage 2, shot planning, is the one I used to pay freelancers for. The storyboarder turns the confirmed script into a shot table, and one request for the complete table is another 0.5 text unit. Every shot carries frame content, cast, scene, framing, camera move, dialogue, and target length; empty slots can stay as typed requirements until I bind something. I split two shots, reordered three, and renamed one "THE BELL, WIDE, no zoom" by hand — edits cost nothing because they aren't generation. Confirming needs one valid shot, and mine had ten.
The old way was a week of DMs and a shared spreadsheet. Now it's two requests and twenty minutes of dragging rows.
Stages three and four: a cast worth reusing
Stages 3 and 4 prepare this episode's view of the series-shared characters, scenes, and props, and both default to whatever the shot table references. Each slot takes three routes: reuse an existing Character and Scene version, upload, or generate fresh inside the desk. Every explicitly requested image is one 1-credit unit, and every candidate waits for a human verdict — the review states are pending review, adopted, needs changes, and intentionally kept, and the system never adopts on its own.
I generated nine images for 9 credits: one reference each for the two leads, a rain-soaked wardrobe variant, a market master shot plus its night-rain variant, one lantern prop, and three composed starting frames for shots that needed the stall keeper and the alley in frame together. The night-rain variant never touches the master scene, and the reference sets carry names like "Mei, soaked, ep3", so episode 4 can find them without me. Two of my ten shots went text-to-video and needed no drawings at all; nothing in the interface forced one.
Nine images for an episode's whole visual world. The old quote for equivalent character sheets started at $150 per artist.
Stage five: where the credits actually go
Shot production is per-shot by design, and there is no generate-the-whole-episode switch to click by accident. Each of the ten shots picks its own route: reuse a Video Studio result, upload footage, generate from text, or generate from one confirmed image. My batch: eight generated, two uploaded. The system suggests one to three test shots first, flagging the risky ones, and I ran the bell wide as a test — a single 10-credit unit. The confirmation for the remaining seven showed the scope, each shot's method, seven units at 10 credits each, 70 total, fixed — and I clicked once.
Here's the trap. A redo because a clip is ninety percent right is a fresh 10-credit unit, and nothing nudges you to stop; technical retries inside a task are free, taste retries never are.
Shot 7 came back failed on the first pass. The other seven didn't care: batches settle per unit, successes stay ready for adoption, and I retried only the failure. I had closed the laptop mid-render, and the task kept running server-side; reopening showed the queue's honest face, "2 of 3 running," with no invented percentages. I run Lite, which means two concurrent units; Free gets one, Pro gets three, and that cap spans every feature and device I'm signed into.
Adopting the eighth clip advanced nothing, which I first read as a stall. It's by design: confirming a stage, a task finishing, and adopting a result are three separate states, and the move to stage 6 happens when I click "confirm and hand to the next stage" and read the handoff card it opens. The handoff is free and calls no model.
One more charge happened here, self-inflicted: the finale shot bugged me until I regenerated it. One click, ten credits.
Stage six: dubbing, music, effects, captions
Stage 6 is one assembly desk with five content types: main video, dialogue and narration, BGM, sound effects, subtitles, laid out as a shot filmstrip beside grouped sound lists. The adopted clips dropped in shot order; reordering is drag or buttons, trimming is contiguous, cuts are hard, and uploaded footage keeps its original audio until I mute it.
Dubbing starts from a series casting table. My stall keeper got the seventh built-in voice in the list — no system to the choice — and the narrator is cast separately. Cloned voices work too, provided they already exist in the current project. Eleven dialogue lines fit inside one audio unit's limit, so one confirmation, one 2-credit charge, covered the whole batch; a bigger cast splits into visible units with the fixed total shown first. Two days later I rewrote one line, and the old audio flagged "text has changed" instead of vanishing. Rebuilding that single line was one more 2-credit unit.
BGM took one 5-credit music unit, bound to the whole episode. My brief, verbatim: "slow strings, rain on tin, no drums, sad but warm, loops." Music has no section redo; I'd change the brief and regenerate the track, and both versions would stay.
