DEV Community: Takashi Abe

Obsidian AI Exporter: The Current Evolution Supporting Multiple AIs

Takashi Abe — Fri, 20 Mar 2026 01:51:03 +0000

Opportunities to interact with generative AI like Gemini and Claude in daily life have been increasing. When doing research or brainstorming, you often find brilliant ideas that make you go "aha!" during exchanges with AI, right?

But the moment you close your browser tab, those precious conversations flow away into the depths of your history. I wanted to keep my dialogues with AI as my own "knowledge asset" that I could search and link to. With that thought in mind, I released the Chrome extension "Obsidian AI Exporter" two months ago.

Thanks to everyone, many people have started using it, and over these two months I have received bug reports and feature requests and have made various updates. Today, I want to share with you the new features of this greatly evolved extension.

What Kind of Tool Is This?

Obsidian AI Exporter is an extension that extracts conversations with AI in your browser with one click and saves them as clean Obsidian documents.

It originally started as Gemini-only, but since I personally use multiple AIs, it now supports all the generative AI services I use. These days, you don't have to stick to a single platform.

Three Major New Features Added

Over the two months since release, it has evolved to be more flexible and more convenient.

1. Expanded AI Platform Support

In addition to Gemini, it now supports major AI services, including Claude, ChatGPT, and Perplexity. No matter which AI you are conversing with, just press the "Sync" button that appears in the bottom right of the same screen to save your entire conversation or research results.

2. Support for Saving Methods Other Than Obsidian

Although it bears the name Obsidian, you no longer necessarily need to use Obsidian. By configuring the Markdown function from the settings screen, you have the option to download as a pure Markdown file to your computer or copy to clipboard and paste into another editor.

3. Appending to Existing Notes and Adding Unique Features

An append mode has been added to write only new conversations to existing notes. Furthermore, fine adjustments tailored to each AI's characteristics have been made, such as saving the report portion of Gemini's Deep Research and the collapsible display of web search results in Claude.

It also supports the LaTeX format, so formulas are properly recorded.

Keep Your Important Data in Your Own Hands

What this extension is committed to is balancing privacy and convenience. Without going through external servers, all processing is completed in your browser and local environment. SSL is supported, but Obsidian-side configuration is necessary.

The more AI becomes our excellent partner, the more the dialogues exchanged there should become irreplaceable assets. Before they fly away and disappear, try pulling them to your hands with one click.

I am placing links to the Chrome Web Store and GitHub below. I would be delighted if you could install it and try out how it works.

Chrome Web Store:

Obsidian AI Exporter - Chrome Web Store

Export AI conversations from Gemini, Claude, ChatGPT, and Perplexity to Obsidian via Local REST API

chromewebstore.google.com

sho7650 / obsidian-AI-exporter

Obsidian AI Exporter

Chrome Extension that exports AI conversations from Google Gemini, Claude AI, ChatGPT, and Perplexity to Obsidian via the Local REST API.

日本語版はこちら

Features

Multi-platform support: Export from Google Gemini, Claude AI, ChatGPT, and Perplexity
One-click export: Floating "Sync" button on supported AI pages
Multiple output options: Save to Obsidian, download as file, or copy to clipboard
Deep Research support: Export Gemini Deep Research and Claude Extended Thinking reports
Artifact support: Extract Claude Artifacts with inline citations and sources
Tool content support: Optionally include Claude's web search results and tool activity as collapsible [!ABSTRACT] callouts
Append mode: Only new messages are added to existing notes
Obsidian callouts: Formatted output with [!QUESTION] and [!NOTE] callouts
YAML frontmatter: Metadata including title, source, URL, dates, and tags
Auto-scroll: Automatically loads all messages in long Gemini conversations
Platform-based organization: Use {platform}…

View on GitHub

I plan to continue adding convenient features little by little. If you have impressions from trying it or requests for features you want, please feel free to leave a comment on a GitHub issue or note.

AI Era: The Power of Writing Meditation - Connecting with Notebooks

Takashi Abe — Wed, 11 Mar 2026 01:34:12 +0000

Lately, I have been feeling a bit overwhelmed by the evolution of AI.

My reliance on AI has made my desires unmanageable. I can immediately start on ideas I have and let troublesome tasks be handled automatically. I feel it is wonderful that getting advice has lowered the hurdle to taking the first step.

However, it is also true that when I pause to reflect, moments of feeling lost have increased, wondering what exactly I am doing all this for.

During such times when I am about to be swallowed by waves of information and ideas, what saves me is not a digital screen but an old-fashioned analog notebook.

I pour my heart into a blank notebook, reconnecting with my life and purpose. What should I be doing now? What should I be thinking about? Should I calm my mind and deal with the task at hand, or should I move toward a future vision? Above all, I need to make a decision about what not to do.

When it comes to these things, no matter how excellent an AI I ask, answers simply will not come back. Unless you are going to walk a life decided by AI, if there is a path you want to walk on your feet, you absolutely need time to open your notebook and deeply face yourself.

For me, this time of self-dialogue is similar to meditation.

Time spent at my desk endlessly writing out characters and time spent slowly running while letting my thoughts wander. Both are beloved moments that I cannot do without.

However, when thoughts are just circling in my head, working memory fills up, and I am driven by the urge to write them out somewhere. That is precisely why a notebook as a receptacle for thought is optimal.

You might think you cannot take notes while running.

That is certainly true, and I have not yet found a good way to record things while running. But I have come to interpret that ideas I have thoroughly thought through while running and still vividly remember when I get home are what is truly important to me right now.

