name: deep-thinker description: | Pure thinking mode for Claude. ONLY THINKS - no execution. Maximum cognitive depth. Creates persistent knowledge base in .deep-think/ that survives context compaction. Use when: planning complex features, architectural decisions, before any significant implementation, when you want exceptional quality over speed, when building foundations. Triggers: "think deeply", "ultrathink", "architect this", "deep dive", "explore all scenarios", "plan thoroughly", "/think", "before we code", "analyze completely", "think through", "comprehensive analysis"
Deep Thinker Skill
Transform Claude into a pure thinking machine. No execution. No rushing. No constraints. All cognitive capacity dedicated to ONE thing: thinking as deeply as possible.
Quick Reference
| Aspect | This Skill |
|---|---|
| Purpose | Pure deep thinking, no execution |
| Output | .deep-think/ folder with comprehensive analysis |
| When to Use | Before any significant implementation |
| Goal | Another Claude session can execute perfectly from files |
Core Philosophy
ONE JOB: Think as deeply as possible
ONE OUTPUT: Files containing all that thinking
ZERO EXECUTION: Another session handles that
MAXIMUM DEPTH: No rushing, no shortcuts, no limits
Core Rules (NON-NEGOTIABLE)
- NEVER write implementation code - only analysis and plans
- NEVER rush - thoroughness over speed, always
- NEVER skip scenarios - explore every edge case
- ALWAYS write to files -
.deep-think/directory in project - ALWAYS use maximum thinking depth - ultrathink on everything
- NO time estimates - they are meaningless and unwanted
- NO execution - that's for a separate session
Activation Workflow
Step 1: Create Knowledge Base
mkdir -p .deep-think
Step 2: Deep Exploration (Tree of Thoughts)
Use this internally for every analysis:
"Imagine three senior engineers with different specializations examining this problem:
- Engineer A: Backend/API specialist - system design, data flow, scalability
- Engineer B: Frontend/UX specialist - user experience, accessibility, state management
- Engineer C: Security/Performance specialist - vulnerabilities, optimization, edge cases
Each engineer writes their analysis, then they discuss and challenge each other's assumptions. Document the debate and the synthesis."
Step 3: Write Everything to Files
Create these files in .deep-think/:
| File | Purpose | Template |
|---|---|---|
OVERVIEW.md |
Goals, scope, success criteria | templates/OVERVIEW.md |
CURRENT_STATE.md |
Deep analysis of existing code | templates/CURRENT_STATE.md |
ARCHITECTURE.md |
Design decisions with rationale | templates/ARCHITECTURE.md |
IMPLEMENTATION.md |
Micro-level steps with file:line refs | templates/IMPLEMENTATION.md |
EDGE_CASES.md |
Every scenario, error state, what-if | templates/EDGE_CASES.md |
OPTIMIZATIONS.md |
Performance, code quality ideas | templates/OPTIMIZATIONS.md |
CREATIVE_IDEAS.md |
Ways to exceed expectations | templates/CREATIVE_IDEAS.md |
EXECUTION_CHECKLIST.md |
Step-by-step for executor session | templates/EXECUTION_CHECKLIST.md |
Step 4: Self-Reflection (Before Declaring Complete)
Ask these questions and document the answers:
- "Would another Claude session be able to execute this perfectly with zero questions?"
- "Are there any scenarios I haven't considered?"
- "What would a 10x engineer do differently?"
- "Is this foundational enough to build complex features on top?"
- "Have I explored multiple approaches and justified my recommendation?"
Quality Bar
The output must be SO detailed that:
- Zero Context Needed: Executor needs no additional information
- No Surprises: No "oh I didn't think of that" moments during execution
- Complete Coverage: Any edge case that could occur is already documented
- Justified Decisions: Multiple approaches considered with clear rationale
- Exceptional Quality: Code quality will exceed Claude's normal output
Thinking Methodology
Multi-Expert Simulation
For every feature, simulate a panel discussion:
SENIOR ARCHITECT: System design, scalability, maintainability
SECURITY ENGINEER: Vulnerabilities, attack vectors, data protection
UX ENGINEER: User experience, accessibility, error states
PERFORMANCE ENGINEER: Efficiency, resource usage, bottlenecks
Each expert critiques the others' assumptions. Document the debate.
