QWE AI Academy By the end of this guide, you’ll have a plan for producing a 90-second vertical short drama episode – the kind that runs on ReelShort or DramaBox – using AI tools that together cost less than a pizza. Not a movie. Not a TikTok skit. A serialized, cliffhanger-ended, 9:16 episode with consistent characters across scenes.
Most tutorials walk you forward from “pick an idea” to “click render.” That skips the part that actually matters – the tool math. We’re going backwards. Finished episode first, then the decisions that get you there.
What a finished episode actually looks like
90 seconds. Vertical 9:16. One lead character who looks identical in every shot. Three to five scene cuts, voiceover, burned-in captions, a hook in the first three seconds, a cliffhanger at the end. That’s the deliverable – everything upstream serves those constraints.
Short drama is not a short film. A short film is horizontal, standalone, built for a festival or YouTube premiere. A short drama episode is vertical, serialized, engineered for binge-watching on TikTok, Reels, and YouTube Shorts – and that distinction drives every tool decision you make. The format looks simple. The serialization constraint is where it gets hard.
Why bother? The money is real. As of 2025, short drama apps generated $2.98 billion in in-app purchase revenue, up 115% year-over-year (Sensor Tower via Filmustage). ReelShort users in the US spend an average of 35.7 minutes per day in the app – beating Netflix mobile at 24.8 minutes (Omdia Q4 2025). Attention is there. The question is whether your episode can hold it past the paywall.
The three-tool stack (working backwards from the render)
Three roles. That’s it. Script generator, video model with character consistency, captions and export layer. Most beginners buy an all-in-one platform and pay for features they never touch. The unbundled stack is cheaper.
| Role | Recommended tool | Why it fits short drama |
|---|---|---|
| Script + beats | ChatGPT / Claude | Free tier handles a 90-second beat sheet in one prompt |
| Video generation | Kling or Veo 3.1 | Both output native 9:16; character reference supported |
| Voice + captions | ElevenLabs + CapCut | Voiceover, then burned-in captions for sound-off viewing |
Video generation is the expensive part. Veo 3.1 costs $0.40/sec (Standard) or $0.15/sec (Fast), per Google’s January 2026 announcement. Kling 2.6 runs $0.07-$0.14/sec – roughly 56% lower than Veo Standard (Laozhang blog comparison). For a 90-second episode, that gap is roughly $10 vs. $36+ per generation attempt. And you will attempt multiple times.
Here’s a question worth sitting with before you commit to a tool: how many episodes do you actually need to produce before the format clicks for you? The answer shapes whether you optimize for cost (Kling) or consistency (Veo) from day one. If it’s fewer than five, start with Kling’s free tier and don’t overthink it.
Character consistency: where most first attempts fall apart
Character drift is the real budget killer – here’s why. Generate ten clips of the same character with the same prompt and you get ten slightly different faces. Eye color drifts. Jawline shifts. Hair changes length between scenes. The illusion breaks, and you’ve burned credits on unusable footage.
The fix isn’t a better prompt. It’s a reference image. Lock one front-facing, neutral-expression portrait of your lead before generating any video. Feed that same image into every clip. Veo 3.1’s January 2026 update added an “Ingredients to Video” feature – up to 4 reference images, improved consistency across shots. Google’s announcement walks through the reference-image workflow in detail.
Now the trap most tutorials miss: Sora 2 scores 4.1/5 on photorealism benchmarks (AI Workflows + OpenCreator). Looks stunning in a single shot. But it has no character reference input – meaning it cannot lock a face across episodes the way Veo and Kling can. Photorealism without consistency is a demo, not a serialized drama. Sora is the wrong primary tool for this format, full stop.
Test your reference image with 3-5 short clips in different lighting before generating any scenes. If the face drifts in the test, no amount of prompt engineering will fix the full episode – regenerate the reference first.
The step-by-step, in reverse order
Standard tutorials go script → storyboard → generate → edit. Work backwards from the paywall instead.
- Start with the cliffhanger. Write the final 5 seconds first – the reveal, the slap, the phone call. That’s what viewers pay to resolve in episode 2.
- Write the hook second. First 3 seconds must establish stakes. “My billionaire husband asked for a divorce” is a hook. “Sarah walked into the office” is not.