Effects came from Sound Design: three cues at 2 credits each — rain drumming on the awning, a wok sizzle at close range, one distant market bell. Effects bind to a shot or a time point and can overlap. The cheapest trick I learned: Sound Design's curated catalog of ready sounds is free to preview and copy into the project, so two of my three cues could have cost 0. I paid for the bell because I wanted it a half-second shorter than the catalog's.
Captions built themselves from the confirmed dialogue with one 0.5-credit recognition request, and my two manual spelling fixes were free.
The ledger
Every number in this loop, added up:
| What ran | Units | Credits |
|---|---|---|
| Script draft, one explicit request | 1 text | 0.5 |
| Complete shot table, one request | 1 text | 0.5 |
| Images: 2 cast refs, 1 wardrobe, 2 scenes, 1 prop, 3 starting frames | 9 image | 9 |
| Shot clips, 8 generated (2 uploads placed free) | 8 video | 80 |
| Dubbing, 11 lines merged into one confirmation | 1 audio | 2 |
| One-line rebuild after a late script edit | 1 audio | 2 |
| BGM, one full-episode track | 1 music | 5 |
| Effects: awning rain, wok sizzle, market bell | 3 audio | 6 |
| Auto-captions, one recognition request | 1 request | 0.5 |
| Copilot assists, 14 successful | shared pool | 0.7 |
| Finale-shot redo, self-inflicted | 1 video | 10 |
| What episode 3 actually cost | 116.2 |
Skip the redo and the late edit and the clean run is 104.2. The pool line needs its own word: ordinary copilot questions bill on a shared assist pool at 0.5 credits per ten successful assists, failed replies don't count — one of mine died mid-sentence and cost nothing — and there is no meter to watch, because it settles server-side. The real discount is episode 4: cast, scenes, voices, and the series' visual direction carry over for 0, so only new images, clips, and audio ever bill again.
What the loop still gets wrong
The honest caveat: nothing in this desk judges taste; the system verifies only whether the next action can reliably execute. Specific gaps I hit. There is no lip-sync, so close-ups talk with soft mouths. Assembly has no clip splitting and no transition library — hard cuts only, and my montage ambitions died there. You can prepare an end frame per shot for continuity planning, but a generated clip still renders from its first frame only, and the interface says so plainly. Music regenerates whole. Each effect unit delivers one result, no A/B pile to pick from. And on Free, the single running slot means an eight-clip batch queues one at a time.
None of that blocked the export. One of them did change how I write shots: I stopped scripting cuts the timeline can't make.
Export, then publish, in that order
Export renders the current assembly into a playable MP4 deterministically: no generation model, no video credits, free retries if a render fails. The first reliable export creates the episode's artifact and its first version; every later export appends a version instead of overwriting. The minimum bar is one shot with an adopted clip. Two of my shots run intentionally silent, which draws a note and lets it pass.
Publishing is a separate button behind that. It binds exactly one fixed version, so re-exporting never quietly changes what the public sees. The public payload holds the chosen version, cover, title, synopsis, series name, episode number, and my public author identity. The script, per-shot notes, fixed prompts, and copilot chats stay private no matter what; the per-episode final-prompts switch, off by default, adds only an editable summary of story premise, visual direction, mood, aspect ratio, and sound direction. I reused a frame as the cover for 0; generating a new one would be a 1-credit image unit. Public copies and visitor downloads carry attribution and platform watermarks, un-publishing doesn't recall files already saved elsewhere, and episodes join the public series page only when I add them myself.
Going deeper
- AI Drama workflow — the operating manual for every stage named above
- Credits system guide — the billing sheet behind every number in the ledger
- Sound Design studio — where the cheap half of an effects layer lives
The group-chat version of this episode cost $430 and died at shot nine. The solo version cost 116.2 credits and one weekend, and it ended at a publish button I pressed on a Tuesday. Steal the shape: one episode, ten shots, one sitting. Or go smaller: two shots, one reused clip, a free export, and see whether the loop fits your story before you write twelve of these.