A living wiki that gets smarter every run

The whole system rests on one idea: instead of a pile of chats and files, you keep a wiki. Every note you've ever written, every build, every conversation gets read across and folded into topic pages.

Raw history is messy and forgettable. A wiki is not. The system distils everything down to what matters and files it under the right topic, so a thought from a session eighteen months ago sits on the same page as the note you wrote yesterday. When something doesn't fit an existing page, it opens a new one — no one has to curate it by hand.

This is the part that makes it a brain and not a search box. An agent doesn't just answer and vanish — it updates the wiki before it talks to you.

Four Hermes agents run against the same wiki and the same data — your articles, your chats, your sessions — but each looks at you from a different angle. Together they cover the money, the making, the mind, and the writing.

Post reads across your articles and builds and looks for the pattern that actually converted. It doesn't guess: it cross-checks against live web research and your own numbers, so the play it hands you comes with proof attached rather than a hunch.

Build mines your Claude Code sessions for the things you made but never published, then points at the one worth finishing. In one run it pulled ideas scraped from sessions you'd long forgotten, spun up a fresh SEO page, and tied the suggestion back to an old project — cloning a site that had generated $50K — with the money proof already stapled on.

Under that suggestion sits an SEO Agent — Ideas link: the raw builds scraped from past conversations, cross-checked against your article data. It even read your post statistics and flagged that one earlier article had converted 3 readers into paid subscribers — so this one has conversion potential too.

The soft part. You work a lot; sometimes you break. To make coaching actually land, the system analysed all of your Claude.ai conversations — using Claude itself — to build a profile of you: how you act when you're angry, what drains you, how you sound when you get pissed. It saved that to a page called Leo-Psychology.

Every day the Stoic agent reads from that page plus your daily journal, then writes notes back to you — the ones that catch the exact mindset you keep losing.

Now the pragmatist. Note reads every X/Substack note you've ever written plus all their stats — shares, likes, restacks — and compiles new wiki pages with Karpathy's method. Then it does its own web research to make sure a note actually solves a problem, and suggests the next one to write.

Raw chats are full of noise. Before anything is filed, each conversation is stripped down to what matters — your messages, your decisions, what you actually built — and turned into one clean note. Then the system reads across all of them and builds a profile of you.

# Turn every chat into one clean note, then a profile of you
Read all my conversations in [folder].
For each one, keep my messages and the real decisions, drop the tool-noise.
Then write me a profile: how I think, what I get excited about, what blocks me,
how I work. Use real quotes.
Save it as a single page.

Before processing, exclude any conversation that mentions
[names, topics, clients you want kept out].
Check both the title and the body. When in doubt, drop it.

Then do the same with your building brain. Your Claude Code history already sits on your machine in ~/.claude/projects/ — same idea, different source: what you actually made, not just what you talked about.

Read my Claude Code sessions in ~/.claude/projects/.
Keep my prompts and what I built, skip the tool-calls and any client project.
Write one clean note per session, and flag the things I built but never shipped.

Here is one Build suggestion, traced end to end. It used your previous Claude Code conversations, created a new SEO page, and referred back to a past project — cloning a website that generated $50K.

SEO Agent — Ideas  # the link right under the build
  ↳ scraped from past Claude Code conversations
  ↳ builds you created but never published as articles
  ↳ cross-checked against your article data
→ new page created: SEO
→ read X post stats: 1 earlier article converted 3 → paid
→ verdict: conversion potential ✓

You end up with four clean piles — a vault, how you think, what you build, who you are — and hand them all to Claude.

Build my second brain using the attached SKILL.md and SOUL.md.

Here's my data:

Obsidian vault: [path to your vault]
Claude.ai export: [path to your unzipped export]
Claude Code history: ~/.claude/projects/
My own work + numbers: [path to your articles / notes / analytics]
Exclude anything that mentions: [names, clients, private topics] — check titles AND bodies.

Process my conversations into clean notes, build my psychology profile, then
compile the Karpathy wiki. Before you wire up any agents, show me the
connections you found and let me approve them one by one. Don't go fully
autonomous until I've reviewed one test run.

Then it gets good. The skill reads everything and starts suggesting connections — and waits for your yes or no:

Every yes builds another connection into the wiki. That's the brain forming in front of you.

The brain is built. Now you make it run on its own. Either path ends the same way: you wake up to fresh thinking every morning.