SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

Teacher's AI Environment, June 2026 Edition: The 4-Piece Set Rebuilt with Claude Code, Codex, Cursor, and Obsidian

[Understand in 10 seconds] The conclusion of this article

✅ In June 2026, the AI landscape is shifting toward a powerful configuration using a 4-piece set: "Either Claude Code or Codex + Cursor + Obsidian" (Gemini users should also consider Antigravity)
✅ Beginners stumble in two stages. The first stage is installation and habituation, and the second stage is the wall of invisible files. The latter can be solved instantly with Cursor
✅ All three major companies—ChatGPT, Claude, and Gemini—have released AI that can be used within spreadsheet and document creation apps. The golden rule for teachers is to "install only one company's tool that matches the conversational AI you mainly use"
✅ When you store conversations, brainstorming, and research with AI in Obsidian, the concept popularized by Karpathy's LLM Wiki of a "second brain you nurture" becomes a reality
✅ Voice input is in the era of the big three: Aqua Voice / Superwhisper / Typeless. You no longer need to type, and "fragmented notes" will start to accumulate


Introduction: The landscape has completely changed in 5 months

This article is written for teachers in the field who want to use AI to streamline school administration and lesson preparation. If you are someone who asks "What is AI?", you might need some prerequisite knowledge.

In January 2026, I wrote "The Complete Guide to Choosing AI Tools for Teachers: Which Should You Choose—ChatGPT, Claude, or Gemini?".

Five months have passed since then.
Honestly, it is no longer a question of "which one to choose."

Five months ago, the main way of using AI was "making round trips to AI": opening the ChatGPT page, writing a prompt, copying the resulting text, and pasting it into Word.

It is different now.

  • AI has started living inside your own folders (Claude Code, Codex, Antigravity)

  • AI has started coming down directly into Excel and Word (All three companies—Claude/ChatGPT/Gemini—provide official add-ins)

  • AI has entered Google Docs, Slides, Gmail, and Drive across the board (March 10, 2026, Gemini major update)

  • You can now send work from your smartphone to your PC and have it finished while you are out (Claude Code's remote control, Codex's iOS app, Claude Cowork's Dispatch feature, etc., which each company has prepared around the same time)

What has changed is not so much the AI side, but the boundary between AI and your own work environment.

In this article, I will organize the set that teachers (and non-engineer business professionals) can actually run every day as of June 2026, narrowing it down to four tools.
ChatGPT, Gemini, various add-ins, and voice input will also appear, but the main players are strictly the four: Claude Code, Codex, Cursor, and Obsidian. The rest will be treated as training wheels to make daily life easier to manage.

To business professionals:
While this article uses teachers' work (grade-level newsletters, student observations, student guidance records, etc.) as examples, the same structure applies directly to non-engineer business professionals who organize their work in folders. Please try reading this while substituting the examples with your own work, such as: grade-level newsletters → internal documents, observations → evaluation comments, parent communication records → customer interaction logs, and research lesson plans → proposal materials. The logic for choosing and combining tools is the same.

Before reading this article:
Please be sure to check in advance whether your school or board of education permits the use of AI for school administrative tasks.
The Ministry of Education, Culture, Sports, Science and Technology (MEXT) has also published guidelines regarding the educational use of generative AI.
Please use this article after prioritizing your school's internal rules.

What you will gain from this article

  • The roles of Claude Code, Codex, Cursor, and Obsidian, and the logic behind 'why these four'

  • How to overcome the 'invisible file wall' that beginners stumble over using Cursor

  • The positioning of Claude and Gemini add-ins that can be used in Excel/Spreadsheets

  • A starter path for teachers that can be set up for 3,000 to 5,000 yen per month


Chapter 1: The Map of June 2026—From 'AI You Visit on a Trip' to 'AI That Lives Inside'

First, let's organize the landscape.

1-1. Divided into three major layers

As of June 2026, the generative AI that teachers interact with has been divided into three major layers.

  • Conversation Layer: ChatGPT, Claude, Gemini (everyday use via browser chat)

  • Work Layer: Claude Code, Codex, Antigravity (runs within your own folders; can be entrusted with creating, editing, and organizing documents)

  • In-App Layer: ChatGPT for Excel/Sheets, Claude for Excel/Word, Gemini in Workspace (can be called directly from within Excel or Google Docs)

Five months ago, there were almost no options other than the conversation layer.
Now, it is a matter of using these three layers selectively.

1-2. The conclusion for teachers

I will write the conclusion first.

  • Conversation Layer: It's fine to start with ChatGPT (free or Plus for $20/month). Use Claude as well once you get used to it.

