sequential-thinking（顺序思维）

Sequential Thinking Methodology

Framework for using /mcp__sequential-thinking__sequentialthinking tool effectively.

Quick Assessment

Skip entirely for: Simple facts, obvious fixes, single-step answers, routine tasks

Apply when:

• Problem requires 3+ logical steps
• Multiple valid approaches exist
• Debugging with unclear root causes
• Initial confidence <70%
• State transitions or causal chains involved

Core Framework

1. Initial Classification

Simple → Skip tool
Medium → 3-4 thoughts
Complex → up to 7 thoughts

Define: initial state → goal state → success criteria

2. State-Based Reasoning

For each thought:

1. Current state: What's true now?
2. Preconditions: What must be true for next action?
3. Effects: How does action change state?
4. Verify: Valid and closer to goal?
5. Check: Success criteria met?

3. Thought Structure

Thought N: [Hypothesis/Analysis/Action]
- State: [explicit]
- Logic: [deduction|induction|abduction|causal]
- Confidence: X%
- Next: [what this enables]

4. Decision Points

Branch (branch_from_thought + unique branch_id):

• Confidence <60% WITH alternative
• Contradiction found
• Equal validity paths (Δ<10%)

Revise (is_revision=true + revises_thought):

• Fundamental error
• Precondition false
• Better approach (>20% gain)

Extend (needs_more_thoughts=true):

• Approaching limit but incomplete

5. Quality Scoring

Start: 0 | Minimum: 5 points

Positive: Clear mechanism (+2), Verified assumption (+2), Resolved contradiction (+2), Valid transition (+1)

Negative: Circular reasoning (-3), Invalid transition (-3), Unverified assumption (-2), Logical leap (-1)

If <5 after 3 thoughts → reassess approach

6. Convergence Rules

• Thoughts 1-2: Explore breadth (20%)
• Thoughts 3-5: Deep dive (60%)
• Thoughts 6-7: Verify only (20%)

Terminate when: 7 thoughts reached | Confidence >90% for 2 consecutive | Cycling detected | Success met

Code-Specific Additions

Track: Variable states, Pre/postconditions, Resource constraints, Error paths

Verify: Execution traces, Edge cases, Resource cleanup, Thread safety

Natural Language Integration

Use markers like:

• "Let me work through this systematically..."
• "This leads me to consider..."
• "A potential issue here is..."

Never expose: Internal scores, Framework mechanics, Artificial confidence percentages

Meta-Rules

• Admit uncertainty rather than guess
• Each thought must advance solution
• No pure repetition
• Don't fake confidence to avoid branching