From Prompt to a Working App A Beginner’s Course in AI App Development: Experiencing Specification, Constraints, and Context

🎥 The YouTube version is also available:

From Prompt to Working App: A Beginner’s Guide to AI App Prototyping

Books: Rapid Prototyping in the Age of Generative AI: Learning with Streamlit　Building with Docker Al Business Systems

Generative AI is not limited to writing and summarizing text.

By organizing ideas, communicating the necessary conditions, and turning them into screens and functions, it is now possible to prototype small applications even without extensive programming experience.

However, simply asking AI to “build an app” does not guarantee the result you expect.

What are you trying to create?
Who will use it, and in what situation?
What conditions must it respect?
What information may it use as the basis for its decisions?

When this information is missing, AI may generate something that looks plausible, but it is less likely to be useful or aligned with its intended purpose.

In this course, participants will use familiar themes—such as travel planners, meal planners, and study-plan builders—to improve prompts step by step and create working Streamlit applications.

The goal is not to memorize code.

It is to develop the ability to turn a vague wish into a specification that AI can implement.

Why Use a “Working App” as the Learning Theme?

You can learn how to write prompts through text generation alone.

For example, when asking, “Please create a travel plan,” it is easy to see that adding conditions such as budget, travel companions, interests, and weather improves the quality of the answer.

But building an application that actually runs makes these differences even clearer.

Are the input fields sufficient?
Does the result change when the conditions change?
Can users operate the app without confusion?
How should the app behave when necessary information is missing?

Gaps in a prompt appear directly as gaps in the application’s interface and behavior.

Conversely, when you carefully communicate the user, purpose, constraints, and data, the application begins to take on real specificity. Through this experience, participants realize that a prompt is not merely a request to AI—it is language for designing a system.

Start with Familiar, Selectable Themes

When teaching a diverse audience—such as university students, new employees, or people in non-IT departments—starting with a highly specialized business scenario is not always appropriate.

This course therefore provides familiar themes that are easy to imagine and that become more interesting as participants add conditions.

At the beginning of the course, participants can choose or vote on a theme. A theme they select themselves is more likely to generate useful ideas such as, “I want it to work this way,” or, “This information is missing.”

A 90-Minute Journey from Prompt to Application

This is not a course where participants simply receive a finished product. They begin with an ambiguous request, add conditions themselves, and improve an app through dialogue with AI.

1. Start with an Ambiguous Request

For a travel planner, the initial request might be as simple as this:

Create a Streamlit app that generates a day-trip plan for Kamakura.
Allow users to enter their destination, budget, and interests,
and display a recommended itinerary.

AI will likely create an app with a basic structure.

At this point, however, participants are asked to consider the following questions:

• Who is this trip for?
• Should departure time be considered?
• What should happen on a rainy day?
• What if the traveler is accompanied by children or older adults?
• Should options that exceed the budget be displayed?
• What should the app do when information is insufficient?

The first version is not a failure. It is the starting point for discovering the information needed for the next improvement.

2. Make the User and Purpose Specific

Next, add details about who will use the app and why.

This app is for people visiting Kamakura for the first time
to explore options for a day trip.

Users may access the app on a smartphone,
so keep the interface simple.

Include input fields for departure time, budget, travel companions,
interests, and preferences for rainy weather.

Display a morning, lunch, and afternoon itinerary,
an estimated cost, and any points to note.

At this stage, the input fields and result display begin to change.

The app is no longer designed only to “create a travel plan.” It begins to reflect a more specific purpose: helping first-time visitors make decisions quickly.

3. Add Constraints to Make It More Practical

Next, add the conditions that the app must respect.

Add the following constraints:

- Do not display recommendations that exceed the budget
- Reduce walking distances for travelers accompanied by older adults
- Include rest areas for travelers with children
- Prioritize indoor or covered locations when it is raining
- Do not overfill the itinerary when the departure time is late
- If required conditions are missing, display the items that need confirmation

Here, participants learn that producing an answer that “looks good” is different from making decisions that follow the required conditions.

The app’s behavior can also be divided into three categories:

This way of thinking is useful not only for travel and meal planning, but also for future AI use in the workplace.

4. Provide Context and Define AI’s Decision Boundaries

For AI to work appropriately, it needs context: the information that forms the basis for its decisions.

For the travel planner, instruct the AI to use only the sightseeing data prepared for the course.

Use only the sightseeing data prepared for this course
when making recommendations.

Do not invent stores, opening hours, or prices that are not included
in the data and present them as facts.

Each sightseeing location should include its name, category,
whether it is indoors or outdoors, estimated visit duration, cost,
and whether it is suitable for children or older adults.

The lesson here is not simply that more information is always better.

It is essential to define what AI may use as evidence—and what it must not use as evidence.

This is a key part of designing systems that use AI appropriately.

5. Run the App and Improve It Through Language

The completed app is launched and checked using Streamlit.

streamlit run app.py

Once the app is running, participants operate it and identify further improvements.

For example, they might ask AI to make the following changes:

Make the following improvements:

- Allow users to select their budget with a slider
- Allow users to select travel companions with radio buttons
- Display why each recommendation was made
- When conditions are missing, clearly show which inputs are required
- Display the points that should be checked by a human at the end

Even without editing code directly, participants can improve the app by communicating the purpose of the change and the desired behavior in words.

This repeated cycle—build, run, and revise—is a new way of prototyping with AI.

What Participants Will Learn

Through this course, participants will develop the ability to:

• Break down an ambiguous request into specific requirements
• Describe users, purposes, and usage situations clearly
• Design constraints and exception conditions
• Define the data AI may use and the boundaries of its decision-making
• Discover improvements through a working prototype
• Turn ideas into tangible forms through dialogue with AI

A longer prompt is not necessarily a better prompt.

What matters is whether it includes the information AI needs to make decisions and implement the application appropriately.

A good prompt is not a piece of writing that makes AI guess the correct answer.
It is a blueprint for sharing human purpose, conditions, and judgment with AI.

Conclusion

In the age of generative AI, not everyone needs to become a programmer.

However, many people will need the ability to organize their ideas, define their goals, express necessary conditions in words, and experiment in small steps with AI.

By starting with familiar themes such as travel, meal planning, and study plans, participants can experience collaboration with AI without feeling intimidated.

That experience can eventually lead to the ability to ask important questions in learning and work: What should we build? What should be entrusted to AI? And what should people still check?

Create something that actually works, starting from a prompt.

That small success can become a confident first step toward using AI effectively.

Chinoba

Chinoba

Intelligence as Relationship

Research Platform

founded by

Masao Watanabe

AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance