Ultrathink

Claude Code

Version: 1.0.0 | License: MIT

Ultrathink Skill

Enhanced reasoning mode that activates deeper analysis, multi-approach generation, and multi-agent coordination for complex tasks.

Usage

Add "ultrathink" keyword to any command to activate enhanced mode:

# With slash commands
/plan ultrathink: design microservices architecture
/cook ultrathink: refactor authentication system
/code ultrathink: implement complex algorithm
/debug ultrathink: investigate memory leak

# Mid-sentence
/plan I need ultrathink for this database migration
Please use ultrathink to analyze this performance issue

What Happens

When ultrathink activates:

Deep Analysis - Exhaustive context gathering before action
Multi-Approach - Generate 3+ distinct approaches with trade-offs
Agent Coordination - Spawn specialized agents for parallel work
Validation - Rigorous edge case checking and self-critique

Modifiers

Modifier	Effect	Use When
`deep`	Extra context, 5+ approaches	Major architectural decisions
`fast`	Reduced validation	Time-sensitive with acceptable risk
`safe`	Extra validation passes	Production changes

/plan ultrathink deep: enterprise data migration
/cook ultrathink fast: quick feature with tests
/code ultrathink safe: production hotfix

Examples

Architecture Design

/plan ultrathink: design event-driven architecture for order processing

Expected output:
- Context: Current sync architecture, traffic patterns, pain points
- Approaches:
  1. Add message queue to existing (conservative)
  2. Event sourcing with CQRS (balanced)
  3. Full microservices decomposition (comprehensive)
- Selection: Approach 2 with justification
- Implementation plan with agent assignments
- Validation checkpoints

Complex Refactoring

/cook ultrathink: refactor payment processing module

Expected output:
- Context: Current payment flows, integrations, edge cases
- Approaches:
  1. Extract to service class (conservative)
  2. Strategy pattern + dependency injection (balanced)
  3. Hexagonal architecture (comprehensive)
- Selection with reasoning
- Step-by-step implementation
- Test coverage plan

Performance Investigation

/debug ultrathink: analyze slow API response times

Expected output:
- Context: Endpoint analysis, database queries, caching
- Approaches:
  1. Query optimization only (conservative)
  2. Add caching layer (balanced)
  3. Async processing + CDN (comprehensive)
- Root cause analysis
- Performance benchmarks
- Implementation recommendations

Complex Algorithm

/code ultrathink: implement distributed rate limiter

Expected output:
- Context: Requirements, scale, consistency needs
- Approaches:
  1. Token bucket per-node (conservative)
  2. Sliding window with Redis (balanced)
  3. Distributed consensus-based (comprehensive)
- Selected implementation with tests
- Edge case handling
- Performance characteristics

Output Format

Ultrathink responses follow this structure:

## Ultrathink Analysis

### Context Gathered
- Files reviewed: [list]
- Patterns identified: [list]
- Constraints: [list]

### Approaches Considered
1. **Conservative**: [desc] - Risk: Low, Effort: Low
2. **Balanced**: [desc] - Risk: Medium, Effort: Medium
3. **Comprehensive**: [desc] - Risk: Medium-High, Effort: High

### Selected Approach
[Name] - [Justification with reasoning]

### Implementation
[Detailed work]

### Validation
- [x] Self-critique completed
- [x] Edge cases: [list checked]
- [x] Tests: [status]

### Confidence: [High/Medium/Low]
[Notes on uncertainties if any]

When to Use Ultrathink

Good candidates:

Architecture decisions affecting multiple components
Refactoring with unclear scope
Debugging without obvious root cause
Features with complex requirements
Performance optimization with multiple options
Security-sensitive implementations

Skip ultrathink for:

Simple bug fixes with obvious cause
Routine CRUD operations
Minor UI tweaks
Well-defined, narrow tasks

Tips

Be specific in your request - ultrathink needs clear goals
Include constraints - budget, timeline, compatibility requirements
Accept the time - quality takes longer, that's the point
Review approaches - you can influence the selection
Check confidence - low confidence means more validation needed

Ultrathink

On this page