second-brain（第二大脑）

Second Brain Skill

Conversational interface to the Second Brain personal knowledge management system. Capture thoughts naturally during Claude Code sessions, query your knowledge graph, and manage your inbox.

Core Philosophy

Capture at the point of thinking, not after.

This skill enables seamless capture during work sessions without breaking flow:

• Capture thoughts as they emerge
• Query past decisions and notes
• Surface today's priorities
• Track what needs review

The system remembers so you don't have to.

Core Capabilities

1. Capture

Capture thoughts, tasks, ideas, and references directly from conversation.

Usage patterns:

• "Remember that the API rate limit is 1000 req/min"
• "Add a task to review the PR from Sarah"
• "Note: decided to use Supabase for sync"
• "Capture this idea: what if we..."

Classification: The system uses AI to classify captures into types:

• task: Actionable item with completion state
• idea: Non-actionable insight worth remembering
• reference: Information for later retrieval
• meeting: Time-bound event with notes
• goal: Outcome you're working toward
• project: Collection of related work
• value: Core principle that guides decisions
• person: Relationship context

Confidence threshold:

• High confidence (≥0.6): Auto-classified
• Low confidence (<0.6): Sent to needs_review

2. Query

Search and explore your knowledge graph.

Usage patterns:

• "What did I say about authentication?"
• "What projects support the 'shipping velocity' goal?"
• "Show me tasks related to the SecondBrain project"
• "Who did I meet with about the budget?"

Query types:

• Semantic search: Find by meaning, not just keywords
• Graph traversal: Follow relationships (supports, blocks, contains)
• Filter by type: "Show me all ideas from this week"
• Filter by domain: "What work tasks are due?"

3. Inbox

Review and triage pending captures.

Usage patterns:

• "Show my inbox"
• "What's waiting for review?"
• "How many pending captures?"

Inbox states:

• pending: Awaiting AI classification
• needs_review: Low confidence, needs human decision
• processing: Currently being classified

4. Digest

Get actionable summaries of what matters.

Usage patterns:

• "What should I focus on today?"
• "Show me today's digest"
• "What's overdue?"

Digest includes:

• Due tasks (today and overdue)
• High priority items
• Today's meetings
• Items needing review
• Recent insights

Constraints:

• Daily digest: <150 words
• Weekly review: <250 words

5. Actions

Update status and manage nodes.

Usage patterns:

• "Mark the PR review task done"
• "Complete task abc123"
• "Archive the old project"
• "Set priority to high for..."

Supported actions:

• Complete/reopen tasks
• Update priority (0-4)
• Change status (active, completed, archived)
• Add domain tag (work, personal, both)

Workflow Integration

During Work Sessions

When user mentions something capture-worthy during natural conversation:

1. Recognize capture intent:

- Direct: "Remember this...", "Add a task..." - Implicit: "I should...", "Don't forget...", "Note to self..."

1. Capture with context:

- Include relevant context from current conversation - Tag with source: "cli" (Claude Code session) - Add any mentioned relationships

1. Confirm capture:

- Brief confirmation with ID - Mention if needs_review due to low confidence

Quick Actions

For common actions, provide shortcuts:

/sb capture "thought or idea"     # Capture immediately
/sb inbox                         # Show pending items
/sb digest                        # Today's actionable summary
/sb query "search term"           # Search knowledge base
/sb done <id>                     # Mark task complete

Meeting Transcript Processing

Use case: Paste meeting transcripts to automatically extract and capture structured content.

Workflow

1. User pastes transcript:

   "Here's the transcript from today's standup:
   [transcript content]"

1. System processes and extracts:

- Meeting summary → MEETING node - Action items → TASK nodes - Decisions made → REFERENCE nodes - People mentioned → PERSON links - Follow-up meetings → MEETING nodes - Key insights → IDEA nodes

1. Confirmation and review:

   Processed standup transcript:

   📅 Created: "Daily Standup - Jan 15" (meeting)

   ✅ Extracted 4 tasks:
   - "Review PR #1234" → assigned to you (high)
   - "Update API docs" → assigned to Sarah
   - "Schedule design review" → due Friday
   - "Fix login bug" → blocked by infrastructure

   📝 Captured 2 decisions:
   - "Using Postgres instead of MongoDB"
   - "Sprint ends Friday, demo at 2pm"

   👥 Linked 3 people:
   - Sarah (mentioned 5x)
   - Mike (mentioned 2x)
   - John (new person, created)

   💡 1 insight flagged:
   - "Team velocity improves when standups are <15 min"

Extraction Patterns

Action items (→ TASK):

• "TODO: ...", "Action: ...", "Need to..."
• "Sarah will...", "I'll...", "We should..."
• "@mentions with action verbs"

Decisions (→ REFERENCE):

• "Decided: ...", "Agreed: ..."
• "We're going with...", "The plan is..."
• "Final decision: ..."