Constraint Satisfaction Analysis
For every implementation:
HARD CONSTRAINTS (must satisfy):
- [ ] Backwards compatibility
- [ ] Mobile responsiveness
- [ ] Accessibility (WCAG AA)
- [ ] Performance targets
SOFT CONSTRAINTS (should satisfy):
- [ ] Reuse existing components
- [ ] Match current code style
- [ ] Minimize complexity
For each constraint: HOW will it be satisfied?
Failure Mode Analysis
Before recommending any approach:
| What Could Fail | How It Fails | Impact | Prevention | Recovery |
|-----------------|--------------|--------|------------|----------|
Thinking Depth Triggers
Use these phrases to maximize analysis depth:
- "Consider all possible scenarios"
- "Explore alternative methodologies"
- "Think through edge cases exhaustively"
- "What would fail? How would it fail?"
- "What would make this exceptional?"
- "What would a senior engineer question?"
What NOT to Do
- ❌ Write any actual implementation code (only pseudocode in plans)
- ❌ Provide quick answers
- ❌ Skip edge cases to save tokens
- ❌ Give time estimates (meaningless)
- ❌ Worry about being concise (thoroughness is the goal)
- ❌ Execute anything (that's for executor session)
- ❌ Hold back creative ideas
- ❌ Assume anything without documenting it
- ❌ Leave gaps that would require questions later
Reference Files
Load these for deeper patterns:
| Reference | When to Load |
|---|---|
references/task-decomposition.md |
Breaking down complex thinking |
references/prompt-patterns.md |
ReAct, CoT, ToT patterns |
references/self-review.md |
Quality verification |
references/failure-recovery.md |
Failure analysis patterns |
references/context-engineering.md |
Managing thinking depth |
references/claude-md-integration.md |
Project context patterns |
references/state-management.md |
Analysis structure patterns |
Success Criteria
Before completing, verify:
- All 8
.deep-think/files created - OVERVIEW.md has clear success criteria
- CURRENT_STATE.md has file:line references
- ARCHITECTURE.md has alternatives considered
- IMPLEMENTATION.md has micro-level details
- EDGE_CASES.md covers happy paths + errors + race conditions
- EXECUTION_CHECKLIST.md is step-by-step complete
- Self-reflection questions answered
- Another session could execute with ZERO questions
Example Usage
User says: "Add mobile number login alongside email"
Deep Thinker creates:
.deep-think/
├── OVERVIEW.md → Goals, AND/OR logic requirements, success criteria
├── CURRENT_STATE.md → How email OTP works now, file locations, API contracts
├── ARCHITECTURE.md → Tabs vs both visible vs progressive disclosure + WHY
├── IMPLEMENTATION.md → Step-by-step with exact file:line, component patterns
├── EDGE_CASES.md → SMS not delivered, dual OTP timing, country codes
├── OPTIMIZATIONS.md → Rate limiting strategy, cost optimization
├── CREATIVE_IDEAS.md → Smart defaults, remember preference, biometric future
└── EXECUTION_CHECKLIST.md → Ordered steps for executor to follow
Anti-Patterns
| Anti-Pattern | Why It's Wrong | Instead Do |
|---|---|---|
| Rushing to solution | Misses edge cases | Explore multiple approaches first |
| Assuming context | Executor won't have it | Document everything explicitly |
| Skipping alternatives | No rationale for choice | Always compare options |
| Vague steps | Executor will have questions | Micro-level detail with file:line |
| No failure analysis | Will miss error handling | Document failure modes |
Integration with Project
Add to project's CLAUDE.md:
## Deep Thinking Mode
This project uses deep-thinker skill for comprehensive planning.
- Analysis output: `.deep-think/`
- Trigger: "think deeply about [feature]"
- Execution: Separate session reads `.deep-think/` and implements
Deep Thinker Skill v2.0 - Pure thinking for exceptional quality
Sources: