[Report] How Small Studios Without AI Engineers Can Compete! Game Development AI Utilization Strategy and Case Studies #CEDEC2025

Hello, this is Irui from the Game Solutions Department.
Today I'll be reporting on the session I attended at CEDEC2025 titled "AI Utilization Strategies and Examples for Small Studios That Can't Hire AI Engineers".

Session Overview

While AI is increasingly being utilized in game development, I feel that small studios face the following challenges:

• They're using AI for basic chat and text generation, but haven't implemented more advanced applications
• Trying to do so would require budgeting and securing resources for high-salaried AI engineers, which is difficult
• Nevertheless, they need to explore ways to utilize AI to keep up with the wave of efficiency in game development

To address these challenges, Blast Edge Games, a small game development company, has been combining various existing AI technology services.

Starting in 2024, we built and implemented a system that utilizes AI technology to perform game planning tasks (creating design documents on Notion pages and the associated specifications).

Specifically, we connected knowledge information created on Notion (RAG) with Dify, a tool well-suited for AI application development, to build a workflow for creating and editing specification documents. We will present this process and the future challenges that emerged from it.

This presentation is not simply an explanation of slightly more convenient ways to use ChatGPT, but rather a description of a game development company seriously embracing AI utilization, taking two or three steps forward by using existing tools while developing in-house solutions to fill the gaps.

Quoted from the CEDEC2025 session page

Memorable Points

AI Implementation Triggers and Resulting Challenges/Strategies

The session began by explaining the background that led to AI adoption and the process of establishing specific policies.

The initial trigger for considering AI implementation came 2-3 years ago with the rise of ChatGPT. Seeing the evolution of AI, they expected that utilizing it could significantly improve the cost, speed, and accuracy of game development, and they felt anxiety about falling behind competitors if they didn't adopt this trend.

However, when they tried to implement AI, they encountered problems: specialized AI engineers were difficult to recruit due to intense competition, and small studios couldn't allocate many resources to research and development as an upfront investment while managing daily operations.
Additionally, they felt that even if they spent resources on in-house development, there was a risk that what they built would quickly become obsolete due to the overwhelming speed of technological advancement by companies specializing in AI development.

From these issues, they arrived at the idea of differentiating themselves through clever usage and workflow construction techniques rather than developing everything from scratch.
Specifically, they focused on improving specification document creation tasks where they had expertise as game planners, and decided to consider operating by connecting existing tools rather than developing in-house.

### Building a Specification Document Creation Workflow Tool Combining Dify and Notion

What I built to realize the aforementioned approach is an automated specification document generation system using AI that combines existing tools, Dify and Notion.

Dify is a framework that allows you to build AI applications with no-code, where you can assemble various process flows like inputting prompts to specific LLMs or making API calls using output results, similar to Unreal Engine's blueprints. The entire workflow logic is created using Dify.

https://dify.ai/jp

Notion is a document management tool used for storing specification documents generated by Dify via API and maintaining documents for RAG purposes.

During the session, a demonstration was given showing the workflow from using project documents to generating game title screen specifications.

Tool Evaluation and Quality Improvement Techniques

When actually using the tool, while adjustments such as repeated instructions to AI and manual edits after output to Notion were necessary to bring specifications closer to completion level, the work efficiency improved 3 to 5 times according to personal experience. However, developing the tool itself cost more than 30 times the effort.

Specification documents generated by AI always require human review, but to streamline the review process, it proved effective to templatize the document structure, output mind maps and flowcharts, and display AI thought processes for debugging-like verification, all to reduce the burden on reviewers.

Additionally, to improve the overall quality and accuracy of the tool, it's necessary to approach from both "the quality of text generated by LLM" and "usability according to user intent." For improving text quality, enhancing RAG and information provided in prompts was effective, while for improving usability, enhancing Dify nodes proved useful.

Since this tool was specialized for personal work style, it was mentioned that it would be sufficiently useful for future work, but customization would be necessary for use by other planners or projects.

Impression

I found it interesting that even in the unique task of creating game development specifications, analyzing the process and delegating appropriately to AI can significantly improve work efficiency.

I felt that a similar approach could be applied to tasks outside of game development as well.