The Solo Creator's AI Short-Drama Workflow: From Idea to One Episode a Day

One person can sustain an episode a day. The rule is to separate one-time setup from the daily loop: parsing a manuscript takes 3–5 minutes and reference art 30–60 minutes of machine time, once per series; after that each day is only "review storyboard → generate → pick → composite → publish". Budget 1.5–2.5 hands-on hours per episode, 40–60 minutes of unattended generation, and roughly $6–11 per 90-second episode at 720p or $13–23 at 1080p (typical, from SinCoSphere editorial testing). Batch storyboard review weekly; generate daily.

Conclusion first: review is the bottleneck, not generation

For one person, machine time is not the constraint. A 90-second episode of 6–8 segments comes back in 15–20 minutes when segments run in parallel. What eats the day is human review: checking storyboards, picking takes, re-reading subtitles and lip-sync — 30–60 minutes per episode at minimum (typical, from SinCoSphere editorial testing). So the scheduling rule is simple: batch everything batchable (parsing, reference art, storyboard review) into a prep day and keep the daily loop to generate → pick → composite → publish.

Cadence depends on the channel. ReelShort-style apps run long series and release episodes in blocks, so you need a bank before launch. YouTube Shorts and TikTok reward a daily post; Shorts accept up to 3 minutes since 15 October 2024, but drama completion falls off sharply past two. Everything below assumes a 90-second episode of 6–8 segments.

Where the story comes from: original, public domain, licensed

Three sources, three risk profiles.

1. Original script. Cleanest rights, slowest start: a 60-episode outline plus scripts is one to two weeks of writing, so bank ten episodes before launch. Under the US Copyright Office's January 2025 guidance, the human-written script is the part of the project that is clearly protectable.
2. Public domain. Older novels are free to adapt (check the year: US terms for older works run up to 95 years from publication), but modern translations and annotated editions are still protected; work from the original.
3. Licensed IP. Fastest to ship, most contractual risk. The license must name AI adaptation and derivative works, the distribution channels and the term; a screenshot of a DM is not a license.

Story selection follows short-drama rules: a hook in the first 3 seconds, a conflict and a cliffhanger in every episode, no more than 3–4 locations per episode, and no crowd scenes or long fight choreography — models still fail at those.

One-time setup: parse once, build the art once

The first day of a new series does two jobs that never repeat.

Parsing. The manuscript becomes a cast sheet, scene sheet, prop sheet and episode outline in 3–5 minutes. End-to-end tools (SceneMixer, LibTV, LTX Studio) build this in; with single-purpose generators (Kling, Hailuo, Runway) you write the four sheets yourself with an LLM. Then spend 1–2 hours reviewing: appearance an image model can reproduce (face shape, brows and eyes, hair, build, one visual anchor), a conflict and hook in every episode, locations within limit. Do not skip this hour — a parsing error is paid again in every segment.

Reference art. Character sheets (white background, frontal full body, neutral light), scene plates (no people, explicit light source), key props: 20–40 images per series, 30–60 minutes of parallel generation, another hour to pick and reshoot. On the single-purpose route use Dreamina, Wan or any image model and keep a folder.

Total: about half a working day. Lock style, 9:16 and every character look on day one; changing them later means starting over.

The daily loop: seven steps and their minutes

Seven steps in a fixed order:

1. Storyboard review (20–30 min; batch weekly). Tools draft a storyboard in 1–3 minutes per episode. Check each segment ≤15 s with 3–6 shots, dialogue timed at about 2.5 words per second, only the people in frame named, lighting written as a source not a landing spot. Text edits are free; regenerating footage is full price, so review comes first.
2. Segment generation (40–60 min unattended, 10 min hands-on). One segment is one model call of 3–10 minutes; eight in parallel return in 15–20 minutes, 50–70% usable on the first pass, then one or two top-up rounds. Concurrency is capped per account (SceneMixer, for example, allows 3–15 tasks depending on tier); Kling, Hailuo, Pika and Veo via Flow cap it too. Use the wait to make the next episode's cover.
3. Picking (20–30 min). Walk the QC list: face, wardrobe and set continuity with the previous segment, lip-sync on the right speaker, intra-segment cuts executed. Fix the storyboard text and regenerate; never rerun a failed prompt unchanged.
4. Voice (5–10 min). Match each character to a preset voice by gender, age and language once per series. CapCut's text-to-speech works on the single-purpose route, with less distinction between characters.
5. Compositing (under a minute, 5 min check). One click on an end-to-end platform; CapCut to concatenate and subtitle otherwise.
6. Cover and title (15–20 min). The cover is the emotional peak frame; the title states the conflict without the spoiler; keep two variants.
7. Publish (10–15 min). Same time slot every day, after the checklist.

