How to Make an AI Micro Drama: The Full Workflow, Script to Upload

Pick a trope, write a one-page series plan, build your cast images, then make each episode shot by shot. That's the whole job. The tools change every month, but these steps don't.

A micro drama is a feature-length story cut into one- to two-minute vertical episodes. Each episode ends on a hook that makes you tap "next." Spirit Studios described the format to Deadline as roughly 60 episodes of 90 seconds, "building in 60 cliffhangers."

You used to need a cast, a crew and a location. Now one person with a laptop can make a convincing pilot episode in a weekend. Here's how, step by step, with links to our deeper guides.

What you'll need

• A video model that makes 9:16 (vertical) video. Ideally one with native audio, meaning it makes the voices too, already lip-synced. See choosing a video model.
• An image model for character sheets and keyframes. A keyframe is the still image you'll turn into a clip.
• An LLM (any chat assistant, like ChatGPT, Claude or Gemini) for outlines and shot lists.
• An editor that exports vertical video with captions. CapCut and DaVinci Resolve both have free versions.
• A budget. Our cost breakdown puts one 90-second episode at roughly $19 to $110, just for video. It depends on the model and how many takes you throw away.

Step 1: Pick a trope with an engine

Pick a setup your audience already searches for, then add one twist of your own. Vertical drama runs on familiar tropes, and that's fine.

Think of the CEO and the new assistant, the betrayed wife who comes back, the rejected mate or the heiress nobody recognizes.

Then test it. Can the premise give you a new cliffhanger every 90 seconds, for 60 or more episodes?

"She's secretly the majority owner" gives you dozens of reveals. "Two people fall in love" doesn't, on its own.

Our script templates cover ten genres. Each has a beat sheet and a sample episode 1, including revenge, werewolf and rich heiress in disguise.

Step 2: Write the series bible and the outline

A series bible is one page that pins down your show. Write it before you prompt anything. Put in:

• the logline (the story in one sentence) and the big secret
• 4 to 6 main characters, each with a one-line want
• a few locations you'll reuse (fewer locations means fewer things to keep looking the same)
• the episode count and target length

Then outline every episode in a line or two. A typical shape starts with a hook in episodes 1 to 3 and rising trouble. Then comes a big twist right before the paywall, a midpoint reversal, a dark moment and the payoff.

The paywall is where free episodes stop. A page on ReelShort's own site says the first 5 to 12 are free, and a UK producer we watched said 10 is common.

So whatever happens around episode 10 has to be strong enough to make someone pay.

An LLM is useful here, as long as you push back. Ask for five versions of the episode 10 cliffhanger and pick one. Then have it rewrite the episodes that lead into it.

Step 3: Script the first ten episodes properly

Put most of your effort into the first ten episodes. They get the most attention from viewers and platforms.

Bethany Thompson of Sea Star Productions said in an interview that her team gets most of its feedback on the first 10 episodes. That's also where they make most of their changes. A Scriptnotes listener on John August's site described the first ten as "pored over with repeated rewrites."

These rules make episodes easier to generate later:

• Open on an image, not a line. Writer Gigi M. Green says you have three seconds to stop the scroll, so write the opening frame first.
• Keep dialogue lean. Green suggests about ten lines per episode, at most.
• One action per shot. "She slaps him and storms out while the guests gasp" is three shots, not one.
• Write the cliffhanger first, then work backwards to it.

The full format, with screenplay examples, is in writing a micro drama script.

Step 4: Build your cast before you make any video

Drift (a face slowly changing from shot to shot) is what makes AI dramas look cheap. Fix it before you start:

1. Write one fixed description per character: age, hair, face and signature outfit. Paste it word for word into every prompt they're in.
2. Make a character sheet per person with your image model. That's the face from the front, three-quarter and side, plus a full-body shot in their main outfit.
3. Pick the best version and never regenerate it. That image is the character now.
4. Load the sheets into your video model's reference feature. Use reference images on Seedance and Veo, and Elements on Kling. Use subject references on MiniMax H3, and image references or a trained LoRA on Wan.

A LoRA is a small add-on you train on one face. In Google Flow, creators save each character once and tag them into every scene. The Uhura Creators Academy walkthrough shows voices saved with them too.

The character consistency guide goes deeper, including how to keep it up across 60+ episodes.

