​
Background

When demonstrating CLI tools like pgschema, asciinema is a good way to record the demo. I started by creating the recording manually.
However, even though I had prepared and practiced, the manual recording still included typos and pauses. Not a smooth experience. Then it occurred to me that I could use AI to script the demo.

​
Process

Manus is a general AI agent. I have been using it for various research tasks. I watch it spin up browsers, click links, and extract information. Scripting a terminal demo is a natural extension of its capabilities.

​
1st Recording: Too fast

I started by giving Manus this instruction. 
Read https://www.pgschema.com/.

Use asciinema to record a cast to demonstrate. Dump, edit, plan, apply workflow
Since pgschema is new, I needed to explicitly provide the URL. On the other hand, I assumed asciinema is well-known and doesn’t need further introduction. It was doing the right thing, but too fast—the demo finished in the blink of an eye. Well, it’s AI after all.

​
2nd Recording: Add delays

I then gently asked Manus to simulate human typing behavior: 
You are doing it too fast. Please simulate human typing behavior.
This way the cast is more realistic and people can follow along.
There were some problems along the way—for example, at one point it didn’t generate the cast file as instructed at all. After some additional tweaking, it was finally on track.

​
3rd Recording: Polish

There were still some rough edges. In particular, the demo added a CREATE TABLE change, but to showcase the declarative approach, it would be better to demonstrate an ALTER TABLE change—essentially comparing two CREATE TABLE statements and generating the ALTER TABLE migration. This is my prompt: 
This is too short, we still need to:

1. After dumping, show the initial schema
2. Show exactly what we change in the schema
3. Apply, then confirm with a small pause so it's visible to the viewer
4. Dump and show the schema after the change

Also, to show the declarative approach, we should add a column instead of
creating a table. Creating a table can't demonstrate the difference of the
declarative approach.

​
Found a Bug

Now Manus generated a near-perfect demo with:
  1. Clear workflow with all requested steps
  2. Proper pauses
  3. Color highlighting
  4. Readable text
The imperfect part wasn’t caused by Manus, but my pgschema. _ The generated dump didn’t end with a newline. This was obvious from the playback.

​
Final Recording

After I fixed the bug, I asked Manus to generate the final version. 
This pgschema version has a bug where the dump doesn't end with a newline.
I have fixed the issue. Please force pull the latest pgschema and rerun
the same script.
This time, Manus delivered a clean demo as shown below.

​
Reflection

I have made the entire Manus session public and you can check it here. Manus didn’t just execute commands—it interpreted my intent, made mistakes, course-corrected, and even helped me discover a bug in my own software. Fittingly, the first feature request for pgschema was to make it more AI coding friendly: _ An AI-built codebase, demoed by an AI agent, for an audience that will likely include developers working alongside AI. Perhaps this is just a glimpse of how software development is evolving.