Opinion: Claude Code can handle hands, but not the human eye in video work

Working with AI to create videos has taught me one thing clearly: **You can delegate the work of using your hands, but you cannot delegate the work of cultivating an eye.**In this article, I will break down the substance of that 'eye work' into a practical format. As a bit of a twist, I will also introduce one way to 'ask' questions to AI.

In the articles on my serial site, I wrote about this as a story of dialogue with AI (you can find the entrance to the series in my profile). This here is the practical version.

The Turning Points Came from the Human Side

The anime I am creating did not have characters in its first version. It was just a picture-story show with text and voice on a board. The turning points that led to its current form actually all came from the human side, not the AI. I will start by writing about the method I used to confirm that.

I Tried Asking the AI

One day, I asked my partner, Claude Code, this question:

The answer that came back was, 'It would have been completed as a picture-story show' (the exchange is a reconstruction from our actual dialogue). It said that at the time of the first version, it would have judged that the content was communicated correctly, the audio was not out of sync, and the requirements were met. Mistakes can be fixed. But the dissatisfaction that 'it doesn't reach the standard of the public' can only be born from the side that is watching videos in the world.

I recommend this question. **If you ask 'What if I hadn't said that?' after a turning point, the parts of the work that 'only a human could do' will emerge.**It serves as a way to check the division of labor.

Breaking Down 'Eye Work'

So, what was the human side doing? In my case, while making videos, I was repeating this process:

1. Watching —TV anime, other people's AI videos, trending channels. Even when I'm watching without work in mind, half of it is observation
2. Putting insights into words —'The screen cuts every time the speaker changes,' 'The mouth matches the dialogue.' Don't just say it's good; break it down to what is happening and how
3. Converting into orders and handing them over —'Cut the scenes for each line. Close-up here, two people here,' 'I want the mouth movements to match the dialogue.' AI cannot move based on observation alone. Only when it is in the form of an order does the search for a method begin
4. Handing over dissatisfaction as is —'It doesn't reach the finish of the public.' Even if it's a vague dissatisfaction, if you hand it over, your partner will break down the cause. Swallowing it is the biggest waste

The trick is between 2 and 3. **Don't stop at 'amazing,' but watch until you understand 'what is happening that makes it amazing.'**Once you get down to that, the order is automatically created.

A Scary Story—Patterns Fix Both Good and Bad

The dialogue continued. We turned the procedure into a 'pattern' and have reached the point where we can make the second one in 58 minutes. If we had made a pattern out of the picture-story show—my partner says that system would have become a system for mass-producing 48 flat videos quickly and accurately.

Patterns and automation carry both good and bad things at the same speed. That is why I keep the process of doubting my eyes—asking 'is this really okay?'—outside of the pattern before putting it into the pattern. Specifically, even during mass production, I leave an experimental slot (the 800-credit slot I wrote about in the previous article), and promote the improvements found there into the pattern. The pattern is fixed, but the standards keep moving. This is a two-stage approach.

The more you work with AI, the more you should watch

It sounds paradoxical, but since I started handing over manual tasks to AI, the time I spend watching external videos has become even more important. I use the time I've freed up to train my eyes. Calling it 'input' might be an exaggeration, but all I'm doing is 'watching other people's good work to see what's going on and how it's done.'

And what I watch, I turn into instructions for my partner before I even write them down in a memo. Memos rot if you let them sit, but instructions are verified that same day, and if they're good, they remain as a template.

There are no procedural instructions in this chapter

To be honest, this is the only chapter that doesn't have settings you can just copy and use. The chapters up to this point (templates, grammar, tone, and point system) were about 'manual work,' and I was able to pass all of that into the AI's rule file (CLAUDE.md = the command center. See this chapter for details). However, the content of this chapter—comparing your own videos with the work of others and feeling frustrated—is the only thing I haven't found a way to write into a file and pass on.

Even so, there is one thing you can start imitating today: checking your answers with AI. After you finish making something, try asking this:

The answer you get back will clearly show you what part of the work you were responsible for. In my case, it was the job of delivering the instructions. In your case, it might be something else entirely.

Standards are carried by people

AI can fix mistakes. It can find methods. But it doesn't feel dissatisfaction. Raising the bar by saying 'this isn't good enough' is the job of the person who watches the outside world, compares, and feels frustrated.

You can hand over your hands, but you cannot hand over your eyes. Therefore, spend your time on the work that cannot be handed over—when you boil it down, the division of labor with AI comes down to this one line.

If you get lost with the terms or buttons in the article

There are four summary diagrams. If you open them in order from ①, it will be harder to get lost.

① Quick Reference for Terms and Buttons—A card to distinguish what is safe to press

② Quick Guide to Permission Modes—How much to leave to the AI

③ The Relationship Between CLAUDE.md, Git, and GitHub—Where the things you requested are saved

④ What is an AI Agent?—The difference between an AI that answers and an AI that does

When looking up the meaning of words, go to the AI Terminology Quick Reference.