I may forget trivial triggers, but essential things will surely remain in my heart. Believing that, I only note down the crystals that remain after finishing my run and connect them to the next step.

I sit at an empty desk, staring at my notebook, tirelessly crafting characters. I love this tranquil time.

We intentionally pause in a time of abundance and convenience. Writing characters in a notebook allows you to meditate and listen to your own heart's voice. Would you like to try it once as well? I highly recommend it.

Accelerating AI and the Courage to Pause: Reclaim Your Core Values Through Essential Thinking

Takashi Abe — Mon, 02 Mar 2026 05:57:02 +0000

Lately, I have been hearing the term "AI fatigue" everywhere.

Bold claims are flying around that AI will change the world and that conventional tools are no longer needed. However, behind all this, I feel that the amount of activity each of us needs to handle has actually increased to an abnormal degree.

When I look back on my past personal development experiences, the pace was much slower. For someone like me who does not code as a main profession, touching programs every day is not easy. Creating one class a day was enough, and as the term "weekend programmer" suggested, development was something that progressed little by little, taking time.

But now, with the help of AI, I have three coding projects running simultaneously. New features are added every day, and issues that come up can be resolved within the same day. This is a speed that my former self could never have imagined.

However, when I stop and reflect, I begin to wonder if this is really what I wanted to do.

I started using AI to save time and improve efficiency, but before I knew it, I was being chased by the outputs that AI keeps producing one after another. To keep up with AI, I have been forcing myself to do unnecessary things. As a result, many outputs have been produced, but my mind has been tossed around and exhausted by them. This is not the future I wished for.

When it comes to work, I have found a certain answer for how to engage with AI. My goal is not mere efficiency but improving quality. As long as I keep that in mind, AI functions as a very reliable partner.

The problem is how I spend my private time. Being driven by code that AI has written or analysis reports that AI has organized feels somehow wrong.

So I decided to return to the concept of essentialism. What is truly important to me? In which areas can I use AI to genuinely enhance my motivation and results? I need to reassess these things.

It is natural to take responsibility for what you have created, but you must not be crushed by that responsibility and lose sight of your original purpose. Being controlled by something you started as a pastime or hobby is putting the cart before the horse.

Therefore, to protect my own time, I established new rules.

First, I place what is essentially important to me at the center of my schedule. Then, for everything else that I could do but is not urgent, I confine it to specific time blocks.

Specifically, I dedicate my morning hours at the start of the day to my most mentally demanding important work and my core activities. I make sure to finish essential tasks within the morning. Then, I allocate tasks that do not require as much concentration and hobby-related activities to afternoon time blocks. If they do not get finished within that time, I simply carry them over to the next day.

In the past, being controlled by other people was a source of frustration, but now that may have simply transformed into being controlled by AI.

Precisely because I have obtained such a convenient tool, I want to have a clear sense of what I want to do and what I should do. Rather than matching myself to the speed of AI, I want to incorporate AI into my own comfortable rhythm. In this way, I intend to continue protecting my own way of moving forward.

When I Hired Multiple Generative AIs as Subordinates, Parallel Development Processing Never Stopped

Takashi Abe — Wed, 11 Feb 2026 15:00:00 +0000

Today is a relaxed introduction to recent Shossan updates.

1. Major Update to Obsidian AI Exporter

The tool I have been publicly releasing for saving Gemini chat results to Obsidian, called Obsidian AI Exporter, has been significantly powered up.

In my previous introduction, it was Gemini-only, but now it supports four sites: ChatGPT, Claude (including Deep Research), and Perplexity. In addition to Obsidian integration, you can now save as pure Markdown files or copy to clipboard.

Obsidian AI Exporter - Chrome Web Store
Export AI conversations from Gemini, Claude, ChatGPT, and Perplexity to Obsidian

While I think it will be in the maintenance phase unless there are major UI changes, I am still looking to add more useful features. If you have any feature requests, please let me know!

Currently, I am considering automatic tagging, expanding the append feature, which is still not quite there, and Wiki Link support for Obsidian.

2. Building Your Own News Curation Site with n8n

I wanted a site where I could read the latest overseas news summarized in a format convenient for me, so I developed my own curation platform.

The mechanism uses n8n to fetch updates from specific RSS feeds and URL lists, then an LLM (generative AI) translates and summarizes the content into Japanese. This is then registered to a display app I built with Docker.

By managing specific news URLs as unique keys, it prevents duplicate posts with a simple mechanism. The main body is simple and just displays content, but I designed it with loose coupling so that LLMs and feeds can be easily changed. I have also published the source code, so please take a look if you are interested.

Around v2, the check/ingest/post article parts are the developed app portions.

I named it tech-news as is since it is for personal use, but really the content to post can be anything as long as it is Markdown, so it is like a simple blog. Since I am developing it component by component for easy reuse, I am feeling like it has become more versatile than I expected.

3. Revamping Home VPN Environment with Tailscale

On the infrastructure side, I am organizing my home access environment.

Tailscale | Secure Connectivity for AI, IoT & Multi-Cloud
The connectivity platform for devs, IT, and security teams.

I had been using the VPN function of my Yamaha router, but considering security and convenience, I am currently introducing and configuring Tailscale.

Being able to safely access home resources via NAT traversal without opening VPN ports externally is safe and comfortable. The app is also easy to use, and functionally very detailed settings are possible. I have no idea why it is free for up to 3 users.

Currently, I am configuring a router with Docker on my home Linux server and managing settings with Ansible. An environment is coming together where I can work without being conscious of the boundary between home and the internet even when out.