  • Work Layer: Either Claude Code or Codex. Choose one (there is no need to run both at all times. If you are a Gemini fan, Antigravity is also a candidate). However, make sure to build it in a way that allows for easy replacement at any time

  • In-app layer: Install the official add-in for only one conversational AI that you use the most (do not install multiple).

And, as the foundation for running these, place Cursor and Obsidian.

In other words, the starter kit for educators in June 2026 is

Claude Code or Codex (choose one) + Cursor + Obsidian + ChatGPT or Claude for conversation

which makes a 4-piece set.

1-3. Estimated Costs: Do not subscribe to everything at once

There is no need to pay for everything from the start. Take it in stages.

🟢 Minimum Start (approx. 3,000 yen~/for those who want to try it first)

  • 1 Conversational AI (ChatGPT Plus or Claude Pro) + Obsidian + Cursor Free Version

  • *Codex or Claude Code is included in the conversational AI fee

🟡 Standard (approx. 4,500 yen~/for those who want to use it daily)

  • The above + 1 voice input tool

🔵 Finishing (approx. 6,000 yen~/for those who want to get serious about AI work)

  • The above + Cursor Pro

Beginners should first subscribe to only one conversational AI and create a "workspace" with the free version of Cursor and Obsidian.
The work layer (Codex or Claude Code) is included there, so one is enough at first.
(This will change due to exchange rates and plan revisions. Please read this as a guideline as of June 2026.)


Chapter 2: Claude Code and Codex—"Choose one + a setup that allows for swapping" is the realistic solution

In "If an educator is starting with AI, ChatGPT 5.5 + Codex app is enough" (May 2026), I wrote that the Codex app alone is sufficient.
This has not changed even now.The first one is enough with Codex. If you are a Claude fan, you can start with Claude Code.

What I want to convey in this chapter is to choose one, and then set up your folders, prompts, and operations so that you can swap to the other at any time.
This is not about running both at all times.

2-1. Positioning of Codex: A sense of continuity from ChatGPT

Codex (OpenAI's coding/work agent) can be used as part of ChatGPT Plus for $20/month. It is a work agent with a wide entry point that can be accessed from anywhere: Web, CLI, IDE extensions, or the iOS app.

  • For those already using ChatGPT,without any additional subscription,you can try out a 'work agent' right away.

  • It is easy to see the immediate benefits by throwing daily tasks at it, such as 'one task in 30 minutes (handouts, communications, formatting tables)'.

  • The basic functions of task-layer agents are all present, such as reading entire folders, providing context via project memory files (`AGENTS.md`), and external integration via MCP.

An educator's daily routine is full of these short-cycle tasks.For ChatGPT Plus subscribers, Codex is the easiest one to choose as a first step, which is my impression as of June 2026.
Training programs for non-engineers are also being organized in various places, and deployment to this segment is progressing rapidly.

2-2. Positioning of Claude Code: The real difference is the 'engine'

I will be honest here.
Claude Code and Codex have very similar basic functions and entry-level pricing (detailed specifications such as usage limits, supported IDEs, and mobile/remote implementation differ for each).

General commonalities in functionality:

  • Reading folders, writing based on past files, and sending requests from smartphones

  • Project memory files (Claude Code uses `CLAUDE.md`, Codex uses `AGENTS.md`)

  • Custom skill/command mechanisms (both have skill directories and allow you to call self-made workflows with '/command name')

  • Operations running separate agents in parallel (Claude Code has a sub-agent mechanism, Codex can do the same with multiple CLI processes)

  • External tool integration (MCP)

Symmetry in pricing structure:

Basic plan (including coding agent)

  • Codex (OpenAI) — ChatGPT Plus $20/month

  • Claude Code (Anthropic) — Claude Pro $20/month

Premium plan (relaxed usage limits)

  • Codex — ChatGPT Pro $100–$200/month (A $100/month tier was added in April 2026. A counter to Anthropic's Claude Max $100/month)

  • Claude Code — Claude Max $100–$200/month

In other words, even the basic $20/month plan includes that vendor's coding agent.If you are starting fresh, the cost is the same whether you are a ChatGPT fan or a Claude fan.

The quality of details and the order in which the latest features arrive will vary depending on the time.It is not a situation where one is clearly superior to the other in terms of functionality.

So, why recommend Claude Code as the second option?Ultimately, there is only one remaining difference.

The nature of the engine (model) is different.Codex is based on the GPT series, while Claude Code is based on the Claude series. Even if you send the same request, the output will differ in writing style, reasoning, and the points raised.

This one point will come into play in the next section (2-3).

2-3. Why keep it "interchangeable"—What Fable 5 demonstrated

If the features and pricing are almost the same, the choice is simple.
Use the coding agent included with the conversational AI you are already subscribed to. If you pay for ChatGPT Plus, use Codex; if you pay for Claude Pro, use Claude Code. That is all.

