I Made a 6-Episode Fantasy Short Film Using ChatGPT and Openseedance (And You Can Too)

 

Let me start with a confession.

I have never made a film in my life. Not a short. Not a music video. Not even a decent birthday montage.

I don't know how to operate a cinema camera. I couldn't tell you the difference between a key light and a fill light. The last time I touched video editing software, George W. Bush was still in office.

And yet, last Tuesday night, I sat on my worn-out IKEA couch, opened my laptop, drank a LaCroix that had gone flat, and "directed" a six-episode fantasy revenge series.

It has dragons. It has a fallen prince who gets a second chance. It has a villain who looks like he walked straight out of a Dark Souls fever dream.

The whole thing took me about four hours from blank page to finished clips.

Here's exactly how I did it — no gatekeeping, no fluff, and definitely no "subscribe to my course" nonsense at the end.

The Honest Truth About AI Video (That Nobody Tells You)

Before I walk you through the workflow, let me level with you.

AI video tools right now are like a gifted but unstable intern. They will absolutely blow your mind with what they can do. And five minutes later, they will hand you a frame where your hero has seven fingers and an eye floating outside their skull.

You cannot control this. You can only work around it.

The secret isn't finding the "perfect" AI tool. The secret is building a workflow that embraces the chaos — one where you generate a lot, keep the good stuff, and ruthlessly throw away the garbage.

This is that workflow.

Step 1: Don't Ask ChatGPT for "Ideas." Give It a Job Description.

Most people open ChatGPT and type something tragic like:

Give me an idea for a short film.

This is how you get boring, generic, been-done-a-million-times sludge.

ChatGPT is not a creative genius. It's a highly obedient mirror. It reflects exactly what you ask for. Ask for generic, get generic. Ask for specifics, get something actually usable.

Here's the prompt I now use. It turns ChatGPT into a battle-hardened short-form content strategist. Steal it, tweak it, make it yours.

```

You are a senior short-form content strategist who has produced multiple viral series on TikTok, YouTube Shorts, and Reels. You understand retention graphs, hook density, and AI production constraints better than anyone.

I need three original concepts for a 9:16 vertical AI-generated short series. Target is 5-10 million views per episode.

Mandatory rules you must follow for every concept:

- First 3 seconds: Unscrollable chaos. Explosion, shocking line of dialogue, or visual spectacle.

- Every 15-20 seconds: A twist, a fight beat, or an emotional payoff.

- No complex continuous action. Design for high-impact single shots instead.

- Every single episode ends on a cliffhanger that demands a "part 2??" comment.

- Characters must have fixed visual descriptions (hair, outfit, key accessories) for AI consistency.

- Avoid subtle facial expressions (crying, smirking, micro-emotions). AI still sucks at these.

Output each concept as: Title + Logline (one sentence) + Visual hook + Main character description.

```

Try this. I promise you'll see the difference immediately.

Step 2: Pick One Concept and Turn It Into a "Production Bible"

Last week, ChatGPT gave me three concepts. I picked the one that gave me the most visual excitement: *a disgraced prince gets sent back in time and systematically destroys everyone who betrayed him.*

Basic? Sure. But basic works on short-form platforms. You're not making Bergman's Persona. You're making something people watch while waiting for their microwave popcorn.

Here's what I asked ChatGPT to generate next. Think of this as your pre-production bible.

A. One-sentence logline

If you can't explain it in one sentence, it's too complicated for vertical video.

B. Character sheets for the hero, villain, and sidekick**

This is critical. You need fixed visual descriptions. Hair color. Outfit style. Distinguishing marks. Without this, your hero will look like three different people across five scenes. I learned this the hard way so you don't have to.

C. Three-act beat structure

But here's the twist — I had ChatGPT label each beat with "anticipated comment bait." Stuff like: *"End of Act 1 — audience will comment 'wait what??'" "End of Act 2 — audience will say 'he's cooking now.'"* This forces the AI to think like a retention-obsessed producer.

D. Text-to-image prompts for every main character

This is the part everyone skips. Then they wonder why their character's face keeps mutating. Do not skip this.

Step 3: Generate the Actual Script (Format Matters More Than You Think)

Now for the fun part.

I told ChatGPT:

Write episodes 1 through 3 in standard screenplay format. No tables. No bullet points. Keep dialogue short and punchy — no monologues longer than two sentences. Every scene must either advance the plot, deliver a payoff, or set up a cliffhanger.

The difference between "table format" and "screenplay format" is huge. Tables make the AI think like a spreadsheet. Screenplay format makes it think like a writer.

The dialogue won't be perfect. Sometimes it's a little stiff. Sometimes it's weirdly formal in a way no human would ever speak.

That's fine. You're not publishing this. You're using it as a blueprint. I spend maybe 10 minutes per episode punching up the dialogue. The structure and pacing are already solid.