From the Joy of Researching to the Joy of Realizing

Running multiple projects in parallel like this, I can feel the changes in development style in the AI era firsthand.

In the past, I found joy in the trial and error of researching technologies one by one. But now, I find it fun to watch my ideas take shape one after another, using multiple generative AIs as my subordinates. Is this a change due to age, or is it new excitement brought about by technological evolution?

The image of what I wanted to do becomes realized with the help of AI. Grinning at that process, I am giving instructions to my new subordinates again today.

How to Transform 3 Years of Thinking into Assets: Accelerate Feedback with Claude Cowork

Takashi Abe — Wed, 11 Feb 2026 13:30:16 +0000

Claude Cowork: The Ultimate Engine for Objectively Viewing Your Own Thought Evolution

I had been pondering whether Claude Cowork actually has any practical uses.

Then I discovered yet another new way to use it. Cowork is a work companion that autonomously executes multiple tasks based on massive amounts of data that surpass what chat can handle.

Claude Cowork: The Ultimate Engine for Objectively Viewing Your Own Thought Evolution

As you know, Claude Cowork is not merely a chat tool. It is an AI agent that directly accesses your local file system (in my case, Obsidian) and autonomously completes tasks.

You can either place Obsidian in a folder that Cowork can access, or use the combination of Local REST API + Obsidian MCP Tools to have Cowork retrieve information via MCP—both approaches work effectively.

What makes this tool truly disruptive is its ability to execute against three years of my living thought data with precision and speed comparable to the Observe and Orient phases of the OODA loop (Observe, Orient, Decide, Act). Sounds pretentious, right?

Enabling High-Dimensional Differential Extraction Across Time—Something Difficult to Do Alone

Why bother using Claude Cowork? Because I know how vague human memory is and how conveniently it gets rewritten.

Breaking Through the Limits of Observe: It is humanly impossible to extract every single description on a specific theme from three years of notes numbering in the thousands. Information obtained through word searches is only a fraction. You need to extract information while considering intent as well. Cowork autonomously repeats searching and browsing, comprehensively observing all data.

Deepening Orient: The thinking habits of my three-years-ago self versus my current self, the subjects of interest, and changes in word choice. By having AI identify these differences, I can confront my own growth or stagnation as objective facts that I could never have noticed on my own.

This is not for reminiscing about the past, but for gaining an advanced metacognitive environment to use my past self as a stepping stone toward new ideas.

This is an effective environment that I could not achieve simply by using Obsidian as RAG with Dify. What was limited to merely retrieving information has now evolved into comprehensive information retrieval and analysis.

Using Three Years of Obsidian as Fuel to Gain Insights for the Future

I have actually shared my three years of Obsidian data with Claude Cowork and am running workflows like the following:

Step 1 (Observe): Scan all notes containing my views on generative AI from 2023 to the present (2026).

Step 2 (Orient): Analyze chronologically at which points and through which technological breakthroughs my thinking changed.

Step 3 (Insight): From that trajectory of change, have it present blind spots in thinking that I should focus on in the next year.

[Image of spinning the OOA loop with Cowork]

Previously, doing this required spending several days reading through past notes. However, with Claude Cowork, while generating intermediate analysis files in the working folder, it writes up a map of my own thinking in just minutes. This is not mere summarization—it is prototyping for the future through dialogue with my past self.

This is an efficient and effective way to reuse PKM (Personal Knowledge Management).

Continuously Updating Your Thinking Together with an AI Colleague

In these uncertain times ahead, there is no time to cling to past success stories. On the other hand, data about what you thought in the past and how you experimented is your own unique asset that belongs to no one else.

Stop using Claude Cowork as just a summarization tool.

Cowork is literally a partner for increasing the speed of your thinking and running the OODA loop faster than anyone else. Whether you let three years of data sleep as mere records or turn it into a power plant for your next ideas—that key depends on how you master Cowork as an agent.

I will continue to gaze into this abyss of the ultimate mirror and refine my thinking day after day.

The Illusion of My Own Efficiency: How Claude Exposed My Arrogance and Lack of Planning at Night

Takashi Abe — Fri, 06 Feb 2026 07:21:29 +0000

Conversations with AI are not for improving skills, but for recognizing your own arrogance

I, Shossan, have a certain pride in having automated my environment reasonably well and controlling it efficiently. However, through analyzing my own shell history while conversing with the trending Claude Cowork, what was thrust before me was an image of myself—utterly unplanned and full of biases. How embarrassing.

Being prompted by AI to improve is a shock close to defeat as an engineer. However, I am convinced that this very defeat is the only gateway to the next stage of growth. Gemini's words are too harsh.

People mistake unconscious repetition for effort

Why couldn't I notice my own inefficiencies? It's because a bias called familiarity had crept into my daily work.

This time, when I had Claude cowork analyze my past command history, the following three blind spots were exposed:

Collapse of time management: Activity peaked at 21:00, a time when I should have been resting. This wasn't diligence—it was merely repaying debt caused by lack of planning during the day.

Hollow automation: I manually executed ansible-playbook 31 times. While thinking I was using automation tools, the execution process itself was extremely analog and dependent on manual work.

Tool dependency: By using convenient tools (Keyboard Maestro, Stream Deck, Claude Code, etc.), I had arbitrarily decided that there was no more room for improvement.

The cold data-driven insights and the automation functions I immediately implemented

According to Claude's analysis, my activity is concentrated at 8:00 and 21:00. The fact that I was hammering away at commands at 21:00 after my bath is evidence of a bad habit that degrades sleep quality and diminishes the next day's performance. I truly never expected such results. I was genuinely surprised.