However, you need to keep the one you choose in a state where it can be swapped out at any time. The reason is simple:
AI services can suddenly become unavailable one day.

Here is one real-life example.
On June 9, 2026, Anthropic released "Claude Fable 5" to the public (the higher-end version mentioned at the same time, "Claude Mythos 5," is a limited release for approved customers).
However, immediately after that, access to Fable 5 and others was restricted due to moves to comply with export regulations, and it has not been fully lifted as of the time of writing this article. Although it is highly likely that it will be lifted eventually, a state where it is unusable for several days after release has actually continued for a certain period.

This is not limited to Anthropic.
It is not strange for a similar situation to occur at any time with OpenAI's Codex or Google's Gemini in the current situation. Due to price revisions, model changes, export regulations, government policies of various countries, and changes in school usage policies, the way you use AI services can suddenly change one day.

If your school work workflow is in a state where it "only works with Claude Code" or "only works with Codex," your work will stop on that day.
Therefore, you must choose what you can use from among multiple engines and keep it in a state where you can switch to another engine at any time. This is necessary.

2-4. Three practices for making it interchangeable

Even if you say "interchangeable," you don't need any special mechanisms. Just these three things.

  1. Do not make your folder structure dependent on AI

    • Organize it in a structure that makes sense to humans, such as "Classes/," "School Affairs/," and "Student Guidance/"

    • Do not separate them into "Codex working directory" and "Claude Code working directory"

  2. Write prompts at a granularity that works for both

    • Do not write in a way that assumes features exclusive to Claude Code

    • If the instruction is at the level of "Read the past 6 communications and create this month's version," it will work regardless of which one you send it to

  3. Keep your AI instructions in a format that both AIs can read.

    • Each AI has a mechanism to read instructions within a folder (files named `CLAUDE.md` for Claude Code and `AGENTS.md` for Codex).

    • If the content is difficult, don't overthink it. Just ask the AI you are using, "Make this content usable in Codex as well," or "Make this work in Claude Code too,," and the AI will organize it for you.

If you keep these three points in mind from the start, even on a day when an incident like Fable 5 occurs, you can switch to another engine and continue your work.
Don't use two at the same time; build your setup so you can choose one and always escape to the other. This is the practical solution as of June 2026.

2-5. If you are a Gemini user, Antigravity is also an option.

For those who prefer Google, I will provide one more option.

Announced by Google in November 2025, "Antigravity" is anagent-driven IDE powered by Gemini 3 Pro.

In a nutshell, it aims to operate witha goal presentation that is one level more abstract than Claude Code or Codex.
It is positioned as a new paradigm for non-engineers where multiple agents work in parallel based on instructions like "make a chat app."new paradigm for non-engineers

However, for a teacher's daily tasks (communication, comments, lesson preparation), Claude Code and Codex currently have the right level of granularity.
Antigravity is a tool that is effective in the "I want to build a slightly larger-scale system" phase.

Judgment criteria:

  • If you already use Gemini as your conversational AI or work primarily within Google Workspace, it is perfectly fine to start with Codex/Claude Code and try Antigravity once you get used to it.

  • Take a look when you feel like, "I want to have an AI manage agents and take on larger tasks."

The details are beyond the scope of this article, so it is enough to just remember that it exists for now.


Chapter 3: Why Cursor and Obsidian "Save Beginners"

This is the part I most want to convey in this article.

3-1. The biggest point of frustration is "until you install and get used to it"

When I recommend Claude Code or Codex, most teachers drop out "before file operations."

  • The terminal is black and scary

  • English appears during the installation process, and I get lost

  • If I don't touch it for a week, I forget how to even launch it

Before "AI is convenient," the reality is thatthe habit of simply launching an app disappears in three days.

And then, the next wall that teachers who finally manage to launch it hit is another one.

3-2. The Next Hurdle: With Claude Code or Codex Alone, 'You Can't See the Files'

Once people get it running, they ask the same questions in the first few days.

'What just happened?', 'Where was the file created?', 'Where did my previous text go?'

This is not a problem of beginner comprehension, but a structural issue where the 'place to look' varies for each AI.

Both Codex (by OpenAI; Web, CLI, IDE extension, iOS app, etc.) and Claude Code (by Anthropic; CLI, IDE extension, Desktop app, Web browser version, etc.) have features to open files, view their contents, and edit them. In other words, it is not that the files are completely invisible with each AI app alone.

However, the screens are different.

