Unified thinking toolkit: Creative frameworks (Critical Evaluation, Diverge-Converge, Deep Questioning), First Principles reasoning (absorbed from moai-foundation-philosopher), and Adaptive Thinking via the `ultrathink` keyword. Use for ideation, strategic analysis, architecture decisions, and deep reasoning on complex problems.
Unified thinking toolkit: Creative frameworks (Critical Evaluation, Diverge-Converge, Deep Questioning), First Principles reasoning (absorbed from moai-foundation-philosopher), and Adaptive Thinking via the `ultrathink` keyword. Use for ideation, strategic analysis, architecture decisions, and deep reasoning on complex problems.
when_to_use
Use for thinking and ideation: creative frameworks (critical evaluation, diverge-converge), first-principles reasoning, five-whys root-cause analysis, trade-off and architecture decisions, technology/library/database selection, and Adaptive Thinking via the ultrathink keyword.
Structured thinking toolkit for creative problem-solving and rigorous analysis. Integrates three complementary frameworks that cover the full spectrum from idea generation to critical evaluation.
Core Philosophy: Generate broadly, evaluate rigorously, question deeply. Creativity and criticism are complementary forces.
Quick Reference
What is the Thinking Toolkit?
Three integrated frameworks for structured thinking:
Critical Evaluation: Rigorous 7-step analysis to assess proposals and detect flaws
Diverge-Converge: Systematic brainstorming from 20-50 raw ideas to 3-5 validated solutions
Deep Questioning: 6-layer progressive inquiry to uncover hidden requirements and risks
When to Use Each Framework:
Evaluating a proposal or recommendation: Critical Evaluation
Generating solutions for an open-ended problem: Diverge-Converge
Exploring an unfamiliar domain or unclear requirement: Deep Questioning
Complex decisions: Combine all three (Question first, Generate second, Evaluate third)
Purpose: Systematically assess proposals, claims, and recommendations to detect flaws before commitment.
Seven-Step Evaluation Process:
Step 1 - Restate: Reformulate the claim or proposal in your own words. Ensures genuine understanding before critique.
Step 2 - Assess Evidence: Examine supporting data. Is the evidence empirical, anecdotal, or assumed? What is the sample size and recency? Are there contradicting data points?
Step 3 - Detect Fallacies: Check for common reasoning errors. Appeal to authority without substance. False dichotomy (only two options presented). Hasty generalization from insufficient examples. Straw man misrepresentation of alternatives.
Step 4 - Expose Assumptions: Identify unstated premises. What must be true for this conclusion to hold? Which assumptions are testable? Which assumptions carry the highest risk if wrong?
Step 5 - Note Alternatives: For every claim, ask what else could explain the evidence. Generate at least two alternative interpretations. Consider the null hypothesis.
Step 6 - Check Contradictions: Look for internal inconsistencies. Do different parts of the proposal conflict? Are there contradictions with known facts or constraints?
Step 7 - Evaluate Burden of Proof: Determine if the evidence is proportional to the claim. Extraordinary claims require extraordinary evidence. Identify what additional evidence would strengthen or weaken the case.
Critical Gaps: What needs more evidence or revision
Recommended Actions: Next steps to strengthen the proposal
WHY: Uncritical acceptance of proposals leads to preventable failures.
IMPACT: Structured evaluation catches 60-80% of flawed recommendations.
Framework 2: Diverge-Converge Brainstorming
Purpose: Generate a broad solution space then systematically narrow to the best options.
Five-Phase Process:
Phase 1 - Gather Requirements: Define the problem space clearly. Identify stakeholders and success criteria. Set explicit constraints (budget, timeline, technology). Document "must-have" vs "nice-to-have" criteria.
Phase 2 - Diverge (Generate 20-50 Ideas): Quantity over quality during divergence. No criticism or filtering during generation. Include wild and unconventional ideas. Combine and build upon previous ideas. Use prompts: "What if we...", "How might we...", "What would happen if..."
Phase 3 - Cluster (Group into 4-8 Themes): Identify natural groupings among ideas. Name each cluster with a descriptive theme. Note which clusters have the most ideas (signals interest). Identify gaps where no ideas exist (potential blind spots).
