A lot of people are building second brains with Andrej Karpathy's LLM Wiki pattern, where an LLM compiles your raw sources into an interlinked wiki. It's a smart pattern, but a graph with thousands of connected notes isn't the point.
That's what I mean by indexing my brain. I delegate everything I can to AI, from voice notes and screenshots to transcripts, links and PDFs. And I keep a clear, indexed model of the system in my own head, so I know what lives where and I'm not helpless without it.
What makes that work is a clear breakdown of roles.
The four layers
| Layer | The question it answers |
|---|---|
| Entities | What are the pieces, and what is each one responsible for? |
| Workflows | How do the pieces connect to get a task done? |
| System | What results does all of it unlock? |
| Domains | Where in my life and work do I apply it? |
You direct the system. Workflows connect the pieces, and results serve goals in your own life.
Most setups skip straight to tools. Starting from roles keeps the system understandable, to the AI and to you.
Entities: give everything one job
| Entity | Its role | Examples |
|---|---|---|
| You | Judgment, taste, goals and final decisions | What matters, what's good, what ships |
| Sources | Where raw input comes from | Articles, videos, conversations, voice notes, screenshots |
| Data stores | Where things live after capture | A notes app, folders, a spreadsheet, a database |
| Tools and agents | What does the work | Transcribing, summarizing, tagging, searching, drafting |
The rule that matters most: each entity has one job, and each type of information has one home. When the same notes live in two places, they drift apart, and the AI finds whichever version it hits first.
Workflows: connect the entities
A workflow is a verb that moves information between entities. A handful cover most needs:
- Capture: a source becomes something stored, like a voice note turned into a transcript in an inbox.
- Index: each item gets tagged by domain and type, with a link back to where it came from.
- Retrieve: an agent finds what's relevant and shows its sources.
- Produce: retrieved material becomes a draft, a brief or a decision.
- Review: clear the inbox, archive what's stale and fix what's misfiled.
For each workflow, define what triggers it, what goes in, what comes out, where the result is stored, and whether you approve it before it counts.
System: tie it to results
The system layer answers one question: what does this let you do that you couldn't before?
- Find something you saved months ago in seconds, with the source attached.
- Turn scattered ideas into finished work instead of a growing backlog.
- Make decisions with your own past notes and research in front of you.
- Carry less in your head, because you trust where things go.
Results beat impressive workflows. A setup can look complex and polished from the outside and still be a tangle with no clear value. If a workflow doesn't show up in your week as time saved, work shipped or better decisions, cut it.
Domains: where you apply it
The same four layers work across very different parts of life:
- Work and projects: decisions, research, meeting context
- Learning: courses, papers, concepts to revisit
- Creative work: ideas, references, drafts
- Everyday life: documents, health records, finances, routines
Start with the domain where the pain is highest. Add the next only when the first one runs without you thinking about it.
What I've learned building mine
Hygiene matters more than features. Every capture tool makes it easy to add things. Very few make it easy to remove them. Regular cleanup, one source of truth and a clear archive keep retrieval trustworthy.
Pace matters too. Building everything at once produces a system you can't maintain. Add one workflow, use it until it's routine, then add the next.
Keep the index in your head. Delegate the storage and the grunt work, not the understanding. If a tool is down or a model changes, you should still know where things are and how the work gets done.
It's never finished. I keep adapting mine as models get better, as my circumstances change and as my goals shift.
This is why I keep stressing building your own secret sauce. It's great to get inspiration from what other people build, but a system you designed, understand and actually use beats one you copied and can't operate without.
To map your own, start here:
Interview me to map my personal AI system. Ask one question at a time.
Cover: the sources I pull information from, where I store things, the tools I use, the tasks I repeat, and the results I want.
Then give me four tables: entities with one role each, workflows with trigger, input, output and storage, the result each workflow serves, and the domains I apply it to.
Flag anything stored in more than one place and any workflow that doesn't serve a result.
Go deeper: Stop Building Your AI System Inside One Tool · How I Keep My Claude Code Setup From Drifting · How to Give AI Taste