You should sleep after your bath. If you have to execute commands at that time, it means your planning has failed. Accepting this fact, I immediately adopted the improvement proposals that Claude suggested.

For example, here is a function for managing my custom LaunchAgents that start with com.oshiire.. Until now, I had been typing commands while trying to remember paths or searching through history, but now this svc function takes over that cognitive load.

# ~/.config/fish/functions/svc.fish
function svc --description "LaunchAgent service control"
  set -l action $argv[1]
  set -l name $argv[2]
  set -l plist ~/Library/LaunchAgents/com.oshiire.$name.plist

  switch $action
    case start
      launchctl load -w $plist
    case stop
      launchctl unload -w $plist
    case restart
      launchctl unload -w $plist; and launchctl load -w $plist
    case status
      launchctl list | grep $name
    case '*'
      echo "Usage: svc [start|stop|restart|status] [service_name]"
  end
end

Additionally, Claude proposed a short function called ap for ansible-playbook. This eliminates small decisions like argument specification mistakes and forgotten options. Such thoughtful consideration.

Looking into the mirror called AI, strategically updating yourself continuously

If you just end up praising AI's improvement suggestions as amazing, you are merely a consumer. For me, the learning from this experience lies in how to convert the negative of having AI point things out to me into the positive of solving through systems.

Abandon the assumption that you are doing well, and set aside time to regularly dissect yourself with objective data. Use AI not just as a code generator but as a behavioral analysis partner.

From here, I will continue to cut away waste, one step at a time. Why don't you also try exposing your terminal history and behavioral history—your embarrassing parts—to AI?

AI Era Survival Strategy: From Mimeograph to LLM. What Experienced Professionals Can Do About Technology's Black Box

Takashi Abe — Thu, 05 Feb 2026 04:16:09 +0000

Generative AI has truly enabled us to do so many things. I can no longer imagine a future where we live without being involved with AI.

This note is a perfect example. The content I am speaking through voice input right now is being rewoven into proper sentences by an LLM, which then drafts this note for me. AI is integral to both work and hobbies. Even the music production I need for my new DJ hobby can be handled by generative AI like Suno, which whips up several tracks in no time.

I no longer manually click through each Google search link when conducting research. I just ask Perplexity. At work, I use NotebookLM and Gemini Deep Research features to handle everything from analyzing massive amounts of documents to creating reports.

It may be too early to call this infrastructure, but thanks to generative AI, the areas I can cover by myself have expanded dramatically. By paying a small fee to an AI partner, I can now substitute tasks that previously required hiring specialists at a reasonable quality level. We live in such a world.

I have already outsourced the tedious work of writing code from scratch to AI. However, there is something we must not misunderstand here. Being able to use AI and having ability as an engineer are completely different things.

The Transition from Craftsman to AI Commander

Looking back, Japanese engineers used to be a group of craftsmen. During that era, people valued the long years spent honing specialized skills.

And now? By lining up excellent generative AI agents, a single person can give them instructions and produce deliverables spanning multiple specialized fields in an unbelievably short time.

I am constantly amazed by the innovations of our time. At the same time, I feel that what one person should do and can do has exploded.

From Mimeograph to TikTok: A History Walking Alongside Tools

When I was still in elementary and middle school, from the end of the Showa era to the beginning of Heisei, it was primarily an analog era.

We barely had word processors and printers, and I remember struggling to create decent printed materials. Come to think of it, my middle school graduation anthology was mimeographed. Young people today probably do not know what a mimeograph is.

In 1995, when I joined my company, that atmosphere lingered. Presentations meant writing on transparent OHP sheets with markers and projecting them. Right after joining, company training included analog protocols like how to address internal and external mail. Email addresses were something only a select few had.

By the end of the 20th century, we could write papers in Word and create slides in PowerPoint. Entering the 21st century brought further refinement, with demands for eye-catching slides using special animation effects. It also became an era requiring design sense from engineers, and I remember the struggle of needing yet another new skill. I was glad I did design work in middle school.

Then YouTube emerged, making it normal for amateurs to publish videos. Now with just an iPhone, you can livestream on TikTok. Even my daughter does regular streams.

Every time tools evolved, individual possibilities continued to expand. And now, with the emergence of generative AI, that expansion is reaching its limits. That is an exaggeration, Gemini.

Elevated Abilities and the Invisible Current Position

However, with all this convenience comes new worries.

In the narrow village society without the internet, comparison targets were only nearby, so the goal of how far you needed to go to be recognized seemed visible.

But now it is different.

We live in a world where you can see endless geniuses if you look up. Since generative AI can easily elevate your abilities, it becomes even harder to understand your true capabilities and growth. In a sense, it has become a cruel world where it is easy to put on air.

I was fortunate to experience the transition from analog to digital and older technologies firsthand.

Old machine language and assembly were 8-bit worlds. The number of instructions was small, and memorizing them was easy. It was possible to grasp everything.

However, with modern 64-bit instruction sets, developing anything with assembly alone is madness. Things have become far too complex, and the learning method of understanding and building everything from zero is no longer realistic. To produce results in a finite lifetime, we have no choice but to skillfully utilize the black box called AI and hone our skills in compressed time.

What We Veterans Can Leave for the Next Generation

AI will do it, so you do not need to know the internals.

This is correct in a sense, but it can also hinder growth as an engineer. In an era when you can create working things without knowing the internals, what should you do to deliberately acquire skills? This is an extremely difficult challenge to consider.

