[AI THINKING LAB – 6th Cohort]Diving into 6 AI Ideas at the “Implementation Level” (March 5, 2026)
This was the 4th session of AI THINKING LAB.
What stood out the most in this session was something simple but powerful:
AI becomes truly interesting when we move from ideas to implementation.
Some participants are trying to create videos.
Others are thinking about building AI-powered apps.
Some want to automate their content creation and workflows.
In this session, we explored a key question:
“How can we actually build these ideas?”
Next Milestone: Midterm Presentation on March 19
During the session, we confirmed our next milestone.
March 19 – Midterm Presentation (open/public session)
Each participant will present:
What they want to build
How they plan to build it
Who will benefit from it
AI THINKING LAB is now moving from the
“thinking phase”
to the
“showing and building phase.”
AI Character Video Project
An Idea That Became a Real Project
One participant shared a project idea for a character-based educational video series.
Following feedback from the previous session — “clarify the target audience” — the concept became much more concrete:
Target audience: preschool children
Series theme: everyday items (toothbrush, eraser, etc.)
Character design and storytelling structure
Educational narrative
At this point, the idea had clearly evolved from a simple concept into an actual project proposal.
The Biggest Challenge: Character Consistency
However, a major challenge emerged.
Maintaining consistent characters across multiple videos.
With generative AI tools, it’s surprisingly easy for a character to change slightly each time.
Even small variations in prompts can result in a completely different character.
For a series-based project, this becomes a critical issue.
Testing with Sora
We ran a quick experiment using Sora.
The process was simple:
Input an English script
Upload the character image
Ask the AI to create an animated historical scene
The result?
Yes — it can generate videos.
But the output was still rough.
The reason was clear:
The prompt wasn’t detailed enough.
To improve quality, we need to provide more structured inputs:
Character description
Background and environment
Story context
Scene breakdown
Sora also has a scene-based generation mode (e.g., multiple 10-second clips).
So the next step is simple:
Create a short 10-second clip first and experiment.
AI Meal Planning App Concept
The UI Looked Like a Real App
Another participant shared an idea for an AI-powered meal planning app.
The design was surprisingly detailed:
Screen transitions
Checkbox-style user interface
“AI is thinking” animation
Result presentation screen
Honestly, it looked like a complete app concept.
It felt like something that people could actually use.
But There’s a Catch
Building an app from scratch requires:
Programming
Application development
So jumping straight into full development would be difficult.
But here’s the interesting part.
Before Building an App
Build a “Personal AI” First
Instead of building an app immediately, there’s a simpler approach.
Use ChatGPT’s Custom Instructions.
For example, for a meal planning assistant, you can define:
Output format
Health conditions (e.g., diabetes, hypertension)
Dietary restrictions (low sugar, low salt)
Available ingredients at home
Family structure
Once this information is stored, you can simply ask:
“Plan five dinners for this week.”
And ChatGPT will generate responses using the predefined format and conditions.
In other words:
You can create an app-like experience using settings alone.
This approach can be surprisingly powerful.
The Key to Effective AI Use
AI Needs Your Personal Data
Another important insight emerged during the discussion.
When you ask AI questions, the answers often become generic advice.
Why?
Because the AI doesn’t know your personal context.
To solve this, we need a place to store personal information and data.
Possible tools include:
Notion
NotebookLM
By storing materials such as:
social media posts
documents
ideas
notes
and connecting them to AI tools, we can gradually build
a personalized AI assistant.
In many cases, successful AI use is less about the tool itself and more about
how you organize and store your data.
Audio Editing Discussion
Should We Remove Filler Words?
We also had an interesting conversation about audio editing.
In podcasts or recorded talks, people often worry about filler words such as:
“uh”
“um”
“well”
We tested some AI tools to see if they could automatically remove them.
Descript
Descript can automatically remove filler words like:
“uh”
“um”
But currently, support for Japanese filler words is still limited.
Adobe Podcast
Another tool we tried was Adobe Podcast.
Its strength is audio quality enhancement.
Even recordings made in a normal room can sound like they were recorded in a studio environment.
But Are Fillers Always Bad?
Interestingly, filler words are not necessarily negative.
They often reflect human thinking processes:
“Well…” → hesitation or uncertainty
“Uh…” → thinking while speaking
Removing them completely can sometimes make speech sound unnatural.
For now, the most practical approach is:
AI + human editing together.
Homework for Next Session
Prepare Three Slides
For the March 19 midterm presentation, each participant will prepare three slides:
What problem are you solving?
How will you solve it?
Who benefits from the solution?
The key is not perfection.
What matters most is being able to explain the idea in your own words.
Final Thoughts
This session was not about discovering new AI tools.
Instead, it was about answering a more important question:
How do we actually build things with AI?
Participants started thinking about:
character consistency in video production
realistic approaches to building AI applications
structuring personal data for AI use
Each idea moved a step closer to real implementation.
AI THINKING LAB is now shifting from the
“thinking phase”
to the
“building phase.”
And the next session will take us even closer to turning these ideas into reality.
