The KB Builds Itself

The most surprising thing about a personal knowledge base powered by an LLM is that you do not have to maintain it manually. You work normally — attend meetings, make decisions, have conversations — and the KB accumulates structured knowledge in the background.

---

The Extraction Pattern

Every interaction with Claude is an opportunity to extract knowledge. The pattern is simple:

You say something. Claude responds AND extracts signals to your KB.

This is not a separate step. It happens during normal use. When Claude processes a meeting, it does not just write a meeting note — it also updates People profiles, adds items to your Work Board, and creates decision records. The extraction is built into the skills.

You: "process my meeting"
         │
         ▼
┌─────────────────────┐
│  Claude processes    │
│  the transcript      │
│                      │
│  Writes: Meeting note│──────▶  Meetings/2025-01-15_Platform-Sync.md
│                      │
│  Extracts:           │
│  ├─ People signals   │──────▶  People/Elaine-Benes.md (position updated)
│  ├─ Action items     │──────▶  Work_Board.md (new tasks added)
│  ├─ Decision         │──────▶  Decisions/DEC-001_Import-Export.md
│  └─ Project context  │──────▶  Projects/Import-Export/status.md
└─────────────────────┘

One input. Five outputs. You did one thing (said "process my meeting"). Claude did five things in the background. Your KB is now richer than it was two minutes ago.

---

What Gets Extracted

Different types of signals get extracted to different places:

People Mentioned

When someone is discussed or present in a meeting, Claude updates their People profile:

• Positions expressed: "Elaine advocated for delaying the migration by one sprint"
• Concerns raised: "George worried about the performance impact on legacy clients"
• Communication style notes: "Newman prefers written proposals before meetings"
• Relationship context: "Reports to Steinbrenner, collaborates closely with the gateway team"

Decisions Made

When a decision is reached — explicitly or implicitly — Claude captures:

• What was decided
• Who was involved in the decision
• What alternatives were considered
• What the rationale was
• Links to the meeting where it happened

Action Items

Tasks, follow-ups, and commitments get pulled to your Work Board:

• What needs to be done
• Who owns it
• When it is due
• Which meeting or conversation it came from

Patterns Observed

Over time, Claude starts noticing patterns across meetings:

• "This is the third meeting where timeline concerns were raised about Import-Export"
• "Elaine and George have disagreed on the migration approach in the last two syncs"
• "The team has been discussing but not deciding on the testing strategy for three weeks"

These are not logged as separate files — they surface in Claude's responses when you ask about a topic or person.

---

How It Compounds

The real value is not in any single extraction. It is in how knowledge compounds over time.

Week 1: Individual Facts

Your KB has a few meeting notes and People profiles. Claude knows basic facts: who attended what, what was decided, who owns which action item.

Elaine-Benes.md:
  Role: Engineering Lead, Gateway Platform
  Recent: Attended platform sync, owns Import-Export migration spike

At this stage, Claude is basically a note-taking assistant. Useful, but not transformative.

Week 3: Connected Patterns

After processing 10-15 meetings, Claude starts connecting dots. People profiles have multiple data points. Decisions link to the meetings where they were discussed. Patterns emerge.

Elaine-Benes.md:
  Role: Engineering Lead, Gateway Platform
  Reports to: Steinbrenner
  Key positions:
    - Advocates for incremental migration (3 meetings)
    - Concerned about testing coverage (raised twice)
    - Prefers async communication for technical decisions
  Working on: Import-Export migration (lead), Platform monitoring (contributor)
  Recent interactions: Platform sync, 1:1 with you, architecture review

Now when you ask "what should I discuss in my 1:1 with Elaine?", Claude pulls from weeks of accumulated context — not just the last meeting.

Month 2: Grounded Responses

After two months, your KB has enough data to provide genuinely insightful answers:

• "How does the team feel about the migration timeline?" — Claude synthesizes positions from multiple people across multiple meetings
• "What decisions have we been postponing?" — Claude identifies decisions that were discussed but never resolved
• "Who should I involve in the testing strategy discussion?" — Claude knows who has expressed opinions and who has relevant expertise

The responses are not generic AI advice. They are grounded in YOUR specific context, YOUR team's actual positions, YOUR real decisions. This is the compound effect — each meeting note makes every future response slightly better.

---

You Stay in Control

The extraction is not a black box. You control exactly what gets extracted and how.

Customize in CLAUDE.md

Your config file defines extraction rules:

## Extraction Rules

When processing meetings, extract:
- Positions: what each person advocated for
- Concerns: what worried them
- Decisions: what was agreed upon
- Action items: tasks with owners and deadlines

Do NOT extract:
- Personal opinions about people
- Salary or compensation discussions
- Off-the-record comments marked as such

Override Per Conversation

You can always tell Claude to adjust:

• "Process this meeting but don't extract anything about the reorg discussion"
• "Update Elaine's profile but skip the part about her team change"
• "Add this to the Work Board but mark it as low priority"

Review the Output

Every extraction results in a file you can read. Nothing is hidden in a database or an opaque system. Open the People profile — see exactly what Claude wrote. Open the Work Board — see exactly what was added. You can edit anything, delete anything, or correct anything.

---

The 80% Rule

Your KB does not need to be perfect. It needs to be approximately right and consistently maintained.

Review for 30 seconds. After Claude processes a meeting, glance at the output. Are the attendees correct? Are the key decisions captured? Are the action items right?

Correct when needed. If Claude attributed a statement to the wrong person, fix it. If it missed a decision, add it. If it added an action item you do not actually need to do, remove it.

Corrections make it better. When you correct something, Claude can learn from it (through the memory system). If it consistently gets a person's name wrong, tell it once and it remembers. If it keeps extracting the wrong level of detail, adjust the extraction rules in CLAUDE.md.

80% accuracy with corrections is better than 0% because you never started. The most common failure mode is not "Claude got something wrong" — it is "I never processed the meeting at all and now I have no notes."

A slightly imperfect KB that gets used every day beats a theoretically perfect system that never gets built.