  • View the results of work done in Codex on the Codex screen

  • View the results of work done in Claude Code on the Claude Code screen

  • As written in Chapter 2, the realistic solution is to choose one and keep it swappable. If you switch, the place where you view the results changes every time

  • For those who subscribe to both and touch the same folder in parallel, the screens end up scattered in two places

In short, for beginners, 'What just happened?' is likely to occur, and for those using both, 'Which screen has the latest version?' is likely to occur.
If there is one common 'place to look' that does not depend on the AI, this problem disappears entirely.

3-3. Cursor Serves as Both a 'Common Place to Look Regardless of AI' and for 'Detailed Editing'

This is where Cursor comes in handy.

Cursor is an AI editor for file editing, but it has two layers of value for faculty members.

First Value: You can view the results on the same screen regardless of which AI you are working with

If you keep your folder open in Cursor (the same location as the folder managed by Obsidian),

  • files rewritten by Codex,

  • files rewritten by Claude Code,

  • and files you corrected by hand

are all reflected on the same Cursor screen. You can also see the differences in color, showing which line was rewritten and how. You can see the contents of Markdown, text, and CSV files as they are (PDFs can be previewed with extensions. Manual editing of Excel, Word, and PowerPoint assumes the original apps).

For example, if you ask Claude Code to "write this month's grade-level newsletter" → Claude Code rewrites the file in the background → You can see the file being rewritten on the Cursor screen.
This alone changes the feeling from "leaving it to AI" to "working together with AI."

If you are subscribed to both, it works even more effectively. For example,while writing a newsletter with Codex, you can simultaneously organize comments with Claude Code, if you keep both Cursor and Obsidian open, you can see the respective edits overlapping on the same screen in real-time. You can tell at a glance "where exactly was just fixed."

And this connects to the "how to build it so you won't be in trouble even if you switch" approach written in Chapter 2. Since Cursor and Obsidian are "places to look" that do not depend on AI, even if you switch from Codex to Claude Code, the Cursor screen does not change.The only thing that changes is the engine running in the background.

Second value: Cursor itself comes with a built-in LLM (Pro $20/month)

When you subscribe to Cursor Pro ($20/month), you can choose from multiplebuilt-in LLMs within Cursor, including Claude, GPT, and Gemini.

What this is effective for isfine-tuning, such as "change the phrasing of just this one sentence" or "make the sentence endings in this paragraph consistent."

Claude Code and Codex are good at "heavy tasks that deal with the entire folder" via the terminal.
On the other hand, Cursor's built-in LLM is good at light tasks likeselecting a range and fixing the phrasing on the spot.

Practical example:

  • Ask Claude Code to "write a grade-level newsletter" to create the outline

  • Open it in Cursor and ask the Cursor built-in LLM, "make just this introduction a bit gentler in tone"

  • Done

Heavy tasks are handled by Claude Code/Codex, and fine-tuning of text is completed within Cursor.
Once you understand this, the back-and-forth between the terminal and the editor decreases significantly.

3-4. Obsidian is a text-based GitHub: Why it works for teachers and business professionals

Next isObsidian.

For engineers, the place to store code is GitHub.
What teachers and business professionals handle is not code, buttext.
Therefore,the storage place for text = Obsidian.

The features of Obsidian are:

  • Markdown format, lightweight and fast

  • Everything can be handled aslocal files (cloud synchronization is an option, but you are not tied to a specific AI company)

  • Notes can be connected to each other with `[[links]]` (described later)

  • Easy to operate from AI agents via community plugins and CLI integration

For those who feel that "creating links seems difficult":
Obsidian is famous for its feature of connecting notes with `[[links]]`, but you don't need to create links from the start. Simply dropping notes into folders is valuable enough. Once you get used to it, you can leave the linking itself to AI by asking, "Read the notes in this Vault and add `[[links]]` between related notes." "Link later, just accumulate for now" is perfectly fine.

For professions where the daily work is not "writing code" but "thinking in text"—such as educators, business professionals, writers, and researchers—Obsidian, with its high visibility and ease of structuring, has become a powerful option.

For educators, create one Obsidian Vault (folder) and organize it as follows:

  • "Lessons/" for subject-specific notes

  • "School Affairs/" for grade-level and administrative notes

  • "Student Guidance/" for guidance record formats

  • "Sounding Board/" for conversation logs with AI (described later)

—Everything is stored using this structure.

Then, open that Obsidian folder in Cursor as well.
Run Claude Code and Codex on that folder.

This establishes a workflow where "all notes are in Obsidian, and editing and AI tasks are done in Cursor + Claude Code + Codex."
Correcting drafts (markdown or text) created by AI while checking the diffs in Cursor—this cycle is difficult to establish with just a terminal.

