On September 10, 2025, a ChatGPT hackathon took place at the event space on the 25th floor of Mercari’s Tokyo office. The theme was “Custom GPT dev for work.”

Approximately 50 members from different companies across Mercari Group gathered for the event. Following an explanation from OpenAI of the tool’s latest updates and advice on creating custom GPTs, participants got to work creating custom GPTs for use in day-to-day work.

This event report provides a rundown of what happened at the hackathon.

Schedule

2:00 – 2:10 PM: Introduction
2:10 – 2:30 PM: Overview of Custom GPTs
2:30 – 5:00 PM: Discussions and prototyping
5:00 – 5:55 PM: Presentations
5:55 – 6:00 PM: Awards and closing

The event began with an overview of custom GPTs. Following this, participants worked on developing ideas individually, and then gathered for discussions in teams of four to five to decide on the custom GPT they would create. They then built their custom GPTs over a duration of about two hours.

Hearing the latest GPT-5 updates straight from OpenAI Japan

New features of GPT-5

Kenta from OpenAI Japan opened the event with an explanation of custom GPTs. He also generously shared with us all the latest information regarding GPT-5.

According to Kenta, while in the past users had to decide whether using GPT-4o mini or another GPT model would best suit their needs, the new update comes with an Auto Mode in which users can simply converse with GPT-5 for all requests. During the conversation, the model will adjust its responses, thinking carefully when presented with complicated issues and answering quickly when given simple ones.  This provides a natural user experience.

Users also have the option of selecting a specific mode for their use case. The following modes are available for when you want AI to think deeply about strategic issues or to set a recommended model for your custom GPT.

• GPT-5 Auto: Automatic adjustment of speed and thinking based on content
• GPT-5 Thinking: Advanced thought processes for complex issues
• GPT-5 Pro: Highest-level thinking skills for strategic issues

New considerations when creating prompts

With the update to GPT-5, there are new things to consider when writing prompts. Kenta explained the following two points.

– Removing contradictions is more important now 

“GPT-5 faithfully follows human instructions. Therefore, any contradictions in the instructions will cause GPT-5 to overthink, affecting its response. You can ask ChatGPT to check whether there are any contradictions in your instructions.”

– Improved quality through self-evaluation 

“We built a mechanism that allows ChatGPT to evaluate its output based on pre-determined criteria and continuously improve the accuracy of its responses. By repeating this self-evaluation cycle, it will be able to produce higher quality output.”

Keys to building custom GPTs

After we received the latest updates from OpenAI, we moved on to the main topic of this year’s event—custom GPTs. Kenta broke down the three components of custom GPTs.

Instructions: Automate repetitive tasks
Increase efficiency by storing instructions for repetitive tasks in a custom GPT.

Knowledge: Organize specialized information
ChatGPT accumulates information about the company and stores it as knowledge in file format.

Actions: Connect to external systems
This feature for connecting to external systems is not currently available for use at Mercari. 

Keys to creating a knowledge file

The quality of a custom GPT is greatly influenced by the quality of the knowledge files it is given. Kenta shared specific information and expertise imparted from developers with us.

Choosing a file format

• Table data: CSV or excel formats recommended; descriptive column headers are key
• Text data: JSON, markdown, txt, and word formats are prioritized; there are limitations with PDF files
• Information placement: Place important information at the beginning (up to approximately 110,000 tokens can be read directly)

Kenta explained the technical background behind why the information at the beginning of a file is so important. 

“Information within files is read directly from the beginning of the file, and the remaining portions that aren’t read directly are vectorized and processed with semantic search. Therefore, it’s easier for AI to understand files with large amounts of information if the important information is summarized at the beginning. You could think of it this way: files that are easy to understand for humans are also easier to understand for AI.”

Let the custom GPT hackathon begin!

After we learned about the technology, the hackathon began. Participants were divided into teams of four to five and followed a three-step process to develop custom GPTs for use in everyday work.

– Individual work (15 mins)
Participants first came up with ideas individually, using a custom GPT provided by OpenAI called “ChatGPT Use Cases for Work” as a sounding board to list out the issues within their workflows. 

– Team discussions (15 mins)
Next, participants formed teams, shared their ideas, and chose one or two to work on as a team. With the teams comprising members in different roles across Mercari Group, even a short discussion brought to light various ideas, solutions, and issues that individual members would have been unaware of otherwise.

– Development and trial and error (120 mins)
Finally, the hands-on development began. Participants worked to develop custom GPTs for the ideas they chose as a team. Beyond creating a working prototype, participants repeated trial and error to develop something that could be used in actual work, while simultaneously preparing to present their results.

Custom GPTs that improve work efficiency, created in just two hours

After about two hours of development, each team presented their custom GPTs. Four minutes were given to each team to present their issues, solutions, and the expected results of their creations.

Three particularly outstanding ideas were selected respectively for the People’s Choice Award (decided by participant vote), the Judge’s GPT Award (decided by @mark and his custom GPT), and the Yurino Award (decided by the custom GPT “Yurino bot” created by new graduate member @Yurino, who facilitated the event).

