Routing Brains

Routing Brains Skill

Purpose

Route meeting items to appropriate Brain CLAUDE.md files:

1. Discover Brains dynamically from Brains/ folder
2. Match items using keyword and semantic scoring
3. Handle unmatched content with user prompts (one at a time)

For product-specific context, see CLAUDE.local.md.

Brain Discovery Protocol

Step 1: Scan Brains Folder

CRITICAL: Never hardcode Brain names. Always discover dynamically.

1. List all subdirectories in /Brains/
2. For each subdirectory:
   a. Check if CLAUDE.md exists
   b. If yes, read and parse the file
   c. Extract keywords and context for matching
3. Build Brain registry with:
   - Brain name (folder name)
   - Core objective (from CLAUDE.md)
   - Key terms (extracted from all sections)
   - Status (Active/Paused/Completed)

Step 2: Extract Keywords

From each Brain's CLAUDE.md, extract:

Step 3: Build Matching Index

Brain: "[Brain-name]"
Keywords: [relevant terms extracted]
Core Theme: "[one-line description]"
Status: Active

Matching Framework

Confidence Levels

Matching Algorithm

For each meeting item:

1. Keyword Match Score (0-0.5):

   • Count keyword overlaps with Brain keywords
   • Weight Core Objective keywords 2x
   • Normalize by total keywords

2. Semantic Match Score (0-0.5):

   • Compare item theme to Brain Core Objective
   • Consider context (what problem is being solved?)
   • Factor in stakeholder mentions

3. Combined Score: Keyword + Semantic

Multi-Brain Matching

Items CAN match multiple Brains if:

• Score >= threshold for multiple Brains
• Item explicitly mentions multiple workstreams
• Action item affects multiple teams

Item Segmentation

Extract discrete items from meeting content:

Routing Protocol

Step 1: Match Each Item

For each item:
  1. Run matching algorithm
  2. If High confidence (>= 0.8) → Route to Brain
  3. If Medium confidence (0.6-0.8) → Route with [Review] flag
  4. If Low confidence (0.4-0.6) → Add to potential matches
  5. If No match (< 0.4) → Add to unmatched list

Step 2: Handle Multi-Brain Items

If item matches multiple Brains:

1. Route to ALL matching Brains
2. Note cross-Brain dependency
3. Flag in both Brain summaries

Unmatched Content Handling

Sequential Questioning Protocol

CRITICAL: Ask ONE question at a time. Do not overwhelm the user.

Step 1: Announce count

I found [N] items that don't clearly match any existing Brain.
Let me walk through each one with you.

Step 2: Present FIRST item only

**Item 1 of [N]**: [Brief summary]

**Type**: [Decision/Action/Discussion/Blocker/Question]
**Content**: [Full item - 1-2 sentences]
**Context**: [Why this came up]
**Potential match**: [Brain] (low confidence) OR "No matching Brain"

What would you like to do?
a) Assign to [Brain 1]
b) Assign to [Brain 2]
c) Create a new Brain
d) Leave in temp folder
e) Discard (not relevant)

Step 3: Wait for response before continuing

Step 4: Proceed to next item

Got it - [acknowledge choice].
**Item 2 of [N]**: [Next summary]...

New Brain Creation (Guided)

If user chooses to create new Brain:

Step 1: "What should this Brain be called? (Suggested: '[name]')" → Wait for response

Step 2: "In one sentence, what's the core objective?" → Wait for response

Step 3: Create folder, generate CLAUDE.md, route item, confirm

Example Input

Items to route:
1. Decision: "Use React for dashboard migration"
2. Action: "Update API documentation before launch"
3. Discussion: "Potential partnership with external vendor"

Example Output

## Routing Results

### High Confidence Routes
- **Decision: React migration** → frontend-modernization (0.92)

### Medium Confidence Routes  
- **Action: API documentation** → api-platform (0.71) [Review]

### Unmatched Items
1 item requires user input:

**Item 1 of 1**: Partnership discussion
**Type**: Discussion Point
**Content**: Potential collaboration with external vendor
**Potential match**: No matching Brain

What would you like to do?
a) Assign to frontend-modernization
b) Assign to api-platform
c) Create a new Brain
d) Leave in temp folder
e) Discard

Quality Gates

• Brains discovered dynamically (not hardcoded)
• Each item categorized by type
• Matching algorithm applied consistently
• Multi-Brain items flagged
• Confidence levels logged
• Unmatched items presented one at a time

Anti-Patterns

• ❌ Hardcoding Brain names
• ❌ Force-matching items to Brains
• ❌ Auto-routing low-confidence matches
• ❌ Presenting all unmatched items at once
• ❌ Proceeding before user responds
• ✅ Always scan Brains/ folder fresh
• ✅ Ask ONE question at a time
• ✅ Wait for explicit confirmation

Success Criteria

1. All Brains discovered dynamically
2. Items routed with appropriate confidence
3. Multi-Brain dependencies flagged
4. Unmatched items handled with user input
5. Ready for summarizing-meetings skill