So what is the role of old-timers like us?

Simply passing down stories of past hardships has no meaning. What is important is conveying the background of how each thing and technology was created and how it works.

The logic behind the code AI spits out, the underlying principles beneath convenient frameworks. Given our intuitive understanding of these concepts, we should be capable of deciphering the hidden meanings and effectively sharing knowledge.

Not just using existing AI, but how do we create new technologies beyond that? As those who know the principles, we have come to need a certain responsibility in educating the next generation and driving innovation.

To Avoid Mass-Producing Working Garbage

Here, the role of people like us who know old technologies changes dramatically. It is not intended to lecture young people about learning assembly. The goal is to use that knowledge to detect code that appears correct but is actually inefficient and dangerous—working garbage produced by AI.

AI lies without hesitation. It proposes implementations full of security holes and presents architectures that wastefully consume resources. At such times, if you have low-level knowledge—the principles of how computers actually work—in your head, you can instantly detect discomfort, like this implementation feels wrong or this coding style will become technical debt.

This aesthetic eye is the only value in this era. It is impossible to beat AI in coding speed. However, in quality assurance and architecture design, humans still have an advantage. But that is a privilege only those who know the internals can have.

Veteran Survival Strategy: From Coder to Auditor

As a reminder to myself, I defined it this way. The job of engineers in the AI era is not creation but auditing.

Knowledge of assembly and memory management does not exist for nostalgia. It is a weapon to break through modern highly abstracted layers and expose the flaws AI tries to hide.

The black-boxing of technology will not stop. Therefore, instead of retelling old war stories, we veterans should delve into the complexity of this technology and analyze its workings.

It is fine to have an AI write code. Let's also maintain a relaxed attitude. However, the moment we can no longer ask why about that output, we stop being engineers and become mere AI operators.

I ask with self-reflection.
That code of yours—do you truly understand its contents before deploying it?

Why Computer Science is the Only Weapon for IT Architects in the AI Era

Takashi Abe — Tue, 03 Feb 2026 14:08:57 +0000

Sometimes, at tech conferences or networking events, ambitious young engineers ask me, “What should I learn next to truly level up my skills?”

I respect their drive. And because I respect them, I give them my most honest, unvarnished answer:

“Go back to school and study Computer Science (CS) systematically.”

That is it. In an era where fields like Data Science and Cybersecurity are becoming increasingly specialized, you cannot advance without the “foundation of all foundations”—CS. This choice isn’t just my personal preference; it is the cold, hard conclusion reached by many professionals fighting on the front lines of the industry.

1. Why Does Only the IT Industry Coddle the “Inexperienced”?

In any other professional field, it is a given that you cannot work without specialized knowledge. If you want to be a doctor, you need a medical license. If you handle hazardous materials, you need a certification. If you want a global business role, you need a proven language score.

However, the IT industry, particularly in Japan, often promotes the notion that experience is not necessary. or “Zero background? No problem!”

As a Principal Technical Architect at a major US cloud company with over a decade of experience, this trend deeply frustrates me. In any legitimate job description, fundamental knowledge is a prerequisite, and experience is what you build on top of it. Mastering the basics is a professional obligation.

I am not saying these words out of bitterness for the five years I spent studying at a KOSEN (Japan’s specialized 5-year engineering colleges). I make this statement because individuals who lack basic understanding can be a significant liability in the field. When a company or a team has to pay the “education cost” to fill that gap, the professional ecosystem becomes unhealthy.

2. The “Gold Watch” at IBM and the Crisis at a major US cloud company

Let me tell you a story from my past. When I joined IBM Japan, I was shocked to find that among the new graduates—who were given six months of intensive training—less than 20% truly understood Computer Science.

To me, the technical training felt elementary. Because I already had the technical foundation, I could focus entirely on sharpening my “human skills.” As a result, I earned the “Gold Watch”—a traditional prestige awarded to the top-performing new recruit at IBM Japan. Even at a global authority like IBM, that was the reality of the talent pool in Japan.

Now, let’s look at the current state of A major US cloud company, where I have spent the last 10 years. In the past, Salesforce was a sanctuary for “Business Engineers.” You were a star if you could demonstrate a single product such as Sales Application Services. But those days are over. Products have become infinitely more complex, requiring deep integration with legacy environments. Business knowledge alone is no longer enough to survive.

Recognizing this “danger zone,” I took it upon myself to develop a CS training curriculum for my colleagues. I mapped out over 50 essential items to help bridge the gap. While the company praised this initiative, it revealed a grim reality: even in a global giant, many employees lack a formal grasp of the basics. This is likely the case for many SIers (System Integrators) and consulting firms that have bloated their headcount without vetting for foundational skills.

In the U.S., a CS degree is like a passport for an engineer. Why does Japan continue to neglect the “foundation”? I simply cannot understand it.

3. Don’t Mimic Geniuses: Arm Yourself with Knowledge

Some might argue, “But I know famous, successful engineers who never studied CS!”

Those people fall into two categories. The first group consists of "handy IT uncles" who have successfully navigated small, niche markets through years of experimentation and learning.

The second category is the Geniuses. These are the rare exceptions who have the innate ability to grasp complex patterns intuitively.

For the rest of us, who are considered "ordinary" professionals, attempting to emulate them can lead to disastrous outcomes. Without a foundation of basic knowledge, application is 100% impossible. If you believe that AI will manage the fundamentals, you are mistaken. AI can handle tasks, but it cannot handle decisions. To understand the underlying mechanisms and take responsibility for your architectural decisions, a backbone in CS is indispensable.