Step 4: The Openseedance Part (Where the Magic Actually Happens)

Okay, real talk about Openseedance and Seedance 2.0.

I've tried almost every AI video tool on the market. Runway Gen-2 and Gen-3. Pika Labs. Kling. Luma Dream Machine. Some of them are genuinely impressive. Some of them are overhyped. Almost all of them share two problems: **slow renders** and character inconsistency.

Openseedance wasn't even on my radar until a friend sent me a DM saying "you need to try Seedance 2.0."

Here's what I found after two weeks of beating on it.

Speed: Legitimately fast.

I'm not talking "fast for an AI tool." I'm talking paste your prompt, hit generate, take one sip of coffee, and the clip is ready. That matters more than people realize. When you're iterating on prompts, speed is the difference between trying three variations and trying thirty.

Pricing: Actually reasonable.

There's a free tier. The paid tiers won't make you cry. I've seen tools charge $0.50 per second of video. Seedance 2.0 is nowhere near that. You can experiment freely without watching a credit counter like a hawk.

Character consistency: Better than most.

This is the hardest problem in AI video right now. Seedance 2.0 isn't perfect — no tool is — but it's noticeably better at respecting reference images than the competition. My main character's face stayed mostly the same across multiple clips. That's a win in my book.

The workflow that worked for me:

1. Copy my ChatGPT script directly into the Openseedance prompt box

2. Upload the reference images I generated in Step 2

3. Hit generate

4. Repeat 10-15 times per scene, keeping the 2-3 good takes

That last point is important. You're not going to get a perfect clip on the first try. You generate multiple versions, pick the winners, and move on. This is exactly how traditional animation works. AI just makes it faster and cheaper.

The Brutal Reality Check (Read This Before You Start)

I've been honest so far, so I'll keep going.

You will generate a lot of garbage.

I mean *a lot*. Weird extra limbs. Backgrounds that shift between cuts. Faces that melt and reform like something out of a David Cronenberg movie. Dialogue that gets randomly inserted into the background as floating text.

This is not a sign that you're doing something wrong. This is just where the technology is right now.

The workflow I just described doesn't eliminate the garbage. It increases your good-output rate from maybe 10% to 25-30%. And when you're working with a tool that's fast and cheap, 25-30% is more than enough to finish actual projects.

The trick is to treat AI generation like shooting on film. You shoot way more than you need. You throw away the bad frames. You keep the good ones. You move on.

What AI Is Actually Good At (And What You Should Avoid)

After burning through hundreds of dollars worth of credits across different tools, here's my honest tier list.

AI is genuinely excellent at:

- Massive, moody environments (ruined castles, neon-drenched cyberpunk cities, creepy forests)

- Characters with strong silhouettes and simple, repeatable costumes

- "Key moment" action shots — the sword swing, the explosion, the dramatic reveal

- Fantasy, sci-fi, period pieces, and light horror

AI is genuinely terrible at:

- Two people having a quiet conversation while the camera slowly zooms in (almost always fails)

- Complex hand-to-hand combat lasting more than 3 seconds

- Subtle facial expressions — crying, smirking, "a single tear rolls down the cheek"

- Any text on signs, logos, or screens

- Consistent object permanence (things disappearing and reappearing between frames)

Design your scenes around the first list. Avoid the second list like it owes you money.

What I'd Do Differently Next Time

I learned a few things from this project that aren't in the main workflow.

Start with the ending. Before you write anything else, know exactly what the final shot of your series looks like. Work backwards from there. This sounds obvious but it's surprisingly easy to forget.

Keep episodes under 90 seconds. The AI holds up better. Viewer retention is higher. Everything is easier.

Sound design is half the battle. AI video gets you 80% of the way there, but bad sound ruins everything. Spend real time on music and sound effects. It makes cheap AI footage feel expensive.

Build a small library of reusable environments. If your series takes place in a castle, generate 20 different angles of that castle upfront. Reuse them across episodes. This saves you from constantly reinventing the wheel.

The Bottom Line

One year ago, making even a three-minute narrative short meant begging friends for favors or spending money you didn't have on gear and crew.

Today, I can sit on my couch with a laptop, spend 30 minutes prompting ChatGPT, paste the result into Openseedance, and walk away with something that looks 80% of the way to a real animated series.

Is it perfect? No. Is it ready for a film festival? Absolutely not.

But for YouTube Shorts, TikTok, and Reels — the places where millions of people actually watch videos — it's more than good enough.

And here's the thing: the technology is getting better every month. The people who are learning these workflows right now, while the tools are still a little janky, are going to be way ahead when the tools get great.

You don't need to wait. You can start tonight.

P.S. — If you try this and make something cool, drop the link in the comments. I genuinely want to see what people build with this stuff.

P.P.S. — Openseedance has a free tier. Go beat on it before you spend a dime. That's what I did.