Two promises before letting AI touch your work
① Do not let it touch the originals directly. Create a copied folder like "AI Work/2026-06_Grade Newsletter/" and run the AI only within that. Until you are used to it, do not allow deletions, overwrites, or batch conversions. Humans should reflect only the final version into the original Word or Excel files. ② Do not place Vaults containing personal information on personal clouds or personal PCs. "Local storage does not mean it is safe." As a rule, only allow AI to read anonymized data, teaching materials, or your own thought notes.

3-5. Storing all conversations with AI in Obsidian—"The Growing Second Brain"

And this is where the real meaning of using Obsidian begins.

In April 2026, famous AI figure Andrej Karpathy published a short note on GitHub called "LLM Wiki," which was widely read in the AI community.

To put it briefly,

Store all results of conversations, research, and brainstorming with AI as markdown files in Obsidian. Point Claude Code and Codex to that folder. It becomes easier for the AI to provide answers based on your past thoughts while searching and referencing necessary notes.

Karpathy himself simply calls it "LLM Wiki," but the community later superimposed the long-standing concept of a "Second Brain" onto this pattern and popularized it, which is the true nature of what is now called "the growing second brain" in the Japanese-speaking world.

Here is the specific way to use it for educators.

  • When you have a good interaction on ChatGPT or Claude → Finally, ask it to "Summarize this conversation into a single markdown file in a format that can be saved in Obsidian" and save it via copy-paste

  • For brainstorming with Codex/Claude Code, instruct it to write it out from the start, like "MyVault/Sounding Board/2026-06-18_Parent Communication.md"

  • Divide folders by domain ("Brainstorming/Class/", "Brainstorming/School Affairs/", "Brainstorming/Student Guidance/")

If you continue this for three months, you will be able to make requests like the following within Obsidian.

  • Based on all the past brainstorming sessions regarding parent communication in the Student Guidance folder, provide an initial response plan for this case.

  • Extract the questioning patterns I repeatedly struggle with from the social studies brainstorming logs in the Class folder.

  • Summarize the changes in my work style from the school affairs memos of the past three months.

Not general information from the internet, but extracting answers from the experience and thinking you have accumulated yourself—this is what a "second brain you nurture" is meant to be.

However, there are two conditions to achieve this.

  1. A place for local storage is required (=Obsidian)

  2. An AI agent capable of reading that entire folder is required (=Claude Code or Codex)

In other words, the 4-piece set of Cursor + Obsidian + Claude Code/Codex is not just a collection of convenient tools, but is built as infrastructure for nurturing personal experience together with AI.”
Once you understand this, it should make sense why you need these four items.

3-6. Why Obsidian alone is not enough

Finally, one common misunderstanding.
I am sometimes asked, "Isn't Obsidian alone enough?"

Obsidian is powerful, but there are parts that fall short of the reality for teachers.

  • PDFs can be opened with standard Obsidian, but the contents of Excel, Word, and PowerPoint cannot be seen by default (you can handle them with plugins, but there is the reality that it is difficult to freely install plugins on school PCs)

  • Most of the files distributed by the school are ultimately Excel, Word, or PowerPoint

  • Even if you ask an AI to "write this in Word based on the last six newsletters," it is difficult to open and check the final Word document with Obsidian alone

When you layer Cursor on top,

  • Markdown, text, CSV → You can use Obsidian or Cursor

  • PDF → Standard viewing in Obsidian. You can also preview in Cursor if you install extensions

  • Excel, Word, PowerPoint → Final editing is done in the original app. Claude Code/Codex can read and write the text content

  • AI editing results (text/markdown) → Diffs visible in Cursor

  • Minor corrections → Completed within Cursor's built-in LLM

This is how it works.
Obsidian is the home base for my knowledge, and Cursor is the window for all files coming from outside plus a place for fine-tuning. These are the two roles.



Chapter 4: Three Companies' Add-ins Are Here—AI Has Descended into Excel/Google Systems

The major development from spring to early summer 2026 is that all three companies—ChatGPT, Claude, and Gemini—have released AI that can be used within spreadsheet and document creation apps (though the forms differ: Gemini is built into Workspace, ChatGPT is an Excel/Sheets add-in, and Claude is a Microsoft 365 add-in).
Five months ago, Anthropic and OpenAI were still focused on "browser chat."
Now, it runs directly inside Excel, Google Sheets, and Google Docs.

I will look at them in order.

4-1. ChatGPT for Excel / Google Sheets (May 2026 GA, OpenAI Official)

On May 5, 2026, OpenAI officially released the Excel and Google Sheets add-ins for general availability (ChatGPT for Excel and Google Sheets — OpenAI Help Center).

There are only two points to identify the official version.

