Making an AI Short Drama Solo, Part 1: Why I Started and How an Episode Comes Together

Behind the scenesOct 16, 2026 · 5 min read

Why a programmer started making a drama series

I write code for a living, and a programmer's day is full of material. The "small change" request on its eighth revision. The pager going off at 3 a.m. The AI-written code that is more confident than any human engineer. Every time I traded these stories with coworkers, I thought: this would make a great one-minute sketch.

The problem was that I had no actors, no studio, and no idea how to animate. Once AI generation tools got good enough, I wanted to test one question: can one person, with an automated pipeline they built themselves, publish a drama series episode after episode on a steady schedule? The answer became 硅谷码农那些事儿 (roughly "Tales of a Silicon Valley Coder"), a Chinese-language series on Douyin, Xiaohongshu and WeChat Channels.

This series of posts shares what I learned along the way. It is not about which buttons to click in a particular tool. It is about the methods and the trade-offs: why each step works the way it does, and the dead ends I hit first.

How one episode gets made: five steps

Every episode goes through roughly five steps. Each of the following posts zooms in on one of them.

1. Script and storyboard. One joke per episode, under a minute long. The script is broken into beats, then into individual shots: who is on screen, who is speaking, and how close the camera is.
2. Voice. Every character has a fixed voice. All the dialogue is recorded first, and the real audio length decides how long each shot runs.
3. Picture. Each character gets a fixed "look" and each location gets an empty set. Then the characters are placed into the set and the shot is generated, moving and talking.
4. Editing and checks. The shots are cut together with subtitles, sound effects and an end card, then checked one by one: is there an extra person in the frame, is the mouth moving with the line, is the expression right?
5. Publishing and review. The episode is exported to each platform's specs and posted. A few days later I look at the numbers and use them to decide what to make next.

Only one of those five steps, the picture, is really "AI generation." The rest are either classic film production work or rules I added to keep the AI under control.

What dozens of episodes taught me

AI is one step; the unglamorous work is most of it

At first I thought the hard part was choosing a good model. It turned out the model sets the ceiling, but the floor is set by the plain, careful work around it: Is the script broken down clearly? Do the timings line up? Did someone actually look at every shot? When I tracked the time on one episode, the first full pass took only a small fraction of the total. The rest was rework.

Finish first, then get better

Early on I spent hours polishing single shots until they were perfect, and as a result I couldn't ship even one episode a week. So I flipped it: make the whole episode the simplest possible way, publish it, then upgrade the pipeline one round at a time. The show's visual style evolved exactly like that, from "moving comics" to partly real video to today's 3D widescreen. Every upgrade happened because the previous version got stuck on a specific problem.

Write down every mistake

AI tools have a habit: if you don't write a mistake down, they will make it again, and you will forget it too. Now every pitfall I hit becomes a written rule, for me and for the AI assistant that helps me work. That ever-growing list is the most valuable part of the whole pipeline. The last post in this series is about it.

Next up

Part 2 starts at the beginning: script and storyboard. How to open a one-minute episode, what to cut, how to break a scene into shots, and why my scripts go through automated checks before anything is generated.