Seaside Cloudworks

Since methods differ between personal and corporate use, we will describe them separately. The usage also accommodates considerations such as corporate compliance and information leakage.

A. Use of ChatGPT Enterprise or Teams Plan for Business

Create an internal-only knowledge base

With ChatGPT Enterprise, you can upload internal knowledge (documents, FAQs, manuals, etc.) so employees can ask ChatGPT questions.
This knowledge is not shared externally and is only accessible to users within the same corporate account.

• Examples：
  • Upload product specifications and development guidelines so employees can ask questions in natural language
  • Compile internal workflows and procedure manuals for use in new employee training and support
  • Centralize common customer service FAQs to streamline customer support

---

Data Security

• Data is managed separately within enterprise tenants
  On Enterprise/Team plans, exchanged data is completely isolated from other OpenAI users.
• Not used for model training
  Conversations and uploaded files are not used to retrain models.
• SOC 2, GDPR compliant
  Data protection meeting compliance requirements is implemented, making it suitable for enterprise use.

---

Management Functions

• Administrator Console
  Enables user management, permission settings, and SSO (single sign-on) integration.
• Access Control
  Allows creation of dedicated knowledge bases accessible only to specific departments.

---

Team Collaboration

• Shared Conversations
  Share conversation threads with internal members to continue asking questions to the same AI assistant.
• Project-Specific Workspaces
  Organize knowledge and history per project, accessible only to relevant stakeholders.

---

In summary, using a corporate account allows you to create an “internal-only knowledge base + secure AI chat environment.” You can leverage AI across your entire organization without worrying about information leaking outside the company.

B. Uses not recorded for personal purposes

For personal use, it seems best to either use methods that leave no history or utilize temporary chats. Temporary chats, in particular, are only available for that moment, making them quite useful.

1.Turn off conversation history

ChatGPT has a setting called “Conversation History and Training.” Turning this off prevents conversation content from being saved to your account history and prevents it from being used to improve the model.

• How to Set Up
  1. Click the ⚙️ “Settings” icon in the bottom-left corner of the ChatGPT screen
  2. Select “Data Management”
  3. Turn off “Conversation History and Training”

In this state, past interactions are not retained and cannot be referenced upon revisiting.

---

2.Utilizing Temporary Chat

OpenAI offers a “temporary chat” feature.
This is a session that leaves absolutely no history and cannot be restored after closing.

To use temporary chat, click the dotted circle in the upper right when opening a new chat. When the checkmark appears, it becomes a temporary chat.

• Function Features:
  • No history is saved at all
  • You can use the same features as regular chat
  • When you close the session, the conversation disappears

---

3.Local Environment Tweaks

• Use Incognito Mode
  When you don’t want to leave browsing history or cookies in your browser, use Chrome’s “Incognito Mode” or Safari/Firefox’s “Private Browsing” feature.
• Avoid Leaving Screenshots or Logs
  If automatic recording features (like screenshot syncing or history backups) are enabled on your PC or browser, disable them as needed.

---

4.Regular History Deletion

Even if you had the setting enabled to keep a history, you can delete it later.

• Delete unnecessary chats from Doba’s “History”
  • You can also delete them all at once via “Settings → Data Management → Delete All Conversations”履歴削除の手順

---

Summary

個人アカウントで安全に使いたいなら、

• Turn off conversation history
• Use temporary chat
• Do not include confidential information / Use anonymized data

By keeping these three points in mind, you can use it with peace of mind without leaving a record.

Finally

Let’s safely and meaningfully use the compelling generative AI whose capabilities improve day by day.

August 7, 2005 ChatGPT-5 was released. Previously, it was necessary to use different models such as 4o and o3 depending on the application, but now they have been integrated into a single model. The official website states, “This is our highest-performance model, integrating inference and chat into a single system.” Please see the official announcement below.

Introducing GPT-5
https://openai.com/ja-JP/index/introducing-gpt-5

At Seaside Cloadworks, we use ChatGPT for simple tasks such as throwing out prompts for planning and image generation and getting responses, but we rely on it heavily for image generation in particular. Therefore, we asked ChatGPT-5 how image generation has evolved since it became ChatGPT-5.

What has improved in image generation with ChatGPT-5?

Since there was no explanation in the official document mentioned above, I asked ChatGPT-5 itself. The following is the answer.
It says “I” as if it were the person’s own answer, so I believe it to be true.

---

ChatGPT-5 (me) has the following improvements in image generation compared to previous models (GPT-4 series).

1.More natural integration of recognition and generation

• Previously, “image reading models” and “image generation models” were separate, and it was sometimes necessary to switch between them during interactions.
• With ChatGPT-5, image understanding (analysis) and image generation (drawing) interactions have become seamless, and the process of loading an image → understanding its content → processing it if necessary → generating a new image has become smooth in a single interaction.

2. Improving the accuracy of instruction comprehension

• Even complex compositions and detailed style specifications (e.g., “A cat playing the piano in a wooden café bathed in the light of dusk, painted in an oil painting style”) can now be easily reproduced as intended.
• The composition’s spatial relationships and atmosphere, which tended to be vague in the past, are now reproduced more faithfully.

