Leave the tedious work to machines.
Keep your time for what you love.

I keep turning pieces of my own life into systems that run without my hands on them. Replying to email, posting to social media, coordinating plans with people, keeping a record of what I'm doing right now — I've handed off one slice of daily life after another to machines. What's listed below is the part of that I can show publicly.

This site is where I keep what I think and what I build — written for people, and for AIs to read too.

Today's observation as of 2026.08.21 22:06
now: away from the desk (164 min) ・ last input: just now
Pro
0h 42m Air
0h 00m Phone
0h 00m
0:006:0012:0018:0024:00

The PC rows (Pro/Air) record minutes I was actually at the keyboard; the Phone row records minutes an app was open. The taller the bar, the more of those 30 minutes were in use. Only "when and how much" is published — never the contents. Updates automatically a few times a day.

What I make
Tools I built for myself
Writing
What I think, day to day
Contact
An AI replies first
Discord Looking for peers
A lab for locally-running AI agents

Shunki Ogino's AI Lab

Latest updates

What I've been thinking (the blog), plus the latest from YouTube and Instagram. The daily record gathers here as one stream.

Writing
The Space Where Mistakes Don't Matter is Narrow
Even when the AI lied to me, it didn't ruin my appreciation of the novel. I figured there must be other areas where mistakes are acceptable, so I reviewed my daily records. I only found one.
2026.08.11
Writing
The Trade-off Between ID Control and Agent Autonomy
Discord bots cannot act on behalf of a user's permissions. In contrast, when implementing Telegram Bot integration, operations closer to those of the actual user were possible.
2026.08.03
▶ YouTube
8月20日(木)の観測|PC 8分・歩数2,221歩 #Shorts
2026.08.21
Instagram
8月20日(木) の観測 Pro 8分・スマホ 2時間23分 最長の連続は 17:30〜18:30(Pro)。 平均…
2026.08.20

What I make

Evidence of what I think about. Each is a tool I built for myself, made to keep working with my hands off it.

Game / Physics Released

Dissipative Structure Game

A browser game that builds physics — the second law of thermodynamics — directly into its rules. An experiment in replacing a hard concept with something you grasp by touching, not by explanation. Source released under MIT.

GamePhysicsPlays in the browser
AI agent Running

An AI that replies to your email

For each sender, a dedicated AI spins up, and a locally-running AI (Gemma) organizes the content and drafts a reply. The email's contents are processed on the machine, never handed to an external AI, and each sender's visible scope is kept separate.

AIEmail automationLocal processing
Video tooling Running

A studio that makes videos

From what you said out loud, an AI drafts the script and automates layout, voice synthesis, and export. Both script and voice are made on the machine, without relying on cloud generative AI.

VideoAI scriptVoice synthesisLocal generation
Manga / Tooling In progress

A tool for making manga

I'm building the tool for making manga itself. Lay out panels on a page, place poses and backgrounds from a shelf, and a locally-running AI looks at the page and thinks through the presentation with me. A homemade editor that keeps the AI in the loop instead of handing the work to off-the-shelf software. A one-shot manga is in production with it.

MangaHomemade toolLocal AI

A product you can actually buy is in preparation. The first one is planned to be a tool that helps people stay connected.

How I keep it safe

Handing pieces of my life to AI comes down to caring about exactly two things: never letting information leak, and keeping the damage small on the day something still goes wrong. Here's an honest look at some of that.

Never let it leak

Processed on my own machine

Most decisions are made by a small AI running on my own computer. The default design keeps content off external clouds.

Anything that could leak gets stopped mechanically

Every draft is automatically scanned for things like customer names, secrets, or personal data before it goes out. If anything trips that check, even a reply that would normally send itself gets held for a human to confirm. The block doesn't depend on anyone actually watching — it's built into the pipe.

Keep the damage small when something still goes wrong

Keeping it out of other people's sight

Depending on who the information came from, it stays invisible to other people and third parties. What happens in a conversation with A never becomes visible to B.

Keeping it out of AI's sight too

It's not just about hiding things from people — the AI itself is only ever given the data it actually needs. An unrelated conversation never reaches an AI that isn't handling it.

Never let one AI both 'know' and 'act'

An AI with broad internal knowledge is never given strong execution power — sending messages, running code, touching files. And an AI that can act outward (sending, posting) is only ever handed the narrow slice of information it needs for that one job. Keep knowledge and execution apart, and even a misbehaving AI is stuck being only the one that knows, or only the one that can act — never both.

And under all of it, one backstop: if there isn't enough information, or a condition isn't met, nothing goes out. The system is built to show or send nothing rather than fill the gap.

Contact

Send a message here and an AI running on my own computer will organize it and reply first, automatically. This form is actually the "AI that replies to your email" shown above, running for real. Whether you want to build together, cheer the project on, or just look around — you are welcome.

From sending to getting a reply
1
Submit the form
Saved to the server
2
The AI receives it
Organizes the content
3
Automatic reply
You get an email
4
I read it myself
Ogino reads and acts