Hands-on total: 1.5–2.5 hours. One episode a day is sustainable, two on a push day; beyond that, review quality slips.

Cost per episode and what to batch

Video generation is over 85% of series cost. At September-2026 platform credit prices a 10-episode, 90-second series runs about $60–110 at 720p and $130–230 at 1080p, so $6–11 or $13–23 per episode including one or two regeneration rounds (typical, from SinCoSphere editorial testing). Upstream list prices set the floor: MiniMax Hailuo H3 is $0.08/s at 768p and $0.13/s at 2K; Wan 3.0 on Alibaba's Model Studio is ¥0.6/s at 720p and ¥1.2/s at 1080p; Seedance 2.0 is reported at about ¥1/s.

For contrast, invideo's own case study claims $1,000 per episode with three people over three days (vendor claim, July 2026) — a labor-inclusive number; ours is generation credits only.

Batch: parsing and art (once per series), storyboard review (5–7 episodes every Sunday), cover templates, voice matching. Do not batch generation and picking — a day later you have forgotten the continuity state of the previous segment and errors slip through.

Asset reuse: one cast library, one scene library

Reuse across episodes is what makes a daily cadence possible. Cast, scene and prop libraries are built once after parsing and every storyboard references the same entries: the same character sheet and face crop travel with every segment, the same plate with every scene. Never edit library images per episode; when the story needs a costume change, an injury or a new haircut, add a new look tagged "from episode N" and leave the original alone.

On end-to-end platforms the library is bound to the project and references attach automatically. On the single-purpose route keep a folder and a naming rule (character_look_view) and upload references by hand every time. Reference limits differ by model (Hailuo H3 takes up to 9 images per call and charges beyond 5), so in crowded frames give priority to the leads. From the second series on, generic plates — hospital corridor, office lobby, night street — carry over and halve the art day.

A sample week: five episodes plus a bank

About 15–22 hands-on hours a week for five episodes plus one or two banked (usable Sunday test segments go straight into the bank). Below three banked episodes, Saturday becomes a production day. For a ReelShort-style launch, run this week four to six times before publishing anything.

Pre-publish QC list and the US legal notes

Before every upload:

1. Lead's face, wardrobe and hair consistent across segments; set dressing does not jump.
2. Lip-sync belongs to the speaker; nobody off-screen talks.
3. Subtitles match dialogue word for word; no repeated lines.
4. No stray extras, extra hands or fingers, warped props.
5. No platform watermark, third-party logo or recognizable real person.
6. 60–120 s; hook in the first 3 s; cliffhanger in the last 3–5 s.
7. Levels consistent between segments; no clipping or dead air.
8. AI-generated content flagged where the platform asks (YouTube's upload flow asks about realistic synthetic content).
9. Title and cover match the content and contain nothing the channel bans.
10. Rights file on hand: license or public-domain note.

Two US notes. The Copyright Office's Part 2 report (29 January 2025) says prompts alone do not make you the author of AI output; the human-written script, your arrangement of segments into an episode and your edits can be registered — disclose the AI-generated material when you do. Second, never use a real person's photo or voice as a reference; right-of-publicity claims are state law.

Key figures and sources

Sources

• https://9to5google.com/2024/10/03/youtube-shorts-3-minutes/
• https://www.copyright.gov/ai/
• https://platform.minimax.io/docs/guides/pricing-paygo
• https://help.aliyun.com/zh/model-studio/model-pricing
• https://finance.sina.com.cn/tech/roll/2026-03-04/doc-inhpvpqk3838479.shtml
• https://invideo.io/blog/how-to-make-ai-micro-drama/

Tools mentioned

Read next