Kaishu Shito

Work

2023 — now

AGI Lab AGIラボ

Co-founder · COO of tempi Inc.

Co-founded a Japanese AI publication and paid community. The publication's X account has 138,500 followers.

What I built: I co-founded this in January 2023 with my partner, who is the representative director of tempi, the operating company. I am COO and own the go-to-market half: membership and pricing, the Discord, the editorial calendar, courses, events, member migration, analytics. In July 2026 I pulled our billing, platform and bank data into one reconciled picture and produced the first revenue baseline the business had ever actually had. I led the community side of the move off note.com onto our own platform in May 2026 — the coupon campaign, the member migration, the Discord identity linkage.
Hardest part: Two billing systems ran in parallel during the migration, so a member moving from one to the other looked exactly like a member leaving. Before proposing anything, I separated channel migration from actual cancellation in the measurement. An earlier analysis of our member records showed most cancellations happened in the first three months, and the most common stated reason was information overload rather than content quality. That is what drove the Discord restructure and the onboarding work.
Learned: Most of what I believed about this business was not measured. Our buyers turned out to be business operators more than engineers. The churn reason was volume, not quality. I instrument before I argue now, and I publish the number I can defend rather than the number I want.
Where it is: Organization accounts: 138,500 followers on X, ~22,000 on note, 18,200 on Instagram. 500+ deep-dive articles over 3.5 years, written across the team. agi-labo.com

2026

Getting Started with Google Antigravity 『はじめてのGoogle Antigravity入門』

Sole writer · published under AGI Lab · Gijutsu-Hyoron

Wrote all 14 chapters of a 224-page book on Google's agent IDE. Out 28 September 2026.

What I built: I took over the publisher thread on 5 February 2026 and carried the project from proposal to print. 14 chapters, about 235,500 characters of Markdown under Git, 343 screenshots I shot and placed myself, and the whole galley cycle: the query lists, replacement figures, the typesetting fix list, the preface, cover feedback, the re-proof. I drafted with coding agents under my own rule that AI-generated chapter text is source material and not a manuscript — I reproduce the behaviour on a real machine, capture the screen, then write. The printed byline is the organization, AGI Lab, not my name.
Hardest part: Google shipped Antigravity 2.0 on 19 May 2026, after my first draft was already with the publisher. It split the IDE, added a standalone app and a CLI, and invalidated a lot of the book. I revised against the new build rather than patching, and moved every volatile fact — pricing, quotas, model names, UI labels — into a dated appendix and a reader support page, so late fact-checking only had to re-verify one section. I also wrote a script that diffed the re-proof PDFs across every chapter for duplicated captions, stray links and leftover dummy text, because one proof arrived while I was travelling.
Learned: A book about a product that ships breaking changes every few weeks needs an architecture, not just good writing. Quarantining the volatile facts was the single decision that made the last three months survivable.
Where it is: 224 pages, ¥2,640. ISBN 978-4-297-15883-5. gihyo.jp

2026 — now

AGI Radar

Spec, design and build

Built a live site that forecasts when AI labs will ship their next models. It tracks 212 models against confirmed primary sources and updates itself every morning.

What I built: I wrote the spec, designed it and shipped it. Next.js on Cloudflare Workers with a D1 database, OTP login, a voting layer, Japanese and English editions. A pipeline runs unattended at 06:00 JST: collect release signals through Gemini search grounding and Grok's X search over a leaker list I curated, score the sources, update each forecast. I built it with Claude Code and Codex as the coding agents, and I reviewed every claim the system is allowed to make.
Hardest part: The product's whole value is whether you can trust a prediction, so the hard part was credibility, not the model. Every signal carries a label (unconfirmed / multi-source / official), mid-tier leakers start at unconfirmed, and a leaked codename stays marked as a leak until an official source confirms it. On the operations side, a routine merge cannot deploy to production or touch the production database; both need a separate, explicit step. That rule came out of an incident where a deploy silently rolled back the live UI.
Learned: Autonomous does not mean unattended. The parts I had to design most carefully were the ones that decide what the system is allowed to assert.
Where it is: Live and updating daily: 212 confirmed models and 792 candidates under review as of 23 September 2026. The code is private, so the site is the verifiable part. agi-labo.com/radar

Sep 2025

Claude Hackathon 2025 in Tokyo

Proposed it · ran operations · mentor

Proposed Japan's first Claude hackathon and ran its operations alongside Anthropic and KERNEL. More than 100 people built with Claude for three days.