Phase 4 - Converge (Score and Select): Rate each cluster against success criteria (1-10). Apply weighted scoring based on priority of criteria. Select top 3-5 candidates for deeper analysis. Document why rejected options were eliminated.
Phase 5 - Document and Validate: Write up selected solutions with rationale. Define validation experiments for top candidates. Identify risks and mitigation strategies. Plan implementation sequence.
Output Format:
Problem Statement: Clear definition of what we are solving
Idea Count: Total ideas generated and cluster distribution
Top Candidates: 3-5 selected solutions with scores
Validation Plan: How to test each candidate
WHY: Premature convergence on the first idea leaves better solutions undiscovered.
IMPACT: Teams using diverge-converge find 3x more viable solutions.
Framework 3: Deep Questioning
Purpose: Progressively uncover hidden requirements, constraints, and risks through layered inquiry.
Six-Layer Progressive Inquiry:
Layer 1 - Surface Understanding: What is the stated goal or request? What does success look like? What are the obvious inputs and outputs? Verify: Can I explain this to someone else clearly?
Layer 2 - Problem Depth: Why does this problem exist? What is the root cause vs symptom? What has been tried before and why did it fail? What would happen if we did nothing?
Layer 3 - Context and Constraints: What are the technical constraints? What are the organizational or process constraints? What are the time and resource limitations? What external dependencies exist?
Layer 4 - User Perspective: Who are the actual end users? What is their current workflow? What pain points drive this request? What would they consider a disappointing solution?
Layer 5 - Solution Exploration: What are the boundary conditions? What edge cases could break the solution? What are the performance requirements? How will this integrate with existing systems?
Layer 6 - Validation and Risk: How will we know if the solution works? What could go wrong? What is the rollback strategy? What monitoring or alerting is needed?
Progressive Depth Indicators:
Shallow: Only Layers 1-2 explored (common in quick tasks)
Moderate: Layers 1-4 explored (sufficient for most features)
Deep: All 6 layers explored (required for architecture decisions)
Exhaustive: All layers with multiple iterations (critical systems)
Output Format:
Understanding Level: Shallow, Moderate, Deep, or Exhaustive
Key Discoveries: Insights from each explored layer
Open Questions: Remaining unknowns requiring further investigation
Risk Assessment: Identified risks by severity
WHY: Surface-level understanding leads to solutions that miss the real problem.
IMPACT: Deep questioning reduces requirement changes by 40-60%.
Combined Workflow
For complex problems, use all three frameworks in sequence:
Step 1 - Deep Questioning: Explore the problem space (Layers 1-4 minimum)
Step 2 - Diverge-Converge: Generate and select solutions based on discoveries
Step 3 - Critical Evaluation: Rigorously assess the top candidates
Decision Complexity Guide:
Simple task (1-2 files): Skip thinking frameworks (direct implementation)
Feature addition: Deep Questioning (Layers 1-3) + brief evaluation
Design decision: Deep Questioning (full) + Diverge-Converge
Architecture change: All three frameworks in full
Integration with MoAI Workflow
SPEC Phase (/moai plan):
Apply Deep Questioning during requirements gathering
Use Diverge-Converge for solution approach selection
Apply Critical Evaluation to finalize SPEC document
Run Phase (/moai run):
Use Critical Evaluation when reviewing implementation options
Apply Deep Questioning when encountering unexpected complexity
Agent Teams:
team-reader (analyst role): Primary user of Deep Questioning framework
team-reader (architect role): Primary user of Critical Evaluation framework
team-reader (researcher role): Uses all three for comprehensive analysis
Works Well With
Agents:
manager-spec: Combined with Philosopher for full decision framework (planning absorbs strategic analysis)
manager-develop: Deep Questioning + Critical Evaluation during run-phase reasoning (DDD/TDD root-cause analysis, architecture trade-offs)
Origin: Integrated from critical-thinking, brainstorm-diverge-converge, and ideation frameworks
Common Rationalizations
Rationalization
Reality
"Brainstorming is for creative tasks, not engineering"
Engineering decisions have design space too. Diverge phase surfaces solutions that linear thinking misses.
