I run my life as a single Git repository, operated by several AI agents through Claude Code. Journaling, self-analysis, investing, hobbies, parenting: each of them started as a small app I built to solve a problem of my own. Parental leave gave me a long stretch of free time, so instead of running them separately, I moved them into one place. And the primary reader of this repository is the AI, not me.
The moment I stopped writing for humans, my journal, my investing notes, and my parenting records all turned into one knowledge base that an AI can read across. Apps that used to sit in separate boxes finally connect here.
This post is a map of the whole thing. I'll split the individual how-tos into their own posts, so here I only cover two things: why bundle everything into one, and how it runs.
What I build almost always starts from a problem of mine or my family's.
At first I built each of these as its own standalone app. But after making a few, I noticed that separate apps kept splitting up the one thing that should stay continuous: me. My investing tool didn't know my values. My journaling app didn't know the quirks in how I think. The same person was using all of them, and none of the context carried over.
Here's where a habit in how I think shows up. When I face a problem, I break it into conditions first and look for the most promising route.
Do I line up a lot of handy tools, or put everything on one foundation? The second option keeps all the context in one place, and the base gets stronger the more I build. Which one compounds? That question is why I decided to put my life inside a single repository.
This is the core premise.
On a team, you write documentation so humans can share it. But solo, there is no such reader. The primary reader of my repository is the AI. So I don't design my documents to be easy for a human to read. I write them in a structure that's easy for an AI to read and search.
Concretely, I do three things:
I cut prose, decoration, and anything meant to be read top to bottom. When I want the content back, I don't open the file; I ask the AI.
This isn't a second brain in the usual sense. A second brain is meant to be read back later. Mine is the opposite: the notes are deliberately unreadable to me. They're written for the AI, and the human-facing interface is conversation, not the files.
The skeleton of the repository looks like this:
life/
├── journal/ # dialogue-style journal (year/month/day.md)
├── persona/ # self-analysis (dimensions/ = cross-cutting, chapters/ = timeline, insights/ = auto-extracted)
├── projects/ # metadata for each dev project
├── ideas/ # seeds of ideas
└── .claude/
├── skills/ # per-purpose agents (/journal, /weekly-review …)
├── rules/ # behavioral rules injected into the agents
└── config/context.yaml # context shared by every skill
Every file starts with metadata the AI can query against. A journal entry, for example, begins like this:
---
date: 2026-03-12
mood: 4
tags: [dev, ideas]
highlights:
- sketched out a new project
---
Extend this to a whole life, and my journal, self-analysis, values, and project records all pile up in one place as structured knowledge the AI can read. That's the foundation I stand on when I operate my life.
Let me draw the line up front. I can share how I built it; I keep my personal data itself private. This post stays inside that line.
| Shared | Kept private |
|---|---|
| Architecture and design ideas | The contents of my journal |
| How I think | Anything about my family |
| Generalized skills / CLAUDE.md | My actual holdings |
| Directory structure | The body of my self-analysis |
One note: the skills and CLAUDE.md I publish are carved out of my repository and generalized. My everyday records are written for the AI, but anything I hand to people I rewrite into a form a human can read and reuse. I treat the two as separate things.
The template below is on GitHub: the CLAUDE.md, the directory structure, and three skills for journaling, weekly review, and self-analysis (/journal
, /weekly-review
, /persona
).
日本語 | English
Claude Code で「人生を1つのリポジトリ」として運用するための、AIネイティブな個人運用フレームワーク（テンプレート）。
日記・自己分析・プロジェクト管理・アイデアを、人間ではなく AI が第一読者となる構造で1つのリポジトリにまとめ、用途ごとの Claude Code スキル（エージェント）で運用します。
設計思想の全体像はこちら → リポジトリは人間のために書かない — Claude Code で日記・投資・育児まで運用する【全体像】
.