What I did: In July 2025 my co-founder introduced us to Anthropic's community team, and I wrote the proposal that started the event. I brought it to KERNEL, the AI incubation space run by DEEPCORE in Hongo, which agreed to host. Then I planned how our team would run it — first-round screening, the online stream, the Discord, recruiting mentors — and worked the floor as a mentor over the three days. Officially the event was co-hosted by Anthropic and KERNEL, with AGI Lab running operations.
What happened: The beginner workshop on day one was planned for 30 people; more than 80 joined. Eleven teams reached the final, Anthropic's community lead joined live for the closing keynote, and winners took home Claude API credits.
Where it is: 5–7 September 2025, KERNEL Hongo and online. Reports: AGI Lab · KERNEL

2026

AI agent hackathons

Organizer · built the AI judging pipeline

Ran two AI agent hackathons, for 50 and then 70+ people, and published the full log of the AI that judged the first round.

What I built: I started AGI Lab's first hackathon in March 2026: the format, the starter repository, sponsor prizes, the venue, Demo Day. For the September joint event with the AI Mokuyokai community I built the first-round judging pipeline — agents read every submission against a fixed rubric, in batches, reconciled row by row so a missing result shows up instead of disappearing.
Hardest part: Making AI judging something entrants would accept. The answer was not a better prompt; it was making the run auditable. I recorded what the AI could and could not confirm for each submission, mapped 36 results back to 36 rows, and kept three human judges for the final.
Learned: People will accept a machine's judgement if they can see its working. Publishing the pipeline's failures bought more trust than a better accuracy number would have.
Where it is: March: 50 applicants, 7 finalists at Demo Day in Shibuya, 3 winners. September: 70+ participants, 17 submissions, 7 finalists. The published judging log

2026

Reading Books with AI 『AIと本を読む』

Co-writer · published under AGI Lab · Forest Publishing

Wrote the first half of a 192-page book on reading with AI, and ran the experiments behind its method. Out 18 September 2026.

What I built: The book claims a method, so every claim had to be run before it was written. I built a 79-question verification sheet and measured it row by row, then ran three more rounds. One experiment took Book I, Chapter 1 of Clausewitz's On War through the method in one pass — 28 of 28 sections, every claim numbered back to the source. Cross-checking caught 3 errors in the AI's own output, which went into the book as the reason the method has a verification step at all.
Learned: A how-to book is only as good as the number of times the author actually did the thing.
Where it is: 192 pages, ¥1,980, first printing of 5,000. I wrote roughly the first half; the byline is the organization, AGI Lab. Press release

2024

How to Make ChatGPT GPTs 『ChatGPT GPTsが作れるようになる本』

Co-writer · published under ChatGPT Lab · SB Creative

Our first book: a 240-page hands-on guide to building custom GPTs. Out 29 March 2024, and it went into a second printing.

What I did: I wrote about half of the book's six chapters. We had spent the months before building GPTs and explaining them to members of our paid community, and the book grew out of that. It is published under ChatGPT Lab's name.
Where it is: 240 pages, ¥2,200. ISBN 978-4-8156-2553-5. sbcr.jp

2025 — 2026

A website and CMS for a local business

Solo, client work

Rebuilt a small business's website and shipped a CMS the owner now runs every day without me.

What I built: A full rebuild and a custom CMS, built alone, with CI and a pull-request review on every change. Every page was checked at phone width before it merged.
Hardest part: The owner is not technical and had to be able to publish without me. That constraint decided the CMS design more than anything else.
Learned: Shipping for one non-technical owner is a harder spec than shipping for a thousand developers. It's done when they stop calling you.
Where it is: In production. The client is not named here, and the code is private.

Jul 2026

Codex 2040

Solo · four days

Built a playable AI-governance strategy game in four days. Placed third at Global Build Week Tokyo.

What I built: You run an AI lab to 2040, and every choice trades capability, safety, public trust and capital against each other: a 50-node strategy catalog, a world-event system, a voice mode. Codex wrote most of the engine, UI, tests and localization; I designed the system, the economy and the failure states.
Hardest part: Making a governance game that is neither doom nor propaganda. Every lever needed a plausible cost, so the interesting decisions sit where safety and shipping speed genuinely trade off.
Where it is: Third at Global Build Week Tokyo on 18 July 2026, a community event run by an OpenAI Codex Ambassador, decided partly by audience vote. The judged build was the MVP; the rest landed over the next four days. Play · Source · Demo

Kaishu Shito

Work

AGI Lab AGIラボ

Getting Started with Google Antigravity 『はじめてのGoogle Antigravity入門』

AGI Radar

Claude Hackathon 2025 in Tokyo

AI agent hackathons

Reading Books with AI 『AIと本を読む』

How to Make ChatGPT GPTs 『ChatGPT GPTsが作れるようになる本』

A website and CMS for a local business

Codex 2040

Also

Timeline

Contact