"I already evaluated this critically, I do not need the framework"
Unstructured evaluation skips dimensions. The framework ensures completeness across all assessment axes.
"Deep questioning slows down the process"
Shallow answers produce shallow solutions. Deep questions reveal constraints the specification missed.
"Converging on the first good idea is efficient"
The first good idea is often a local optimum. Spending 5 minutes on alternatives can find a global optimum.
"This decision is reversible, I do not need rigorous thinking"
Reversible decisions still cost time to reverse. Rigorous thinking upfront prevents the reversal.
Red Flags
Architecture decision documented with only one option considered
Diverge phase produced fewer than 3 alternatives
Converge phase selected an option without evaluation criteria
Critical evaluation missing feasibility or risk dimension
Deep questioning stopped after one level of "why"
Verification
Diverge phase produced at least 3 distinct alternatives
Converge phase used explicit evaluation criteria to select
Critical evaluation covers feasibility, risk, and impact dimensions
Deep questioning reached at least 3 levels for the primary question
Selected approach references the alternatives it was compared against
First Principles (absorbed from moai-foundation-philosopher)
Five-phase strategic analysis framework for architecture decisions and technology selection.
When to Use
Use when facing decisions where the solution space is broad, trade-offs are non-obvious, or existing patterns may not apply. Particularly valuable for: architecture decisions, technology selection, breaking changes, performance vs maintainability trade-offs.
Five-Phase Process
Phase 1 — Assumption Audit: List every assumption underlying the current approach. Challenge each: is it empirically verified or conventionally accepted? Which assumptions carry the highest risk if wrong?
Phase 2 — First Principles Decomposition: Strip away analogies and precedent. Break the problem into fundamental truths. Ask "What must be true for this to work?" rather than "What has worked before?"
Phase 3 — Alternative Generation: From first principles, generate alternatives that ignore prior constraints. Aim for 5-10 distinct approaches before filtering. Include "absurd" alternatives — they often surface overlooked fundamentals.
Phase 4 — Trade-off Analysis: Score each alternative across: correctness, performance, maintainability, operability, reversibility. Use explicit weights based on project context. Document which trade-offs are acceptable vs dealbreakers.
Phase 5 — Cognitive Bias Check: Before finalizing, audit for: anchoring bias (overweighting the first option), confirmation bias (seeking evidence for the preferred option), sunk cost fallacy (keeping a bad choice due to investment).
Adaptive Thinking (Opus 4.7+ built-in, including Opus 5 and 4.8)
The model's built-in reasoning mode (Opus 4.7 and later, including Opus 5 and 4.8) that dynamically allocates reasoning tokens based on task complexity.
Reasoning Modes
Mode
Trigger
Mechanism
GLM Compatible?
ultrathink
Keyword
Claude native extended reasoning
YES
Adaptive Thinking
Automatic
Opus 4.7+ built-in (incl. Opus 5 and 4.8)
YES
Rules: ultrathink → ALWAYS use Claude native extended reasoning with xhigh effort.
When to Activate
Architecture decisions affecting 3+ files, technology selection between options, breaking changes under consideration, performance vs maintainability trade-offs, complex problems with multiple viable approaches.
On Opus 4.7+: Adaptive Thinking handles reasoning depth automatically — no fixed budget needed. The orchestrator activates deep reasoning by prepending the ultrathink keyword to the user prompt.
Steering thinking down (optional)
A large system prompt (loading CLAUDE.md plus many rule files) can cause the model to think more than a routine turn warrants. To suppress over-triggered reasoning on simple turns, you may add a steer such as: "Thinking adds latency and should only be used when it will meaningfully improve answer quality — typically for problems that require multi-step reasoning. When in doubt, respond directly." The ultrathink keyword still overrides this to force deeper reasoning when needed.
Backward Compatibility Marker
Sequential Thinking MCP support retired in an earlier deep-reasoning consolidation. For deep reasoning use the ultrathink keyword (triggers Opus 4.7+ Adaptive Thinking). The earlier "Sequential Thinking MCP (absorbed from moai-workflow-thinking)" content is superseded; this skill's deep-reasoning path is consolidated onto ultrathink + the creative frameworks above.