3. Enhanced high resolution and detail rendering

• The resolution and detail of the output image have been improved, and the expressiveness of small objects and background textures has been enhanced.
• This makes it possible to create images that are less blurry even when enlarged.

4. Flexible editing functions

• Partial editing (inpainting/outpainting) such as “replacing part of the image,” “changing only the color tone,” and “removing only the background” is now possible with more natural synthesis.
• It is now possible to modify existing images without destroying their atmosphere.

5. Consistency and reproducibility of style

• I am skilled at reproducing the same atmosphere and characters across multiple images with the same characters and compositions.
• This makes it easier to create series works and comic-style continuous cuts.

My impression after trying it out a little

Until now, I provided the original image and gave prompts in Japanese (perhaps the prompts themselves were vague), but the resulting images sometimes looked strange (the shape of the hands) or just didn’t feel right. In some cases, the people in the foreground looked good, but the background was rough, so I tried again and again until I got a good result.

This time, I used Ayase and Honda, and was able to produce the following images in one shot. I feel that the accuracy has improved.

① Creating a series of shots of two people walking

Prompt
Can you create a series of shots of these two people walking along the beach? It would be nice if there was a slight breeze blowing their hair and clothes.

Mr. Honda (left) is in the same posture as in the original image, but his facial expression is natural, and Ms. Ayase (right) is very natural in the way she waves her hands and the direction of her face. The number of arms and legs is correct, and I think it’s very good. I think the accuracy is good considering that the prompt is rough.

② hug

Prompt
The two are hugging each other from the front, looking toward the camera. Please make the background transparent in a bust shot. Their expressions should be soft and smiling.

I didn’t give any specific instructions, but I think it looks quite natural. The hands and fingers look natural, and I was worried that it might look strange, but I’m relieved that it looks healthy. Mr. Honda’s expression is the same as in the previous image, but Ms. Ayase has her mouth closed, so there is a slight change. I think the stability is much better than before. It’s good.

Summary

I’ve only generated these two images so far, but I feel a sense of stability. I want to use it more and see how it feels. Also, I’ve only tried it with images, so I want to experience the improved reasoning capabilities of ChatGPT-5 in the future. I have high expectations for this.

If you use ChatGPT regularly, you may have encountered messages like:
“Your chat is full”
or
“Starting…” that never proceeds.

While refreshing the page or opening a new tab may help temporarily, these solutions often fail to fix the underlying issue. In this article, I’ll show you how to use the Project feature in ChatGPT to organize, store, and continue your conversations more reliably.

Common Problem: ChatGPT gets stuck on “Starting…”

As I wrote in a previous post, there are quick workarounds for this.
However, you may have noticed that even if it works once, reopening the same thread often gets stuck at the same point again.

The Solution: Use “Projects” to Manage and Continue Your Conversations

ChatGPT now offers a handy feature called “Projects.” By using it, you can break free from thread limits and instead build up a long-term, structured interaction.

Benefits of using Projects:

• Continue conversations even if one chat gets stuck
• Organize information by topic or file
• Accumulate knowledge that persists across chats

Step-by-Step Guide: Moving to a Project-Based Workflow

Here’s a step-by-step guide (with screenshot suggestions) to help you make the switch.

1.When you see “Your chat is full” or it’s stuck on “Starting…”

If the chat is just full, can I continue using it afterwards? However, the message saying “full” keeps appearing, which is annoying. Also, if “Starting” remains, it will always resume from there, which is useless.

If there is content in that thread that I want to reuse, please output it as knowledge. I think either a table or Markdown format would be fine.

2.Create a New Project

From the left sidebar, click “+ New Project”.
Give your project a name and click Create.

3.Add a Chat to the Project

Inside a project, you can create multiple chats. This makes it easy to separate conversations by topic or task.

4.Upload Files to the Project

Projects allow you to upload files such as PDFs, Word documents, or text files. This enables ChatGPT to access and reference your materials during the conversation.

Thread Memory vs. Project Knowledge

You might feel that continuing in the same thread helps ChatGPT “remember” better.
However, when using Projects with files, you’re building more reliable and structured knowledge.

Rather than relying on a thread’s short-term memory, you’re feeding in referenceable knowledge that persists and is accessible across chats.

I also add things that I have stopped using to my project under “Starting.”

ChatGPT has added an “agent function” that is now available with Plus ($20/month), so I tried it out to organize my unread Gmail messages. I’ve seen demos that demonstrate “cross-service operations” such as “ordering lunch from Uber Eats,” but this time I focused on the Gmail connector to see how much automatic organization it can do.

Agenda

• What we did in this article (prerequisites)
• Preliminary preparation: Enable the Gmail connector
• Experiment 1: Are there any important emails that require a reply?
• Experiment 2: Can labels be automatically assigned to detected DMs?
• Experiment 3: Can DMs be automatically moved to the trash?
• What the Gmail connector can and cannot do
• Experiment 4: Count the number of emails by sender and month, and create tables and graphs
• Experiment 5: Can this be automatically executed on the 1st of every month and notified via email?
• Bonus: It provides instructions for the DM opt-out link (but the final click must be done manually)
• Things learned from using it and ideas for utilization
• Summary

Preparation: Enable Gmail Connector

First, I sent the following prompt to the agent.