  • Google Workspace Marketplace: App name "ChatGPT", developer "OpenAI"

  • Microsoft Marketplace: App name "ChatGPT", creator "OpenAI, LLC"

Third-party add-ons with similar names like "ChatGPT for Google Workspace" or "GPT for Sheets and Docs" are listed, but please identify them by the developer name.

What you can do:

  • Create tables from scratch using natural language / Read and summarize existing tables

  • Analyze data spanning multiple tabs and sheets

  • Create formulas, explain the meaning of formulas, and fix broken formulas

  • Organize duplicates and inconsistent notations, and aggregate submissions

  • Generate drafts for rubrics, budget tables, and progress management charts

Pricing: Available on all plans including Free, Go, Plus, Pro, Business, Enterprise, ChatGPT Edu, and K-12 (Free and Go have usage limits). Business, Edu, and K-12 have a free preview until June 2, 2026.

For faculty members who already have ChatGPT Plus ($20/month), this is the most natural option to start using without additional fees.

4-2. Claude for Excel / Word / PowerPoint (Officially released in Spring 2026)

On May 7, 2026, Anthropic officially released "Claude for Microsoft 365," an add-in that allows you to call Claude directly from Excel, Word, and PowerPoint (a beta version is available for Outlook).

While the ChatGPT version also covers Google Sheets, the Claude version is characterized by its focus on the Microsoft 365 ecosystem (Excel, Word, PowerPoint + Outlook in Beta). It also has a reputation for
being stronger in terms of readability for long texts and precision in proofreading,

  • Opening a Word document for a grade-level newsletter and saying, "Proofread this to match the writing style of the last three issues"

  • In a PowerPoint for a field trip, saying, "Create three supplementary slides for parents"

  • In a Word document for evaluation comments, saying, "Write these separately based on the three perspectives of proactivity, thinking, and knowledge"

—the Claude version performs a step better in tasks that involve polishing long Japanese text.

4-3. Gemini in Workspace (Major update on March 10, 2026)

Furthermore, on March 10, 2026, Google released a major update for Gemini integrated with Google Docs, Sheets, Slides, and Drive (Education Plus / Teaching and Learning licenses were covered in a separate rollout in February 2026).

This might be the biggest change in the landscape this time.

Main new features:

  • Automatically create documents tailored to the user's writing style by referencing Gmail and existing files

  • Build spreadsheet tables and slides using only prompt instructions

  • Summarize and answer questions by searching across files in Google Drive

There are many ways for teachers to use this.

  • Google Docs: Reference past grade-level newsletters to create a draft for this month's issue. Learn the writing style of your Gmail to standardize the "template" for parent communications (Handle documents containing personal names or specific circumstances only in environments approved by the school or municipality)

  • Sheets: Use "Fill with Gemini" to bulk-complete empty cells based on context. You can ask in natural language for monthly absence trends, grade tendencies by class, or to identify correlated items

  • Slides: Ask it to "Create a 6-slide presentation for the class observation day based on last month's class newsletter"

  • Gmail: Have it generate "3 drafts in polite honorifics" for parent communications

  • Drive Cross-Search: Ask, 'Collect and summarize notes related to truancy support from the grade-level folder.'

In terms of a teacher's daily routine,

  • Grade Analysis: Show a table of raw scores and ask in natural language, 'What are the weaknesses by unit, the polarization within the class, and the trends appearing by gender?'

  • Timetable Rebuilding: When progress is off due to school events, ask, 'Rebuild the special timetable for next Monday based on current progress.'

  • Classification of Free-form Responses: Ask to 'Categorize class survey responses into three layers: students in trouble, students who want to grow, and students who are dissatisfied.'

  • Field Trip Preparation: Generate itineraries, budgets, packing lists, and group-based checklists all at once.

Pricing/Eligibility: Eligible for Education Plus or Teaching and Learning upgrade licenses. Since free upgrades for students vary by time and region, please check the official Google page for each country for the latest eligibility criteria.

If your school or municipality uses Google Workspace for Education and you use Google Classroom on a daily basis, Gemini in Workspace is your top candidate.
However, rather than opening all features to students at once, it is safer to have teachers use it for drafting and verification first, and then open it to students after safe operation is established.

4-4. Decision Criteria for Teachers (Updated Version)

If you are unsure, ask yourself the following two questions in order.

Q1: Which AI do you talk to most often?

  • ChatGPT → ChatGPT for Excel / Sheets

  • Claude → Claude for Excel / Word

  • Gemini → Gemini in Workspace

Q2: Is your school an Excel-based or Google-based environment?

