Appearance
Building a Personal Knowledge Base
Context engineering, not a quick fix. Every meeting you process, every person file you update, every decision you capture makes Claude smarter about your work. The setup takes ~20 minutes. The value compounds over weeks.
What's Inside
This guide walks you through setup, daily use, and advanced patterns. Start at the top and work down, or jump to what you need:
- Before You Start – Verify tools are connected, learn how to talk to Claude (~3 min)
- Setting Up Your Workspace – Install the plugin, run the setup (~10 min)
- Your First Skills – Process meetings, file documents, capture notes
- Daily and Weekly Rituals – Morning briefing, end-of-day, weekly review
- Adding Tools and Customizing – Connect more tools, customize CLAUDE.md
- Workflows and Patterns – Meeting prep, stakeholder updates, research briefs
- Power Tips and Advanced Patterns – Custom skills, community marketplace, advanced patterns
- Real Examples – Actual output from processing meetings, filing docs, running workflows
- How It Works – Architecture, layers, and how the KB builds itself
Why Build a Knowledge Base?
The same AI, with and without your context. The difference is everything.
Every knowledge management system eventually gets abandoned. Not because people don't want organized knowledge, but because organizing knowledge is work. The maintenance cost exceeds the perceived benefit, and the system quietly dies.
PKB bets on a different outcome: what if the garden had a gardener?
The gardener is your AI agent. It files what you capture. It links what you name. It archives what you finish. You plant seeds by doing your work. The gardener cultivates what you plant.
How knowledge compounds
After 3 meetings
📁 People/
George.md
Elaine.md
Kramer.md
📁 Meetings/ 3
📁 Decisions/ 1
What you can ask Claude
"What did George say about the architecture?"
→ Pulls his position from yesterday's standup
"Summarize today's meetings"
→ TL;DR of each, action items collected
"Prep me for my 1:1 with George"
→ Not enough history yet. After 5+ meetings.
keep going
After 20 meetings
📁 People/ 12
George.md ×8
Elaine.md ×6
+ 10 more
📁 Meetings/ 20
📁 Decisions/ 7
📁 Projects/ 3
What you can ask Claude
"Prep me for my 1:1 with George"
→ 8 meetings together, 3 open actions, his position shifted since Week 2
"Who's aligned on the gateway decision?"
→ Elaine: strong yes. George: shifted from no to neutral. Kramer: hasn't weighed in.
"Draft a status update for leadership"
→ Project history, recent decisions, open blockers. First draft in 30 seconds.
Same 5 minutes per meeting. The knowledge compounds.
After 3 meetings, Claude answers questions. After 20, it connects dots. After 80, it knows your work.
The 80% Rule
AI gets to about 80%. Your expertise fills the rest: which decisions actually matter, what was said between the lines, which action items are real commitments vs. polite nods. Expect to review and edit. The value isn't perfection. It's avoiding blank pages and scattered context.
What this is NOT
- Not a new app to learn. It's local markdown files + your AI agent. No vendor lock-in, no data leaving your machine.
- Not a finished system. You'll customize CLAUDE.md, add skills, adjust folders over time. It grows with you.
- Not instant magic. The value is in accumulated context, not any single interaction. Give it two weeks before judging.
Why I built this (the backstory)
Like most PMs, my context was scattered across 15 tools: project trackers, wikis, docs, chat, video recordings, email, data warehouses, spreadsheets, meeting notes in various formats. Every time I needed to prepare for a meeting or recall a decision, I spent 20-30 minutes just finding the pieces.
AI chatbots helped with individual tasks, but every session started from zero. I'd explain my role, my projects, my stakeholders over and over. The 100th conversation was no smarter than the first.
The idea: what if an AI agent could read my accumulated context instead of starting fresh? What if processing a meeting also created people profiles, logged decisions, and updated my task board? What if each interaction made the next one better?
That's what this system does. It's structured context that compounds. The AI still gets things wrong (80% accuracy, your expertise fills the rest). But the difference between a generic AI and one that knows your stakeholders, your projects, and your decisions is enormous.
Take what's useful, ignore what isn't.
Start here: Before You Start →