QWE AI Academy Why does every vertical drama feel like it ends right when you need the next swipe?
That’s the question that actually matters. A vertical drama isn’t a tiny TV show. It’s a retention machine for a phone held upright: 9:16, episodes usually 60-120 seconds, seasons often 60-100 chapters. Each cut lands on a question so the thumb stays put.
Key takeaway: Stop asking ChatGPT for a “screenplay.” Force a beat engine – cold open in the first 3 seconds, one clean emotional turn, cut on the unanswered question – then nest those questions so episode N answers the last hook and opens a bigger one. The format pays for that structure, not pretty prose.
Vertical drama in one tight background hit
China built the pattern as duanju. Western catalogs people name most: ReelShort, DramaBox, and peers. Pulp on purpose – secret identity, revenge, forbidden romance, sudden wealth. Close-ups win. Wide establishing shots die. Per Character App’s format breakdown and write-ups like The Story Mill, every creative choice exists to stop the swipe.
Money side, as of 2026 industry write-ups: first free block often sits in the 5-15 episode band, then coin purchases commonly discussed around $0.20-$0.50 per episode. Many pipelines put the hard commercial squeeze near episode 10. Category size estimates cluster near $14B for 2026 depending on how analysts draw the box – treat any single figure as a snapshot, not a law.
Live-action US runs still land roughly $150k-$300k and wrap principal in about 7-10 days. LLM-first scripting plus AI render pipelines showed up fast for independents because of that math. Writing is the cheap control point.
Method A vs Method B: full screenplay prompt vs beat engine
Method A is the first paste most people try: “Write a 70-episode vertical drama screenplay about a fake marriage to a billionaire.” You get act structure, scene headings, laptop-friendly dialogue. On a phone feed it stalls. Too much walk-up. Soft middles. Endings that resolve when they should wound.
Method B never asks for the full script first. Each episode is five timed beats aimed at ~60 seconds – tight end of the industry range, practical for AI-friendly length. You lock a season spine of nested questions, then expand one episode at a time with hard limits: no establishing shot, one location preference, 1-2 short dialogue lines, emotion only if the camera can see it.
| Dimension | Method A (screenplay dump) | Method B (beat engine) |
|---|---|---|
| Opening | Often setup/arrival | Forced mid-conflict 0-3s |
| Turns per ep | Multiple or muddy | Exactly one |
| Ending | Soft or resolved | Cut before answer |
| Paywall awareness | None | Ep 1-10 built for conversion |
| LLM reliability | Drifts by ep 20+ | One-page question map first |
Method B wins for beginners and for anyone shipping to vertical drama apps or Shorts. Faster to revise. Cheaper to generate later. Matches how the apps actually keep users.
Walkthrough: build a vertical drama spine in ChatGPT
Open ChatGPT (Claude works the same way). Skip episode 1 dialogue. Build the addiction loop first.
1. Season question map (do this once)
Act as a vertical drama showrunner for phone 9:16 series.
Genre: [e.g. revenge romance, PG-13].
Lead: [one-sentence visual tell - age, signature clothing, one mark].
Antagonist: [binary conflict in four words].
Output ONLY a table for episodes 1-20:
Ep | Cold-open image (first 3s) | Single turn | Cliffhanger question that forces the next episode | Free or paywall?
Rules: Ep 1-10 = free block, escalate hard by ep 8-10. Every cliffhanger is answered in the NEXT cold open, then a bigger question opens. No two consecutive same-type buttons (not two reveals in a row). One location family max (office/apartment/car). ~60s target later.Save the table. That is the bible. Episodes 21-80 only extend the same nested pattern. Skip it and long-context models still lose character goals and visual tells by the back half.
2. Expand one episode into shootable beats
Using ep [N] from the table, expand to this exact skeleton. Runtime target ~60s (~120 words total action+dialogue).
BEAT 1 COLD OPEN 0:00-0:03: mid-conflict or reveal. Zero setup.
BEAT 2 SETUP 0:03-0:15: one location, who wants what.
BEAT 3 TURN 0:15-0:40: secret lands / status flips.
BEAT 4 ESCALATION 0:40-0:55: consequence hits.
BEAT 5 CLIFFHANGER 0:55-0:60: cut on the question, one beat before answer.
Format each beat as:
LOCATION:
CAST:
ACTION: (present tense, what camera sees, externalize emotion)
CAMERA: (tight close-up / slow push-in - vertical only)
DIALOGUE: (max 2 short lines)
No monologues. No wide shots. No "she feels betrayed" - show the jaw tighten.Same skeleton as the one in Character App’s vertical drama script template. Fill brackets. You get something an AI director or a phone shoot can run.
Pro tip: After the draft, one more pass – “Delete any sentence that could go without weakening the hook or the button.” Vertical drama punishes connective tissue.
Episode 1 is special: put the season’s most cinematic image in the cold open – the slap, the contract, the wrong face in the wedding photo. First 15 seconds work like a built-in trailer. You’re training the thumb to stay.
3. Paywall episode check
Before ep 11: does the free block end on the highest open loop so far, and does ep 11 cold-open answer it then raise stakes immediately? No → rewrite the map. Conversion is a writing decision, not only a platform toggle.
Sometimes the cleanest draft still feels like a soap you wouldn’t admit you watch. That’s not a bug. The format sells heightened emotion at phone distance. Subtlety is a different medium.
Edge cases that break vertical drama scripts
- Setup addiction: Models default to “Maya walks into the lobby.” Kill it. If beat 1 isn’t already the crisis, three seconds and they’re gone. Industry templates ban establishing grammar in 0-3s for a reason.
- Soft free block: Eps 1-10 that only tease leave no reason to buy coins near the usual paywall zone. Too brutal too early also wrecks the free-to-paid funnel – escalate, don’t nuke the lead in ep 2.
- Question map missing: Eighty episodes with no one-page nested-question list → tone drift, forgotten scars, repeated button types. Lock the map before batch generation.
- Word count bloat: Aim near 120 words of action+dialogue for a ~60s cut. Longer pages force rushed delivery or messy multi-clip AI renders.
- Trope clone fatigue: Genre promise can stay. Change the trackable object or scar so the title doesn’t read like a thousand others.
Coin prices and free-episode counts aren’t universal – they shift by app, region, and promo. Write for the pattern (free hook → paid continuation), not one dollar figure from a single thread.
FAQ
How long should one vertical drama episode script be?
About 60 seconds on screen. Roughly 120 words of action plus dialogue combined. Shorter hooks harder.
Can ChatGPT write a whole 80-episode season in one go?
It can dump text. Coherence won’t survive. Map 20 episodes of nested questions first, expand in batches of 5-10, and re-paste visual tags plus open loops every batch. Treat the model like a fast junior writer who forgets the bible the moment you stop handing it back.
Do I need film gear after the script?
No. For a lot of freelancers the script is the product – rough reported band $40-$100 per finished minute as of 2026 snapshots, highly variable. Others push the same beats into AI video with locked character references, or shoot phone-native on tight close-ups. Beat sheet stays identical either way. Catalog apps still want a pitch/greenlight path; TikTok or Shorts are open-feed gates. Different doors. Same retention grammar.
Paste the season-table prompt with your own four-word conflict. Generate episodes 1-10 before any video tool. Ship the free block first. That’s the only test that counts.