  • Excel-based → ChatGPT version or Claude version

  • Google-based (Workspace for Education) → Gemini in Workspace or ChatGPT version (Sheets compatible)

Golden Rule: Use only one company. Do not use multiple.
Efficiency in school administration is determined not by 'how many you subscribe to,' but by 'how thoroughly you use them.'
For the first 1-2 months, use one service thoroughly, and only consider a second one after you feel its limitations. This is the realistic solution.

4-5. Add-ins are for 'within one file,' while Claude Code and Codex are for 'the entire folder'

One point of caution.
Add-ins are very convenient, but they are not a substitute for Claude Code or Codex.

The difference in roles is easy to understand if you think about it in terms of distance.

  • Add-ins = a secretary that works within the single file you currently have open. They can fix the attendance book or the Word document you are currently working on right there.

  • Claude Code/Codex = a partner that looks over the entire folder. They can perform cross-file tasks, such as "read the last six newsletters before writing this month's edition" or "collect and summarize records related to absence handling from the entire grade folder."

For example:

  • "Calculate the absence count from the Excel attendance book in front of me" → completed with Add-ins

  • "Write this month's grade newsletter based on the last three months' newsletters, while also specifying the placement of photos" → work for Claude Code/Codex

One file in front of you = Add-ins,
tasks spanning multiple files = Claude Code/Codex.

Once you understand this distinction, you won't have to wonder which one to rely on each time.

4-6. Never compromise on the handling of personal information

No matter how convenient add-ins are, you must never carelessly input student names, grades, attendance, family circumstances, health information, or support information. Even with official add-ins, since they read and process the contents of Excel or Sheets, you should be cautious about how you handle them.

The line drawn here is a "Guidelines for the Use of Generative AI in Elementary and Secondary Education Ver. 2.0" by MEXT, "Points to Consider Regarding the Use of Educational Data," and the operational rules of each Board of Education (such as the Tokyo Metropolitan Government's "Guidelines for the Use of Generative AI in Metropolitan Schools"), serving as a safe benchmark.
Always follow the rules of your workplace and local government for final decisions.

✅ Relatively easy to handle

  • Textbook scope, public teaching materials, general lesson plans, dummy data

⚠️ Handle with conditions

  • Event item lists that do not contain personal information, anonymized surveys, summary tables for creating teaching materials

❌ Do not handle outside of approved environments

  • Student names, grades, attendance, family circumstances, individual guidance records

❌ Handle with extreme caution

  • Highly private information such as health, disabilities, support, bullying, school non-attendance, and family economic status

It is safest to get used to them in the following order.

  1. Creating teaching materials (creating unit-specific problems, rubrics)

  2. Anonymized survey aggregation (categorizing free-text responses)

  3. Event preparation materials and accounting sheets (items unlikely to contain personal information)

  4. Creating aggregation sheet templates with dummy data

Only input real data after confirming the operational rules of your school/municipality and the available AI environment.
Rather than what you can use, the line between what you delegate and where humans must verify is the most important aspect of using AI in schools. This is the most critical point for using AI in a school setting.



Chapter 5: Voice Input as a "Common Gateway"

I have listed the tools so far, but there is a gateway common to all of them.a gateway common to all tools.

It is voice input.

5-1. As of June 2026, there are three leading voice input tools

For teachers, the biggest bottleneck with AI is that "it is a hassle to type the text to give to the AI yourself."
Between classes, during breaks, or in the 10 minutes after school, there is almost no time to sit at a desk and carefully write prompts.

As of June 2026, the top group of AI voice input tools has narrowed down to three. Until recently, it was a two-horse race between Aqua Voice and Superwhisper, but with the release of Typeless in November 2025, it has become a three-horse race.

  • Aqua Voice: Fast startup and fast text insertion. Pro starts at $8/month (billed annually). Supports Mac and iPhone (iOS version released in April 2026). The "Instructions" feature allows the AI to automatically format text.

  • Superwhisper: Supports offline processing using the Apple Silicon (M1/M2/M3) Neural Engine. Lifetime license available for a one-time payment of $249.99. Suitable for those who prioritize privacy and dislike recurring costs.

  • Typeless: Its strength lies in its broad compatibility, spanning Mac/Windows/iOS/Android (released in November 2025 by YC-backed Simply CA LLC). The free tier offers 8,000 words per week, and Pro is about $12/month billed annually. It uses cloud processing and excels at filler removal and self-correction detection.

The guidelines for choosing between them are as follows.

  • If you focus on Japanese-language tasks and Apple products, Aqua Voice or Superwhisper

  • If you want to use Windows alongside or across multiple devices, Typeless

  • If you have to choose just one for personal use, Aqua Voice Pro (from $8/month) is the most recommendable option as of June 2026.

5-2. Voice input is not "4x faster typing," but "4x easier verbalization of thoughts"

The true value of voice input lies not in speed, but in being able to get what's in your head out through dictation.

