SuperThink

Confidence-driven workflow preventing misaligned execution: invest 100-200 tokens upfront to save 5,000-50,000 tokens.

When to Use

USE for:

Complex features requiring changes across multiple files/systems
Architectural decisions with long-term implications
Unclear requirements needing exploration
Multi-system integration
Breaking changes
Security-sensitive implementations

SKIP for:

Simple, well-defined single-file changes
Obvious bug fixes with clear solutions
Routine tasks (formatting, linting, etc.)

Core Philosophy

Confidence-driven workflow: Requires ≥90% confidence score before proceeding to implementation.

Context Gathering (30% effort) - Understand what exists
Confidence Assessment (20% effort) - Validate readiness
Structured Planning (30% effort) - Break down approach
Implementation (20% effort) - Execute with tracking

Five-Phase Workflow

Phase 1: Initial Context (30 tokens)

Parse requirements | Identify systems | Create TodoWrite tasks

Phase 2: Deep Context Gathering (60 tokens)

Search codebase (Grep/Glob) | Read existing code | Check docs | Analyze dependencies

Phase 3: Confidence Assessment (40 tokens)

| Component | Weight | Validates | | ---------------------------- | ------ | ---------------------------- | | No Duplicate Implementations | 25% | Existing functionality check | | Architecture Compliance | 25% | Pattern alignment | | Documentation Verified | 20% | API/library validation | | Data Model Understood | 20% | Schema/type confirmation | | Dependencies Mapped | 10% | Side effects understood |

Requires ≥90% score to proceed.

Phase 4: Structured Planning (50 tokens)

Order by dependency | Follow existing patterns | Include validation | Expand TodoWrite

Phase 5: Execution with Validation (Variable tokens)

Mark in_progress (ONE at a time) | Execute | Validate immediately | Mark completed immediately | Handle errors at root cause

TodoWrite Hygiene

Create tasks at Phase 1
Expand at Phase 4
Mark completed after each step
ONE in_progress at a time

Success Metrics

60-70% fewer tokens | Zero failures | First-try success