├── CLAUDE.md # Claude Code への共有指示テンプレート
├── journal/ # 対話式日記（年/月/日.md）
├── ideas/ # アイデアストック
├── projects/ # 各プロジェクトのメタ情報
├── persona/ # 対話型ペルソナ（自己分析）
└── .claude/
├── config/context.example.yaml # 全スキルが共有するコンテキストの雛形
├── rules/ # エージェントに注入する振る舞いのルール
└── skills/ # 用途別エージェント（/journal, /weekly-review, /persona）
CLAUDE.md
— Claude Code への共有指示テンプレート.claude/skills/
— 対話式日記 / 週次レビュー / ペルソナ構築の3スキル（汎用版）.claude/rules/
— 方法論ルール（mcp-vs-skills）＋対話の振る舞いルール（unknown-term-lookup）＋各種フォーマットの型.claude/rules/correction-rules/
— AI に自分を正しく読ませる補正ルール。汎用化した実例を1本同梱（example-understated-expression.md
）.claude/config/context.example.yaml
— 全スキルが共有するコンテキストの雛形journal/
ideas/
projects/
persona/
— 各ディレクトリの役割を README で説明.claude/config/context.example.yaml
を同じ場所に .claude/config/context.yaml
としてコピーし、自分の関心・プロジェクトを記入する。/journal
/weekly-review
/persona
を起動する。CLAUDE.md
・rules
・skills
を調整する。.claude/rules/correction-rules/
参照）。残りは個人の例を含むため、汎用化してから順次おろします。設計の考え方は上記の記事と続編で書きます。/trend
（トレンド収集）・/improve
（環境改善）などの周辺スキル — 本テンプレには未収録です。context.yaml
の sources
/…Correction rules that contain personal examples, and domain-specific skills, get added there as I generalize them.
On top of this foundation, a per-purpose agent (a skill) runs for each job: one that writes the journal, one that deepens self-analysis, one that reviews everything weekly, one that collects trends. Each reads the same knowledge base.
flowchart TB
subgraph AG["Agents (skills)"]
direction LR
SJ["/journal"]
SP["/persona"]
SW["/weekly-review"]
ST["/trend"]
end
subgraph KB["Shared knowledge base (one repository)"]
direction LR
J["journal/"]
P["persona/"]
PR["projects/"]
ID["ideas/"]
end
AG -->|read / write| KB
SP -.correction rules.-> SJ
SJ -.output becomes input.-> SW
KB --> M1["self-understanding"]
KB --> M2["money (investing)"]
KB --> M3["hobbies (recommendation)"]
KB --> M4["parenting"]
Here's what the "multi-agent" part actually means: one agent's output becomes another agent's input. The weekly-review agent reads across my journal, my project progress, and my health data at once. And running through all of it, my way of breaking a problem into conditions is baked into the instructions the agents follow. The quirks of how I, a specific person, think get reproduced as the behavior of the system. A generic AI tool can't reproduce that.
Of the four areas, this is the one I put the most into.
Most "AI assistants" work on the world outside a person: schedules, tasks. This one works on the thing before that: getting the AI to stop misreading me. Not in a deep, mystical sense — concretely.
Left alone, an AI flattens a person into a convenient summary. Say I mention that a day was "the usual." The agent reads that as "a light day" and moves on. But one of my correction rules says: when I call something "the usual," don't treat it as light — go check what I actually did that day. Before the rule, the journaling agent would treat such a day as low-effort. After it, the same phrase makes the agent open that day's record first, so it stops filing the day as light. Its follow-up questions tend to get sharper, too.
I build these rules from a dialogue journal. My values, thinking quirks, and emotional patterns get structured slowly, each with evidence attached: a quoted journal line and a date. That structured picture is my persona, and the correction rules are the part that keeps the AI from reading it wrong. None of this stays inside self-understanding, either — the same persona and rules are what let the other agents act with my context in mind.
I currently have apps for four areas. Each deserves its own post, so here's the map.
The four look scattered but come together as one, because they share the same AI-read knowledge base and the same way of thinking. The tendencies I've built up from self-analysis are visible to the other agents too, since they all read the same base. Four apps sitting side by side can't do that.
My approach is "grab only as much as I need to make it work, then bundle it across areas." I don't fully understand something before I build; I move forward with what it takes to run it. So I don't compete on the theoretical depth of any single area, the deepest parts of deep learning or financial engineering. I draw an honest line at "this is as far as I got." The value I can offer is in the cross-cutting design, not the depth.
And not everything runs smoothly. My development comes in waves. Building something new is fun and it moves; the cleanup phase afterward, tidying up what's already built, is boring, and my hands stop. Of the four areas, some are running now and some are sitting where I left them. The reason they're stopped is "I got bored," and there's no story beyond that.
So I'll write these up one module at a time, starting with the ones I still use every day.