If you are determined to thrive in this challenging and rapidly evolving industry, avoid shortcuts. Go back to the roots. Study Computer Science systematically.

It will be the only “weapon” that allows you to survive in this cruel world.

If you found this post helpful, you can support my work here:

👉 Buy Me a Coffee

🌏 Overseas Reader’s Supplement

KOSEN (National Institute of Technology) : Japan has a unique educational system where students enter 5-year engineering programs at age 15. It is known for producing highly practical and specialized engineers, distinct from standard 4-year universities.
The “Inexperienced” Hire Phenomenon : Unlike in the U.S. or Europe, many large Japanese corporations hire “new graduates” regardless of their major (e.g., hiring a Literature major as a Software Engineer) and attempt to train them from scratch. This trend has led to a lack of formal CS backgrounds in the workforce.
The IBM Gold Watch : In the golden era of IBM Japan, the top recruit in the entry-level training program was gifted a high-end watch (often a gold Seiko or similar). Veterans of the Japanese IT industry continue to view it as a symbol of elite status.
SIer (System Integrator) : A term commonly used in Japan to describe large technology firms that manage massive, end-to-end IT projects for corporate clients. These firms often prioritize project management over deep technical expertise.

Voice Input Gemini Gem: Achieving Uninterrupted Thought Flow with Shossan-style Writing Workflow

Takashi Abe — Tue, 03 Feb 2026 02:55:28 +0000

Recently, my output method has been completely established as a style that starts with voice input.

In addition to Aqua Voice, which I have been using for some time, I have tried various tools such as Whisper Transcription, which runs locally. Through this trial and error, I have found the optimal solution for myself, so I would like to introduce that setup this time.

The combination of voice input and Gemini Gems achieves both consistency and speed in writing

My current writing style has been consolidated into the process of dumping out rough thoughts with voice, structuring them with Gemini Gem, and finally putting my soul into it with my own hands. The part about putting my soul into it sounds very AI-generated. I like it.

By building this workflow, I have been able to dramatically lower the psychological hurdle of writing while maintaining a certain level of quality for my notes. Above all, by using different tools for different purposes, the accuracy of my drafts has improved incomparably compared to before.

The difference between transcription and text generation affects the quality of output

Why is just transcription not enough? It is because there is a big difference in the ability to refine Japanese depending on the tool.

Aqua Voice: Very excellent. The instructions function accurately, and there are almost no Japanese conversion errors or kanji mistakes.

Whisper Transcription: There is a sense of security that it runs locally, but since it specializes in pure transcription, I still feel it is weak in correcting misspeaking and kanji misconversion.

Fillers are not included in either, but I hardly say things like uh or um anymore, so I have no idea whether it is being controlled or whether I have been trained.

The raw text that has just been transcribed, no matter how accurate, does not become my writing as it is. This is where the process of structuring with generative AI becomes important.

The Shossan-style workflow that fully utilizes Gemini Gems and TextExpander

Let me introduce specifically how I prepare my output.

1. Automating structuring with Gemini Gem

Previously, I used to input prompts via snippets, but now I utilize Gemini Gems.

I created a Gem dedicated to notes, and I just throw in the text I voice-inputted as is. With just this, a draft that has been organized into a readable structure while taking into account my writing style is completed. It is overwhelmingly faster to open a dedicated Gem than to type instructions each time.

## Role
You are a renowned blogger and technical writer with 1 million PV per month.
You take the rough text based on voice input provided by the user (Shossan), grasp its true intent and supplement it, and elevate it into an attractive technical blog that captivates readers.

## Thought Process
Before generating a response, execute the following steps internally:
1. Supplement typos and misconversions: Infer and correct conversion errors unique to voice input from context.
2. Structural analysis: Check whether there are logical leaps or deficiencies in the content and supplement as necessary.
3. Restructuring with PREP method: Organize the structure in the order of Conclusion (Point) → Reason → Example → Conclusion (Point).

## Goals
1. Create a blog draft:
- Attach an attractive title.
- At the end of the body, extract 10 or more tags separated by spaces that do not contain hyphens (-).

## Context and Constraints (Style and Tone)
- First person: Basically watashi. Shossan when emphasizing.
- Tone: Friendly and humorous, conversational desu-masu style.
- Technical density: Do not cut specific proper nouns, processor names, version numbers, etc., but make use of them.
- Readability:
- Thoroughly implement paragraph writing and avoid paragraphs becoming too long.
- Create headings according to PREP

## Output Format
1. [Title]
2. [Body (Markdown format)]
3. [Tags (space-separated)]

The Gem that Shossan is using right now

I created this Gem together with Gemini. It is also included as a project in Claude.

2. TextExpander for that extra touch of convenience

Not everything is completed with Gemini Gem alone. For fine adjustments, I use TextExpander, which I have been using for many years.

Welcome to TextExpander.com

Generation instructions for header images that match the content of the article
Listing appropriate tags
Inserting specific standard formats

I call these as snippets and assemble parts while interacting with Gemini.

3. Finally, revise it in your own way as an editor

The level of AI has improved, and seemingly perfect sentences now come out. However, they are still sometimes slightly off from my intention or have unnatural phrasing.

I always read it back with my own eyes at the end and revise it to my own expression. Only through this process does the text created by AI become elevated into Shossan notes. Sometimes I do not know whether I am a writer or an editor, but I believe this final touch is my sincerity to readers.

Someday, a world will come where text typed on a keyboard feels warmer. Now, which paragraph in this note did I actually type on a keyboard? Do you feel the warmth?