Follow-ups (→ MEETING):

• "Let's meet again...", "Schedule a follow-up..."
• "Next week we'll discuss..."
• Explicit dates/times mentioned

People (→ PERSON links):

• Names mentioned in context
• @mentions
• "talked to...", "asked..."

Insights (→ IDEA):

• Observations about patterns
• Hypotheses mentioned
• "I noticed...", "Interesting that..."

Post-Processing

After extraction:

1. Create meeting node with summary
2. Create task nodes with assignments and due dates
3. Link people (create if new)
4. Store decisions as references
5. Write to Obsidian with wikilinks

Configuration

# ~/.config/secondbrain/daemons.yml
transcript_processing:
  auto_assign_unassigned: true  # Assign to self
  default_task_priority: 2
  flag_low_confidence: true     # Mark uncertain extractions
  link_to_meeting: true         # Connect all items to meeting node

Graph Model

Node Types

Edge Types

Implementation

CLI Integration

This skill wraps the sb CLI commands:

sb capture "content"      # Capture a thought
sb inbox                  # List pending captures
sb process                # Classify pending captures
sb digest                 # Generate daily digest
sb list [type]            # List nodes
sb show <id>              # Show node details
sb status                 # System health check

Database

• Local SQLite at ~/.local/share/secondbrain/secondbrain.db
• Obsidian vault for markdown output
• Graph model: nodes + typed edges

Configuration

# ~/.config/secondbrain/config.yml
node_id: "home"
vault_path: "/path/to/vault"
classification:
  model: "claude-sonnet-4-20250514"
  confidence_threshold: 0.6

Context Awareness

Use Session Context

When capturing, leverage current conversation context:

• File being discussed: Link to related codebase notes
• Error being debugged: Capture as reference with solution
• Decision being made: Note with rationale
• Meeting notes mentioned: Parse for action items

Domain Detection

Infer domain from context:

• In work project directory → domain: work
• Personal topics mentioned → domain: personal
• Both contexts mixed → domain: both

Example Interactions

Capture During Work

User: "The deadline for the API migration is March 15th"
Assistant: [Captures as task with due date]
Captured: API migration deadline - Due March 15th (ID: abc123)

Query Knowledge

User: "What did we decide about authentication?"
Assistant: [Searches knowledge graph]
Found 3 related items:
- [reference] "Use JWT with refresh tokens" (Jan 5)
- [task] "Implement OAuth flow" (in_progress)
- [meeting] "Auth design review with Sarah" (Jan 3)

Morning Digest

User: "What should I focus on today?"
Assistant: [Generates digest]

📋 Today's Focus (Wed, Jan 15):
- [OVERDUE] Review PR from Sarah (due Jan 14)
- [HIGH] Prepare API migration plan
- [DUE TODAY] Call dentist to reschedule

📅 Meetings:
- 10:00 1:1 with Mike

⚠️ Needs Review: 2 items in inbox

Anti-Patterns

Don't:

• Capture every single thing mentioned (be selective)
• Force classification when context is unclear
• Interrupt flow for minor captures
• Create duplicate entries for same concept
• Over-classify simple notes

Do:

• Capture when user expresses intent or importance
• Ask for clarification if capture intent is ambiguous
• Batch confirmations when capturing multiple items
• Link to existing nodes when relationships are clear
• Respect user's domain boundaries

Integration Points

With beads issue tracker:

• Cross-reference tasks with beads issues
• Import epic/task relationships

With Obsidian vault:

• Generated markdown syncs via Obsidian Sync
• Wikilinks enable navigation
• Daily notes include digest

With SiliconDoppelgangerActual:

• Deep queries via agent conversation
• Complex graph traversals
• Multi-step reasoning about priorities

Success Metrics

Skill succeeds when:

• Captures happen naturally without flow interruption
• User finds past information quickly
• Daily digests surface actionable items
• Inbox stays manageable (<10 items needing review)
• Classification accuracy >85%

User feels:

• Confident nothing important is lost
• Informed about what matters today
• In control of their knowledge system