Step 5: Lock your locations

Do the same for your three to five main sets. Make one master image per location at 9:16, like the office, the mansion hall or the forest clearing.

Reuse that image as a reference or a first frame every time. A room that stays the same makes small face changes less noticeable, and the reverse is also true.

Step 6: Turn each episode into a shot list

A 90-second episode is usually about 10 to 15 shots. For each shot, write down:

• the shot type (close-up, medium, wide, over-the-shoulder, tracking, establishing or insert)
• how long it is, matched to your model's clip limit
• who's in it and what changes
• the exact line of dialogue, if there is one

Our script templates include a shot table for episode 1 you can copy. Vertical video has its own rules: more close-ups, less side-to-side movement, and room for captions. See vertical framing and shot language.

Some tools turn a script into a shot list for you. See storyboarding with AI and the storyboard projects in our directory.

Step 7: Make a keyframe per shot

Make one still image for each shot, with your cast and location references attached. Then animate those stills with image-to-video, where you give the model a picture and it brings it to life.

Kai Prompt's revenge-episode walkthrough works exactly this way. Every scene gets an image prompt and a video prompt, made as a pair.

A rejected still costs much less than a rejected clip. So get the framing, faces and outfits right at this stage.

Step 8: Generate the clips

Now animate. These habits save money:

• Match the clip length to the model. Veo 3.1 makes 4, 6 or 8-second clips (8 when you use reference images or 1080p). Kling 3.0 makes 3 to 15 seconds (up to 6 shots in one go), and Seedance 2.5 and Wan 3.0 go up to 30.
• Use multi-shot prompts where the model supports them. Write "Shot 1, Shot 2, Shot 3" in one prompt, and you get several angles of a scene with the same faces. ByteDance's Seedance guide and Kling's 3.0 guide both explain this.
• Put one speaker in each clip. When two people talk in one clip, lines get swapped. Cut between them instead.
• Redo only what's broken. In Isa does AI's walkthrough, three of five clips came back wrong, and only those three were redone.

Each model guide has its exact settings: Seedance, Kling, Veo, Hailuo, Wan. Ready-made prompts for each genre are in the prompt library.

Step 9: Voices and lip sync

If your model makes dialogue itself, you're mostly done. Just check that each character's voice stays the same from clip to clip, because it can change.

Using a silent model, or dubbing into other languages? Then make the lines with a TTS tool (text-to-speech, which turns written lines into a voice). After that, run a lip-sync pass on the dialogue shots so the mouths match the words.

The pros and cons, plus the consent rules for cloning a voice, are in lip sync and dubbing.

Step 10: Edit, score and caption

Put the clips in order, then tighten. Vertical drama cuts fast, so trim any dead air at the start and end of each clip.

When you chain extended clips, ByteDance's Seedance guide suggests cutting a few frames at each join to hide the seam. Then add a music bed and a short sting on the cliffhanger. Add room tone (quiet background sound) too, so the cuts don't jump out.

Last, burn in captions, because a lot of people watch with the sound off. See editing and captions and music and sound.

Step 11: Check it, then publish

Watch the whole episode on a phone, once loud and once muted. Check these:

• Does each face match its character sheet in every shot?
• Do clothes, props and injuries carry over between shots?
• Is there any garbled text on signs, screens or papers?
• Does the cliffhanger land in the last few seconds?

Then publish. Publishing and monetizing covers where you can post AI drama, YouTube and TikTok labeling rules, and what apps want from outside creators.

Read AI content copyright basics before you sell anything.

Shortcut: all-in-one agents

Some tools now run several of these steps for you. In the tutorials we watched, OpenArt's Director built a cast, a story and a 90-second episode from one chat. Google Flow has an agent that drafts the story and storyboard.

Kling launched a Canvas Agent for multi-shot storyboards in January 2026. Open-source drama platforms are listed under platforms, and agent skills under skills.

They're fast, but the story is still your job. In every walkthrough, the creator steered each step and redid shots that didn't work. Treat the agent as your crew, not your writer.

A realistic first weekend

• Day 1: premise, bible, outline for 10 episodes, the episode 1 script, character and location sheets.
• Day 2: shot list, keyframes, clips, edit, captions, export.

Post episode 1 and see how people react. Then write episodes 2 to 10 with what you learned. That's how the apps build series too: the opening episodes get rewritten the most.