Use tools appropriately and ensure consistent quality

The general flow has not changed since I started outputting with voice input. However, by preparing a dedicated mold like Gemini Gem, I have been able to stabilize the quality of my output at a high level.

If you have tried voice input but gave up because corrections were troublesome, please try creating a dedicated AI Gem. Your thoughts should crystallize into text surprisingly smoothly.

Audio-Technica AT2020 CWH Condenser Microphone

MOTU UltraLite mk5 18in 22out USB-C Audio Interface

Quality voice input comes from high-quality devices.

If you found this post helpful, you can support my work here:

👉 Buy Me a Coffee

Ultimate Formation to Never Stop Thinking! Starting a Fully Automated Voice Input Life with MacWhisper Pro and AquaVoice

Takashi Abe — Mon, 02 Feb 2026 09:04:24 +0000

Recently, voice input has been at the center of my life.

Don't you sometimes feel that the act of typing on a keyboard can block the flow of your thoughts?

In light of this, I introduced a new app.

It's MacWhisper Pro (available as Whisper Transcription on the App Store). This time, I want to talk about how I use it alongside my beloved AquaVoice and why voice input using a local LLM is the best option right now.

🎙️ MacWhisper

MacWhisper - Quickly and easily transcribe audio files into text

The Two-Pronged Approach of Real-Time AquaVoice and Batch Processing MacWhisper Is the Ultimate Solution

To get straight to the point, the right answer for me now is to use different voice input environments for different purposes.

When I want to turn my thoughts into text in this very moment, I use AquaVoice. When I want to capture thoughts during reading or long thinking sessions, I use MacWhisper. By combining these two, I can now keep my output running at full capacity at all times.

Why Did I Add MacWhisper?

AquaVoice is currently the best voice input app available for Mac in terms of both accuracy and real-time performance. However, it had some weaknesses.

Recording time limitations: Not suitable for long continuous recordings.
Manual intervention required: For long sessions, you need to periodically stop and restart, and each time keyboard operation is required, which interrupts your thinking.

What I needed was to transcribe recordings that stream continuously for 30 minutes to an hour while reading—and MacWhisper answered that need.

Local Processing and External Integration Eliminate Friction in Intellectual Production

Why is MacWhisper so easy to use? The reason lies in its design that does not interfere with the user's thinking.

One-time purchase option available: Instead of going through cloud APIs, it uses the local LLM (Whisper model) on your Mac itself, making a one-time purchase possible with reduced running costs.
Seamless app integration: You can summarize transcribed data with LLM or send it directly to external apps like Obsidian. The fact that it is designed to handle post-writing processing is a nice touch.
Privacy and speed: Since you don't need to send data to the cloud each time, you are freed from the waiting time and security concerns typical of web services.

My Personal Voice Input Workflow

Here is how I actually use these tools:

Situation	Tool	Reason
Thinking and writing in front of my Mac	AquaVoice	Real-time input provides the best experience
Thinking while reading or during long thinking sessions	MacWhisper	Can handle long recordings, batch transcription
Taking notes outdoors	notebooks	Noisy environments make desktop-level accuracy difficult

Failure Story:
What about voice memos while running? I thought it would be great to take notes while running, so I tried it, but it was a failure. Wind noise and car noise significantly reduced accuracy. When outside, the better options for now are bringing the iPhone close to your mouth or using Siri on Apple Watch.

TIPS tested by entering in quote format

Voice Input Is a Thought Accelerator

After trying various things, voice input is just right for us older folks. Above all, it's fast, and the biggest advantage is that your thinking doesn't easily stop.

A tool that lets you convey what you think, exactly as you think it, immediately.

Of course, when organizing and systematizing things, I still use analog notebooks and pens. What matters is distinguishing between outputting (voice) and organizing (handwriting), don't you think?

MacWhisper has a universal environment available on iPhone and iPad as well. A system that catches your thoughts without missing any, whether at home or outside. Why not try building one yourself?

M4 Max Mac Studio Terminal_ Mastering Lightning-Fast CUI Life with Ghostty and Rust-Made Tools

Takashi Abe — Mon, 02 Feb 2026 09:00:31 +0000

Why do I center my work around the Terminal, despite using an expensive Mac Studio?

![Comfortable terminal environment unique to ultrawide displays]

The answer is: because it is comfortable.

GUIs are still heavy on processing, no matter what. I hardly ever feel slowness with the M4 Max, but there are often unnecessary processes that make me feel a slight delay. I can sense the screen redrawing, so to speak.

In contrast, terminals like Ghostty are comfortable. They are always fast. The rendering performance is extraordinary.

Ghostty

Ghostty is a fast, feature-rich, and cross-platform terminal

And almost all applications running on it have the same screen redrawing performance. No frustration builds up. There are no unnecessary interrupt processes, and nothing hides your cursor. Nothing suddenly pops up either. For handling single tasks, not being interrupted by anything is the optimal structure.

Always Keep Terminal in Your Heart

I believe that as an engineer, you should be able to speak through commands.

It is not that you cannot do anything without a mouse—I hope you can be an engineer who can always manage with just commands.

Because it is so much fun.

The ease of processing with just commands. The joy of combining commands like a puzzle. The comfort of completing everything with just a keyboard.

Commands I Have Been Using Lately

Yazi

Blazing fast terminal file manager written in Rust, based on

This is an incredibly comfortable file manager. It can do more than you would expect.

eza is a flashy replacement for ls. It adds color to a monochrome screen.

bat provides line numbers and syntax highlighting for code, making things look flashy as well. It can even show git diffs. Amazing.

lazygit makes git commands flashy—or rather, it displays git status in a list view and lets you easily check changes. It is better than expected.

btop is a stylish top that displays things in a somewhat cyber way. It is fun just to watch. And it is a bit sad to see that all those resources are barely being used.

