# Story

Episode 6 is not about creating one long narrative film. It uses the 26-second “Mount Fuji and Its Reflection” answer version at the beginning of the full making-of as an entry point into a larger production activity: creating many short Aha and spot-the-difference trials, then carrying human feedback into the AI agent’s next batch.

## The text first given to the AI agent

The creator began with one short request. The following is an English translation of the original Japanese text, with its roughness intentionally preserved.

```text
I want to make Aha videos featuring Japanese landmarks, cute cats, and cute dogs.
```

The creator did not begin by supplying before-and-after images, internal change-region data, a storyboard, or a technical specification. The AI agent proposed subjects and changes, then produced the first trials.

## The production flow developed through discussion

1. Use the one-sentence request to propose Japanese landmarks, cultural scenes, cats, and dogs.
2. Create before-and-after images with the same composition and limit each work to one meaningful change.
3. Lock everything that must remain unchanged, then create a 22-second challenge with a two-second hold, an 18-second gradual change, and another two-second hold.
4. Add a four-second comparison showing the answer, producing a 26-second answer version.
5. Let the AI agent inspect intermediate frames, full decoding, duration, resolution, and frame rate.
6. Have the creator watch each trial at normal speed and return an OK or NG decision with a reason when possible.
7. Carry confirmed patterns involving change size, location, naturalness, and overlap with earlier ideas into the next batch.
8. Repeat trials and reviews, producing more than 50 candidates in half a day.
9. Keep technical validity separate from the human judgment of whether the result feels natural, challenging, and interesting.

## Opening example: Mount Fuji and Its Reflection

The finished making-of opens with Mount Fuji and its reflection in a lake. The overall scene stays calm while a change in the reflected mountain progresses over 18 seconds. The viewer first attempts the challenge without the answer, then sees the before-and-after comparison with the changed area identified.

This example shows the method in one finished video. The public production materials also cover the larger cycle behind more than 50 trials made with the same general approach.

## What changed through evaluation

- Moving or replacing a large object could produce a double image at the midpoint or make the answer immediately obvious.
- A change could also be too small to understand even after seeing the answer.
- A technically valid video could still be rejected for naturalness, difficulty, interest, or overlap with a previous idea.
- Keeping the large subject fixed and changing one less-attended surrounding detail worked better for this format.
- The AI agent did not invent rejection reasons that the creator had not stated.

The production record contains 57 evaluated videos: 32 accepted, 22 rejected, and three excluded because their ideas duplicated earlier work. The batches used different subjects and conditions, so they are not a controlled performance comparison. They do document a real process in which evaluation changed the criteria for later trials.

For a clear viewer-facing summary, the finished making-of explains that only two or three videos out of ten were initially accepted. After roughly ten rounds, eight out of ten were fully accepted, and the remaining two were usable with minor adjustments.

## What the person decided and what the AI agent did

The person first chose the broad subjects: Japanese landmarks, cute cats, and cute dogs. After each trial, the person judged acceptance, obviousness, overly small changes, and duplication with earlier ideas.

The AI agent proposed candidates, produced images, created gradual changes and video files, prepared intermediate-frame reviews, performed technical checks, and recorded evaluations. It avoided overgeneralizing one comment and added only confirmed patterns to later trial criteria.

The final acceptance decision remained human. Technical correctness alone did not make a video complete.

## The special rule used in this episode

The standard NijiUnit workflow asks a person to approve an Excel storyboard before video production. Episode 6 prioritized testing many short videos with one shared specification, so the creator approved a special rule that allowed direct video trials without a storyboard.

This does not mean that storyboards are always unnecessary. It demonstrates that an AI agent can adjust the workflow to the work instead of forcing every project into one fixed process.

## Wording caution

These videos invite viewers to notice a slow visual change and enjoy the moment when the answer becomes clear. The page must not claim that this is identical to the psychological Aha experience or promise cognitive, medical, or educational benefits.