There is only one reason why you are getting average answers from the strongest AI.
Because that AI knows nothing about you.
It doesn't know your business, your customers, or what you decided yesterday. So it guesses from scratch every time. And guesses, no matter how good, only ever reach an "average score."
I solved this problem by creating a "copy of myself" using Claude Code. In technical terms, it's a Second Brain. It's a system where your memories, judgments, and writing style are stored on a disk in a format that AI can read and write.
I've been running this for several months and completely rebuilt it for Fable 5. Version 2.0. I'll write everything in this article—what I made, how I made it, the code, and the folder structure. I've put all the templates and scripts on GitHub, so please take them and grow them for yourself.
→ github.com/chaenmasahiro0425/exbrain
What is a Second Brain anyway? ── Just a folder on top of Obsidian
It's not a grand device. Its true identity is a bundle of Markdown files on a Mac. You view this with Obsidian (a free Markdown app), and Claude Code directly reads and writes to the folder. No plugins or connectors needed. Humans read with Obsidian, and AI touches the folder. Both are looking at the same "brain."
It works even when the PC is closed (automated cloud tasks). You can read it on an iPhone (iCloud sync). And no matter what model you switch to, this brain survives. This is the most important part. If you put the mechanism into one sheet, it flows like this: Store → Weave → Use.
Even with the same Fable 5, the moment it connects to this brain, it becomes a different machine. The code follows your architecture, the text is written in your voice, and the proposals stand on the facts you hold.
One lesson learned from running it for several months
Last year, I built the first generation centered on three files: SOUL / MEMORY / DREAMS. It was a system where daily logs, X clips, and meeting minutes were automatically accumulated. After running it for a few months, I learned something very clearly.
Raw logs accumulate on their own if left alone. But "compiled knowledge" rots if left alone.
Daily notes and clips grow fatter every day through automation. However, customer pages and indexes became obsolete in a few weeks.
A customer list created in April remained frozen, not knowing a single new project from July. A brain that only grows when you remember to feed it dies in three weeks.
So I divided it into "4 layers" ── The 2.0 Design
In 2.0, I stopped putting everything in one box. I divided it into four layers with different characteristics and fixed only one "writer" per layer. If there is only one writer, synchronization conflicts structurally do not occur. Identity (personality) is written only by humans, digest (summary) only by the cloud, wiki (knowledge) by a nightly compiler, and raw (raw logs) by automatic collection.
Why make the raw layer inviolable? Because if the same AI repeatedly reads and rewrites the same note, details melt away and errors increase with compound interest. By freezing the raw logs, the knowledge layer above can be rebuilt as many times as needed.
This update was made with reference to the following overseas article.
https://x.com/masahirochaen/status/2073548158270144705
The heart of 2.0 ── The "Nightly Compiler"
This is the main event. I left the job of turning raw logs into knowledge to an automatic compiler that runs every night at 23:30. Every night, this thing wakes up, reads the day's logs and new clips, and automatically updates pages for customers, people, and tools mentioned, complete with source links. If it says "Agreed with CyberAgent for 4.5 million," that fact piles up on the corresponding page.
There are three mechanisms to prevent it from running wild. ① Do not let the LLM touch git. ② Limit where it can write by layer (don't let it touch raw or identity). ③ If a change occurs in raw, automatically restore it and stop the commit. And cost is key. Compilation is routine work, so run it on the cheapest model. Premium models only need to come out for the weekly integration pass.
The "Loop" that keeps the brain alive
A Second Brain lives by schedule, not memory. At the start of each session, a primer injects today's context, daily notes are generated morning and evening, X bookmarks are gathered every 4 hours, compiled every night, and decay detection (lint) and weekly summaries run on Sundays. Premium models only work once a week. Everything else is handled by cheap models and shell scripts. Throwing routines at top-tier models is like throwing money down the drain.
You don't have to choose the "memories you saw" daily
"Do I need to choose which articles are good myself?" ── No. That is the answer of this design.
If something catches your eye on X, just bookmark it. The clip function every 4 hours automatically flows it into the raw layer (1,467 items have accumulated now).
Even just dropping an article link into a Slack DM is fine. All humans do is "see." Selection was made the job of the nightly compiler. We separated reading from remembering.
And then Fable 5 arrived ── I threw away 80% of the "instructions" I had grown
The rules I had written over several months had become enormous before I knew it. The first thing I did when switching to Fable 5 was to "delete" 80% of them.
Old models would return to average scores unless you gave them step-by-step instructions. So I kept adding "don't do this, don't do that" endlessly. Fable 5 is different. It grasps intelligence and speed through effort (man-hours), and the desired behavior only needs a "short sentence." When I stopped listing things, it actually became smarter.
For the Second Brain, this was a tailwind. Skills and prompts that had bloated for older generations actually lower the quality in Fable 5.
So in 2.0, I shortened the compiler prompts and CLAUDE.md. You only need to store "facts" in the brain. Behavior becomes shorter as the model gets smarter.
Summary ── 7 steps to create a "copy of yourself"
① Create a vault (Obsidian folder)
② Divide into 4 layers (raw / wiki / digest / identity), fixing one writer per layer
③ Throw everything you have into raw
④ Set up nightly compilation (every night, with a cheap model)
⑤ Detect "decay" with a weekly lint (Sunday, no LLM)
⑥ Turn research into assets with verification
⑦ Use INDEX as the entrance and prohibit full scans
The model in the driver's seat will change again. But the vault will survive every transition. And the written feedback will make the brain smarter every week, no matter who is driving. You can make the smallest version in an hour. One folder, ten files written about your business, and an agent told to "read that first." After that, the output will tell you the rest.
I've made it all open source. The templates, the scripts, and the README in both English and Japanese. Please feel free to take them. → github.com/chaenmasahiro0425/exbrain