With typing, you edit as you write, which makes the text stiff.
With voice, you get your "raw thoughts" before any editing kicks in.
AI actually finds it easier to read the context of unpolished, spoken language.

As the ChatGPT Research Institute points out, if you talk in detail about the background and requirements for 3 to 5 minutes via voice before feeding it to the AI, the quality of the AI's output improves dramatically.
This is the moment it shifts from "the effort of writing prompts to the effort of generating ideas."

The important thing is not to try to speak perfectly. It is perfectly fine to keep talking while including words that come out as you think, like "let's see," "um," or "oh, that's not quite right." The AI will sort it out for you in the background. If the result you get is not quite right, just say, "no, that's not what I meant," and rephrase it. That's all there is to it.

5-3. The premise is that "fragmentary notes are fine"

In the world five months ago, the premise was "let's write memos properly in full sentences."
That is no longer the case.

Since the AI will transcribe, summarize, and structure them, fragmentary notes are fine.

  • Voice input to smartphone: "For the next social studies class, I want to introduce World Heritage sites with a photo quiz. Also, I rushed the explanation at the end of today's homeroom. I'll supplement it tomorrow morning."

  • After getting home, tell Claude Code: "Organize today's voice memo into class improvement notes and a plan for tomorrow morning."

  • By the next morning, you have a "today's plan memo" that you can review in one minute before heading to work.

At first, it is safest to start in areas that do not contain personal information. Class reflections, teaching material ideas, event planning, and your own training notes—these are the easiest entry points to try out.

Do not leave information regarding students' real names, grades, family situations, health, or support needs in voice input outside of environments approved by the school or local government. Please also be aware that personal information remains in the recorded data itself.

Instead of high-hurdle suggestions like "write a daily report in Notion," voice → smartwatch recording → LINE to yourself → memo app—any of these can be an entry point.

Not raising the hurdle is the biggest trick to keeping it up.


Summary

  • In June 2026, the AI landscape has completely moved past the stage of "which one to choose: ChatGPT, Claude, or Gemini"

  • A teacher's starter kit is (either Claude Code or Codex) + Cursor + Obsidian + (one for conversation). For Gemini fans, Antigravity is also an option

  • Beginners hit a wall in two stages. The first stage is installation and habituation, and the second stage is the invisible wall of file management. Cursor solves the latter

  • Cursor works in two roles: a monitor that shows the AI's work and a built-in LLM for detailed editing.

  • Obsidian is a text-based GitHub. It is most effective for text-centric work by teachers and business professionals

  • When you store conversations, brainstorming, and research with AI in Obsidian, the second brain you cultivate, popularized by Karpathy's LLM Wiki, becomes a reality

  • Claude Code and Codex are almost symmetrical in function and pricing. The realistic solution is not to use both constantly, but to choose one and set it up so it can be swapped at any time (as with the Fable 5 case, there is always a possibility that it could suddenly become unusable one day)

  • Excel and Google-related add-ins have been released by all three companies: ChatGPT, Claude, and Gemini. Install the official version of only one company—the AI you use the most

  • Voice input is in the era of the big three: Aqua Voice (personal recommendation) / Superwhisper (one-time purchase) / Typeless (multi-OS).

AI is not a substitute for judgment. It is support for a system that prevents hesitation.

The goal is not to increase the number of tools. The goal is
to cultivate your own experience together with AI.



Things to do tomorrow (3 items)

  1. Create one folder for AI work. Only put copied materials in it, not the originals. Set that folder as the 'MyVault' in Obsidian, and open the same folder in Cursor (free or $20/month Pro)

  2. Save just one impressive conversation you've had so far on ChatGPT or Claude in Markdown format (you can just ask, 'Summarize this conversation in a Markdown format that can be saved in Obsidian')

  3. Launch the coding agent included in the conversational AI you are already paying for tonight. If you pay for ChatGPT Plus ($20/month), you have Codex; if you pay for Claude Pro ($20/month), you have Claude Code. If you don't pay for either, start by subscribing to just one conversational AI service. The first tasks to delegate are tasks that do not contain personal information (drafting class handouts, event supply lists, categorizing anonymous surveys). At the same time, install just one official Excel/Sheets add-in for the conversational AI you are using.

Once you get used to it, adding voice input (choose one from Aqua Voice, Superwhisper, or Typeless) will instantly eliminate the effort of typing out text to send to the AI.


The terms of service and privacy policies for the AI services introduced in this article
vary by company. Please prioritize your school or local government guidelines and
exercise sufficient caution when handling personal information.
The prices and features listed are as of June 2026.


いいなと思ったら応援しよう!