Since many commands can replace existing functionality, when using fish, I set up aliases as follows. Feel free to use this as a reference. By the way, I am displaying this with bat. This is what it looks like, folks.

$ cat ~/.config/fish/conf.d/alias.fish
─────┬─────────────────────────────────────────────────────────────────────────────────────────────────────────────────
     │ File: /Users/sho/.config/fish/conf.d/alias.fish
─────┼─────────────────────────────────────────────────────────────────────────────────────────────────────────────────
   1 │ if status is-interactive
   2 │     # --- Aliases (speed-boosting commands) ---
   3 │     # Replace ls with eza (with icons and Git status)
   4 │     alias ls='eza --icons --git'
   5 │     alias ll='eza -l --icons --git --group-directories-first'
   6 │     alias la='eza -la --icons --git --group-directories-first'
   7 │     alias tree='eza --tree --icons'
   8 │
   9 │     # Replace cat with bat
  10 │     alias cat='bat'
  11 │
  12 │     alias lg='lazygit'
  13 │
  14 │     # --- Yazi settings (shell wrapper) ---
  15 │     # Type "yy" to launch yazi, and cd to that directory on exit
  16 │     function yy
  17 │         set tmp (mktemp -t "yazi-cwd.XXXXXX")
  18 │         yazi $argv --cwd-file="$tmp"
  19 │         if set cwd (cat -- "$tmp"); and [ -n "$cwd" ]; and [ "$cwd" != "$PWD" ]
  20 │             cd -- "$cwd"
  21 │         end
  22 │         rm -f -- "$tmp"
  23 │     end
  24 │ end

30 Minutes, 5000 Words. The Blog Lightning-Fast Creation Method Using Aqua Voice x Gemini to Boost Writing Speed 10x

Takashi Abe — Thu, 29 Jan 2026 08:05:38 +0000

Recently, my blog writing style has completely shifted to voice input.

Before, I used to sit at my keyboard and write, but now I have transitioned to a flow of speaking with Aqua Voice and having Gemini edit it.

Since switching to this method, my writing efficiency has dramatically improved.

To be specific, it used to take me over 30 minutes to write an article of 1000-2000 characters, but now I can finish around 1000 characters in about 10 minutes, and with 30 minutes, I can complete a long-form article of 5000 characters. By simple calculation, it feels like my productivity has increased 5 to 10 times.

This time, I will introduce the specific details of this lightning-fast output method that I practice.

1. The Input Revolution: Aqua Voice Is Not Just Voice Input

First, what I use in the input phase is Aqua Voice.

Aqua Voice - Fast and Accurate Voice Dictation for Mac and Windows

You might think, if it is just voice input, why not use the standard features from Google or Apple? However, Aqua Voice has a decisive difference.

That is the ability to apply real-time preprocessing and reconstruction by an LLM (Large Language Model) to the voice stream.

It does not simply convert voice to text. By giving instructions in advance such as fix unnecessary sentences or delete redundant repetitions, it cleans up the noise and contextual disorder unique to spoken language at the input stage. Even if you blurt out something extra or stumble over your words, by the time it is converted to text, it has already been reconstructed into readable sentences.

Currently, I feel that Aqua Voice is the most excellent tool that can do stream voice input and utilize LLM at such an advanced level.

2. The Proofreading Revolution: Making Gemini a Ruthless Editor

Once Aqua Voice produces reasonably organized text, I then throw it to Google Gemini. What is important here is not simply asking it to clean up the writing, but having it point out flaws as an editor.

The actual prompt (instruction) I use is as follows:

I prepared a blog article draft using voice input.
- Since it is voice input, there may be parts where the conversion is incorrect. Please interpret while filling in the gaps.
- Please point out any structural flaws or deficiencies.
- After interpreting what I want to convey, please restructure it.
- Please prepare the text as a blog draft.

The key to this prompt is that it asks for pointing out flaws and deficiencies before requesting structuring. Especially when you include this, you get incredibly sharp critiques.

ChatGPTの「良い人フィルター」を外して本音を引き出してみた - Qiita

For example, while writing this article, I received the following sharp feedback from Gemini:

The superiority of Aqua Voice (differences from other tools) is not explained.
The prompt remains a black box. Readers cannot reproduce it.
The quantitative basis for became 5 times faster is vague.

These are very accurate critiques. By receiving this feedback and adding supplementary information verbally, it prepares a high-density article with no logical gaps.

3. Finishing: Fine-tuning in the Editor

Finally, I paste the draft that Gemini outputs into note and make fine adjustments to match my own writing style. This is important. If you do not adjust it yourself to match Shossan style, you end up with text that obviously looks like AI wrote it. Even though you packed in all your essence, it feels like your heart is not in it.

That said, since the structure and logical development have already been completed through back-and-forth with Gemini, here I usually just need to adjust particles and nuances of sentence endings.

Summary

Aqua Voice: An input device that turns thoughts into organized text in real-time
Gemini: An editor that fills logical gaps and organizes structure

By combining these two, I was able to eliminate the bottlenecks of writing - typing and struggling with structure - all at once.

If you are spending too much time writing, please try this voice input flow with an LLM as your partner. By continuing to speak into the microphone, you can train not only your writing ability but also your speaking ability at the same time.

I seriously feel like I have gotten better at speaking recently.