Prompt
I want to extract emails that require a response from my unread emails in Gmail. I want to sort DMs.

However, I encountered an error saying “Unable to connect to Gmail.”
So I connected Gmail from [Settings → Connectors] and tried again.

Experiment 1: Are there any important emails that require a reply?

Set up the Gmail connector and try again. I will also let you know when the setup is complete with the following prompt.

Prompt
We have configured your device to connect to Gmail. Please try again.

Result：

• Important emails from people that require a reply: 0
• 1 security notification from Google (no reply required)
• The rest are mostly direct messages (advertisements/mail magazines).

The agent’s actions appear to be live streaming, as shown in the figure above. As a result, the loneliness of not communicating with others via email is highlighted. Since I do not use Gmail for actual work, I concluded that this is understandable.

Experiment 2: Can labels be automatically assigned to detected DM?

While understanding that Gmail also categorizes emails as promotions, I will experimentally try assigning labels. This is because assigning labels will make it easier to filter emails in Gmail.

Prompt
Please set labels for the above DMs and detected emails. These will be used for filters in Gmail.

Result：

• It was not possible.
• The Gmail connector does not have a “label assignment” function (i.e., the operation API is not provided).
• As an alternative, it is possible to create your own filter on the Gmail side by “providing search conditions.”

Experiment 3: Can DM be automatically moved to the trash?

If you cannot add a label, try deleting it as junk mail.

Prompt
Please move the DM detected above to the trash in Gmail.

Result：

• This is also not possible. Similarly, the API for deletion and movement is not provided.

What the Gmail Connector can and cannot do

I am currently trying out Gmail Connector rather than an agent, but I realized that I needed to know in advance what it could do, so I checked it out.

possible

• search_emails: Search by criteria (unread, date range, from, subject, etc. Gmail search operators can be used)
• search_email_ids: Quickly retrieve email IDs only
• read_email: Retrieve the body and header of individual emails

impossible

• Labeling, marking as read, deleting, moving, replying, sending
• Automatic execution according to schedule (regular batches) and automatic notification of results via email

Experiment 4: Count the number of cases by sender and month, and create tables and graphs.

If it supports the reference system API, we will try to see if we can display how many emails have been delivered in an easy-to-understand manner. We will count the number of emails sent per month by sender.

Prompt
Can you count the number of emails received in Gmail by sender on a monthly basis and summarize them in a table and line graph?

Result：

• Compile the number of messages sent from each source into a table for May to July (the last three months).
• Visualize the increase or decrease with a line graph.
• In addition, it summarized the content of messages sent from each source, which was useful in deciding which sources to stop sending messages from.

Experiment 5: Can you automatically run this on the 1st of every month and notify me by email?

If this works, it should make it easier to sort out unnecessary direct mail. I’ll ask if they can do this regularly every month.

Prompt
Can you automatically perform this analysis on the 1st of every month and notify me of the results by email?

Result：

• NG. ChatGPT (agent) alone cannot perform “scheduling” and “email sending.”
• Alternative:
  • Automate equivalent processing with Google Apps Script (GAS) + Gmail API.
  • Or use automation tools such as Zapier/Make.

Extra: They will provide you with a link to stop receiving DMs (but you have to click on it yourself).

Finally, if it could automatically stop streaming…

When I requested to stop receiving direct mail from a mall, they found the “unsubscribe link” in the email body and provided it to me.
However, the actual unsubscription process must be performed by the user themselves (external site screen transitions and login operations are not possible in this environment).

However, this agent guided me to the site where I could stop distribution and even displayed the browser. For some reason, my email address was garbled, so I couldn’t actually stop distribution, but I thought it was quite helpful.

Things I learned from using it/Ideas for utilization

Positive points

• Extremely powerful for “reading” tasks.Unread sorting, keyword monitoring, and notification extraction from accounts are quite practical.
• Visualize a large number of DMs at a glance by “sender × month.” It is easy to see at a glance where to stop for maximum effect.
• Output “totals & graphs” with just a prompt. Convenient for initial analysis.

Shortcomings (current limitations)

• Unable to perform operations on Gmail (labeling, deleting, replying, sending)
• Unable to perform scheduled execution and automatic notifications
  → Must be combined with external workflows/scripts

It seems like a good way to use it.

• Leave the “visualization” part to ChatGPT.
  → Set up Gmail filters and unsubscribe while viewing the results.

Summary

• ChatGPT Agent + Gmail Connector is powerful for “searching, summarizing, aggregating, and visualizing” emails.
• However, it is currently unable to perform email “operations (such as labeling and deleting) or automatic sending/scheduled execution.”
• It is sufficient for the purpose of “first visualizing and then deciding where to stop.”
• In the future, if operation-related APIs are released, it has great potential to become a “practical automation tool.”

I am still not fully proficient in using the agent, but I feel that it is possible to connect it with other services and request support. I look forward to its future development and would like to try out various features.