Making an AI Short Drama Solo, Part 2: Script and Storyboard, One Joke per Minute

Behind the scenesOct 18, 2026 · 6 min read

One minute, one joke

The single most useful rule I have is also the simplest: one episode serves one joke. Not a joke and a side gag, not a joke and a little lesson at the end. Every shot exists to build toward that one punchline, and if I can't say how a shot helps, it goes.

The whole episode has to fit in under a minute. Early on I measured that against the durations I had planned for each shot, and I was wrong almost every time. Planned durations are wishes. The real length is set by the recorded dialogue, so now I record the voices first and measure the actual audio (more on that in the next post). If it runs long, I cut lines, not pauses.

How the script is written

Open on the conflict

People scroll fast. The first shot or two have to show a conflict or a contrast right away, and the very first line of dialogue has to carry it. No "Good morning, how was your weekend?" warm-up. If the episode is about a "tiny change" request on its eighth revision, the first thing you hear is the eighth revision.

Cut what the audience already guessed

A lot of my first drafts had scenes that acted out something the viewer had already figured out. The manager asks for a change, and then we watch the engineer slowly open the laptop, sigh and start typing. Everyone knew that was coming. Those transition scenes feel safe to write, but they are dead air on screen. Now I jump straight to the next surprise.

Nothing after the punchline

This one took me a while to accept. In one early episode I had a final shot where a character added one more line after the joke landed, a little "see what I mean?" When I watched the cut, that extra line made the joke weaker. It explained something the audience had just enjoyed figuring out on their own. I cut the shot, and the episode got better by getting shorter. Now the rule is written down: once the punchline lands, the story is over. The end card can follow, but no more dialogue.

No silent shots

Every shot has a line or a narration. A silent establishing shot feels cinematic when you write it, but in a one-minute sketch it is three seconds of nothing, and that is where people swipe away. The only exception is a short reaction right after the punchline, which is part of the joke, not a pause before it.

Before I move on, I go through the script shot by shot and ask one question for each: what does this shot do for the joke? If I can't answer in one sentence, it is cut.

From script to shots

A script is not something a generator can work from. It needs to be broken down, and I do that in three layers.

1. Beat sheet. The script is split into beats: the smallest units of story, usually one line or one reaction each. This is where the lines get locked.
2. Storyboard plan. For each scene, a plan of the camera setups: the wide master angle, the reverse angle, the close-ups. This decides where the cameras stand before anyone is drawn.
3. Shot list. Each beat is assigned to exactly one shot, in script order. No beat is left out, no beat appears twice. If someone edits a line later, the shots that claimed it are flagged as out of date.

The rules every shot list follows

Most of these came from mistakes that cost me a full regeneration.

• The speaker is on screen. A line spoken by someone who isn't in the frame confuses the viewer and confuses the lip sync.
• One speaker per segment. If two people talk, that is two segments, cut together.
• At most two people in a shot. Three or more is where generators start cloning faces and inventing extra strangers.
• Introduce people wide. When a new character appears, at least one of their first shots is a medium shot or wider, so the audience knows who they are and where they stand.
• Big movements happen across a cut. Standing up, turning around, moving a hand from head to desk: these look bad when generated inside one shot. I end one shot before the movement and start the next one after it, the way editors have always done.
• No text from the generator. Anything that has to be read on a screen or on paper is added afterwards in editing. Generated text is always gibberish.

Basic shot grammar

I am not a trained director, so I borrowed the basics and turned them into checks. A scene should use a few different shot sizes, not the same framing over and over. Don't stack several shots of the same size in a row. Don't jump straight from a wide shot to an extreme close-up. After the punchline, cut to a reaction or a close-up so the joke has a place to land. And a scene with many shots needs more than one real camera angle. One wide image cropped into pieces is not multiple cameras, and viewers can feel it even if they can't name it.

Describe what you want, not what you don't

When writing the description for each shot, I only use positive statements. Image generators tend to ignore "no" and "without," and sometimes do the opposite: mention a crowd, even to forbid one, and a crowd appears. I treat any negative word in a shot description as a bug. Instead of saying the room is empty of other people, I describe exactly who is in the room and what fills the space around them.

Keep it small, then lock it

I also cap how complicated an episode can be. Usually it is one location and a handful of shots, with only a couple of shots that have two people in them. The rest are single-person shots and screen inserts. The important part: I don't cut jokes or lines to meet the cap. I cut the shots that are hard to film and find a simpler way to show the same thing.

Then the script gets locked. I read it one last time and approve it before anything is generated. When I once changed a scene halfway through production, it meant redoing images, video and voice for every shot it touched. A change on paper takes a minute. The same change after generation can take an afternoon.

Finally, the shot list goes through an automated checker before generation starts. It knows all the rules above. Hard errors, like a speaker who isn't on screen or too many people in one frame, block the run until I fix them. Warnings, like the episode running a little long, get a look and a decision. It takes seconds, and it catches the kind of mistakes that would otherwise show up an hour later as a broken shot.

Next up

Part 3 is about voice: why I record every line before generating any picture, how each character keeps the same voice across dozens of episodes, and how the voice track drives the lip sync.