Let’s dive into the winners, starting with the People’s Choice Award!

(1) People’s Choice Award 

• Selection criteria: Participant vote　
• Reason for award: High practicality, elimination of dependency on individuals, and potential for use across multiple job types

And the winner was…

Automated legal reviews (Team 4)

This custom GPT automated the legal check process for marketing strategies. Team 4 was a mix of CS and Legal Team members, and they presented a solution for an issue common to both of their teams—lightening the burden for both parties when handling a inquiry. Their solution is estimated to save 130 hours per person per year!

Tasks the custom GPT takes care of include:

• Automatic evaluations based on laws, internal policies, and user understanding
• Output of three evaluation outcomes: Pass, Fail, and Pending
• Automatic generation of suggestions for improvements and materials for escalation

The solution is designed to not only output simple pass/fail evaluations, but automatically generate concrete suggestions for improvements, and even write Slack messages to make escalation seamless.

(2) Judge’s GPT Award

Fitting for the event, this award by our judge @mark featured winner selection via a custom GPT. The winner was selected against the following criteria.

• Selection criteria: Custom GPT that reflects a CEO mindset
• Reason for award: Designed with safety in mind, and significantly reduces work time

And the winner was…

Generation of accounting requirements for new businesses (Team 7)

This custom GPT provides a system for listing up accounting requirements when planning a new business. It is designed to allow even members with only one or two years of accounting experience to handle tasks currently dependent on specific members of the team. This solution is expected to speed up tasks that currently take around 2 hours to just 30 minutes! The system developed includes the following features:

• A guide in question format for defining requirements
• Presentation of comprehensive accounting processes
• Prevention of backlog

Overlooked requirements can cause major reworks if discovered during development. This custom GPT is designed to prevent such issues from occurring. The solution was praised as “a system that eliminates omissions and allows for comprehensive consideration.”

(3) Yurino Award

Like the Judge’s GPT Award, the Yurino Award winner was also selected using a custom GPT, the “Yurino bot,” created by new graduate member @Yurino.

• Selection criteria: A fresh, new graduate perspective
• Reason for award: Immediate practical application and structuring of knowledge

And the winner was…

Networking assistant (Team 6)

This custom GPT provides a tool that is expected to see frequent use in preparation for events with other companies. It can provide support for the entire process of holding networking events, from planning to following up on post-events tasks. By simply entering company names and the event format, users can conduct company research through web searches, create specific to-do lists, and automatically set schedules.

This solution was commended for its capability to prevent dependency on specific members and omission in tasks during event preparations.

And more ideas!

In addition to the award-winning ideas, there were many other custom GPTs created which could be implemented in day-to-day work.

Efficiency for market research
This custom GPT conducts cross-sectional analysis of domestic and international cases and outputs standardized reports. The solution aims to streamline research work in the complex regulatory environment of the financial industry. It is estimated to reduce monthly research time from 680 hours to 340 hours. It was also evaluated highly for its capability to overcome the language barrier that existed until now when gathering information to analyze cases from overseas.

Efficiency for social media operations
This custom GPT supports the operation of the official Mercari account with 1.6 million followers. The solution generates a database of past posts that received high engagement so that they can be referenced to produce future posts in Mercari’s friendly tone. It leverages the web search feature to instantly generate posts that take into account the latest trends, such as the launch of new smartphone models. The solution was also commended for its ability to automatically produce images for posts through integration with the image generation feature.

Assistant for creating specifications
This custom GPT is designed to help PMs and PMOs save time and ensure more consistent quality when creating specifications. It incorporates fintech-domain templates to create high-quality specifications in an interactive format. In addition to reducing work time by 30%, it also provides qualitative benefits, like making it easier to start writing and helping to identify perspectives that are often overlooked. Once the specs are completed, the text can also be directly transferred to Confluence or Notion, making it immediately usable for work.

Standardized event preparation
This custom GPT standardizes the preparation process for holding events with other companies. It provides a system that covers all processes in the event organization flow—from preparation and logistics on the day of, to follow-up tasks after the event—and helps ensure a smooth experience for even those holding events for the first time. The solution automatically generates a customized task list and schedule based on basic information, such as the number of participating companies, whether catering is needed, and the location of the event. It was designed with a focus on the idea “Move Fast without oversights.”

In conclusion

This hackathon produced 10 practical custom GPTs in the limited time of about two hours. Kenta concluded the event with his closing remarks.

“I’m so glad to see the variety of ideas that were developed today, each with the potential for further expansion, and many of which are close to being used in your day-to-day work already. I’m sure there were also challenges and questions that couldn’t be addressed today. OpenAI will continue to work with Mercari to consider development of additional features.”

Once again, we would like to thank everyone at OpenAI who helped make this event a reality. We’re confident that the 10 solutions developed here will provide hints for practical usages of AI across industries and job types. We hope the coverage of this event has provided you with ideas for developing your own custom GPTs, too!