- Fill in 3 middle beats. Not scenes – beats. Each beat becomes one 5-10 second clip.
- Lock the reference image before generating anything. One portrait per character. Save it.
- Generate clips one at a time. Not in a batch. Watch each one before spending credits on the next.
- Assemble in CapCut. Add voiceover, captions, music bed. Export vertical.
No elaborate storyboard. For a 90-second episode with 3-5 clips, a written beat sheet is enough. Storyboards matter at 20-episode scale, not at experiment scale.
Common pitfalls (the ones that actually cost money)
Three specific traps. Each has a workaround.
Trap 1: Trusting Kling’s advertised clip length. Turns out Kling’s 2-minute ceiling has a catch. Per AI Tool Analysis, quality degrades noticeably after 30 seconds of extensions – character consistency weakens, lip-sync on multi-person dialogue falls apart. Kling technically supports 2-3 minutes and outpaces Runway (40-sec cap), Sora 2 (35-sec cap), and Veo 3.1 (60-sec cap). But for drama, that ceiling is irrelevant. Use Kling for 5-10 second individual clips, then assemble in the editor. Never let it run long.
Trap 2: Paying Veo Standard for every shot. $0.40/sec Standard vs. $0.15/sec Fast – per Veo 3.1’s official pricing page, that’s roughly an 8x cost difference on the low end per second compared to Fast tier. For establishing shots, transitions, and B-roll, Fast is indistinguishable from Standard. Reserve Standard for the shot with dialogue or the emotional climax. This alone cuts a 90-second episode’s generation cost by more than half.
Trap 3: Uploading horizontal and cropping later. Veo 3.1 outputs native 9:16 vertical – no crop, no quality loss (per the January 2026 announcement). If your tool doesn’t support vertical natively, switch tools before you switch workflows.
Kling vs Veo vs Sora – honest read for drama specifically
- Kling – cheapest per second ($0.07-$0.14), best for high-volume clip production, weakest on lip-sync. Use for voiceover narration scenes, not on-screen dialogue.
- Veo 3.1 – best character consistency via Ingredients to Video, native vertical, most expensive at Standard tier. Use for dialogue scenes and hero shots. Fast tier ($0.15/sec) works fine for everything else.
- Sora 2 – highest raw photorealism, no reference-image input. Use only for one-off establishing shots or dream sequences where the character doesn’t need to match across cuts.
For most beginners: Kling for 80% of clips, Veo Fast for the rest, skip Sora until you’re producing episodes that need a specific hyperreal moment. Total generation cost for a 90-second episode this way is roughly $8-$15 (illustrative estimate based on published per-second rates – your actual cost depends on how many retakes you need).
Where the money actually comes from
Making an episode is one thing. Getting paid is stranger than most creators expect.
On ReelShort and DramaBox, users pay $0.20-$0.50 to enable individual episodes after the first 5-8 free ones (Genra + Streaming Lens, 2025). Your revenue depends on episodes watched past the paywall – not raw views. A series with 100,000 views and a bad episode-6 hook earns less than a series with 20,000 views and a great one. Retention past episode 5 is the only KPI that matters here. View counts are a vanity metric for this format.
FAQ
Do I need to pay for multiple AI subscriptions?
Not to start. ChatGPT’s free plan handles script generation. Check Kling’s current free tier before committing – availability changes. A first episode at zero dollars is realistic. Paid plans matter once you’re generating 5+ episodes per week.
Can I sell my AI short drama to ReelShort or DramaBox directly?
Sort of – but the realistic path looks different from what most people expect. Both platforms commission originals, though they favor established creators or studios with a track record. A better first move: publish your first 10 episodes on TikTok or YouTube Shorts, build watch-time data, then approach smaller platforms with proof of retention. Why English-language content specifically? The US drove roughly 60% of global short-drama mobile revenue in 2024 (Sensor Tower) – the paying audience is largest there, so optimizing for American viewers first makes commercial sense before expanding.
How long does one episode actually take to produce?
First episode: 4-6 hours. Episode five: closer to two hours – once your reference-image library exists and your prompt template is repeatable, most of that time shrinks to clip review and assembly.
Your next move: open Kling or Veo, generate one 5-second clip of a single character from a reference image, and see how consistent the face stays across three variations. Don’t write a script yet. Prove the consistency works for your character first – everything else is downstream of that test.