FOUNDTHEWEI
Back to all posts

How to Build a Second Brain That Your AI Can Actually Navigate

August 2, 20254 min read
AIKnowledge ManagementObsidianClaude

Context: Managing 210+ documents across multiple projects. One mixed-up context could cost me a client. This system prevents that.

Everything in our world can be translated into words or numbers. This simple realization is the key to building a knowledge system that AI can traverse as naturally as you navigate your own memories.

The Setup: Obsidian + Claude Code

I'm running 210+ documents through an Obsidian vault, but here's what makes it work:

CLAUDE.md documents in every folder.

These files tell AI:

  • ▸Where to find specific information
  • ▸Current file structure
  • ▸Folder-specific instructions
  • ▸[Added after incident]: Which client owns this context

Think of them as navigation beacons for your AI assistant when you're too tired to remember.

Sub-agents that process everything.

Using Claude Code, I created agents that:

  • ▸Review and chunk transcripts into logical pieces
  • ▸Maintain interconnected file structures
  • ▸Update documentation as the system evolves
  • ▸Keep different contexts from bleeding into each other

The Four Principles That Actually Matter

1Dense interconnection beats perfect organization

Link everything. The more pathways between documents, the faster AI finds what it needs.

Real application: When client asks about "that thing we discussed in March," I find it in seconds, not hours.

2Context is multi-dimensional

Words are just the start. Capture:

  • ▸Who was involved (which version of me was in that meeting?)
  • ▸When it happened (billable? already invoiced?)
  • ▸Why it matters (revenue impact? risk mitigation?)
  • ▸What situation prompted it (CYA documentation is essential)

Different documents extract different contexts. Use that redundancy.

3Version control isn't optional

Git everything. Your knowledge evolves—track that evolution.

Real commit history showing evolution of a knowledge system:

# Morning: Identifying issues
$ git commit -m "Fix vault data integrity issues identified by tools"

# Afternoon: Major feature release
$ git commit -m "Release v2.0: Add AI-powered context providers for integration"

# Testing and refinement
$ git commit -m "Fix and test automated tools with API key"

# Documentation updates
$ git commit -m "Review and document all analysis tools with clear limitations"
Code

Each commit is a snapshot of your knowledge state. When something breaks, you can trace back exactly when and why.

4Embrace messy evolution

Your structure will break. Projects will demand different setups. That's fine. Have your agent periodically review and update the navigation docs.

Perfect structure is procrastination. Working structure pays bills.

Real-World Results (and Problems)

What's working:

  • ▸AI gives responses that feel like they're from someone who knows my entire history
  • ▸Complex project context stays intact across conversations
  • ▸Cross-pollination of ideas happens naturally
  • ▸Haven't mixed up clients in 6 months (was happening weekly before)

Current challenges:

  • ▸Can't pinpoint where AI pulls specific facts from (dangerous when asked "where's that from?")
  • ▸Unknown if this scales beyond 1,000 documents (currently at 400+)
  • ▸Local models can't handle the context length (yet)
  • ▸Dependency on Claude API (single point of failure I'm working to eliminate)

The Bigger Picture

Yesterday's impossible is today's normal. Since ChatGPT 3.5, keeping up isn't optional anymore.

The goal isn't perfect capture—it's maximum navigability.

But here's what most miss: yes, there's data loss when translating reality into words and numbers. That's fine. An interconnected database is still the closest thing we have to a digital brain that AI can actually use.

More importantly: An 80% accurate external brain beats a 100% accurate brain that can't handle the cognitive load.

Start Here

  1. ▸Pick your tool (I use Obsidian - it's free and local)
  2. ▸Create navigation documents for AI
  3. ▸Build sub-agents for processing
  4. ▸Link aggressively
  5. ▸Accept imperfection

Because in this new dawn of AI, your knowledge is only as powerful as your AI's ability to traverse it.

And your income is only as stable as your ability to manage multiple contexts without bleeding them together.