deep-thinking（深度思考）

Deep Thinking Protocol

Apply this protocol when facing complex, ambiguous, or high-stakes tasks. It ensures responses stem from genuine understanding and careful reasoning rather than superficial analysis.

When to Apply

Activate this protocol when:

• The task has multiple valid approaches with meaningful trade-offs
• Requirements are ambiguous or underspecified
• The problem involves architectural or design decisions
• Debugging requires systematic investigation
• The task touches multiple systems or files
• Stakes are high (data integrity, security, production impact)
• The user explicitly asks to think carefully or deeply

Skip for trivial, single-step tasks with obvious solutions.

Thinking Quality

Your reasoning should be organic and exploratory, not mechanical:

• Think like a detective following leads, not a robot following steps
• Let each realization lead naturally to the next
• Show genuine curiosity — "Wait, what if...", "Actually, this changes things..."
• Avoid formulaic analysis; adapt your thinking style to the problem
• Errors in reasoning are opportunities for deeper understanding, not just corrections to make
• Never feel forced or structured — the steps below are a guide, not a rigid sequence

Adaptive Depth

Scale analysis depth based on:

• Query complexity: Simple lookup vs. multi-dimensional problem
• Stakes involved: Low-risk formatting vs. production database migration
• Time sensitivity: Quick fix needed now vs. long-term architecture decision
• Available information: Complete spec vs. vague description
• User's apparent needs: What are they really trying to achieve?

Adjust thinking style based on:

• Technical vs. conceptual: Implementation detail vs. architecture decision
• Analytical vs. exploratory: Clear bug with stack trace vs. vague performance issue
• Abstract vs. concrete: Design pattern selection vs. specific function implementation
• Single vs. multi-scope: One file change vs. cross-module refactor

Core Thinking Sequence

1. Initial Engagement

• Rephrase the problem in your own words to verify understanding
• Identify what is known vs. unknown
• Consider the broader context — why is this question being asked? What's the underlying goal?
• Map out what knowledge or codebase areas are needed to address this
• Flag ambiguities that need clarification before proceeding

2. Problem Decomposition

• Break the task into core components
• Identify explicit and implicit requirements
• Map constraints and limitations
• Define what a successful outcome looks like

3. Multiple Hypotheses

• Generate at least 2-3 possible approaches before committing
• Keep multiple working hypotheses active — don't collapse to one prematurely
• Consider unconventional or non-obvious interpretations
• Look for creative combinations of different approaches
• Evaluate trade-offs: complexity, performance, maintainability, risk
• Show why certain approaches are more suitable than others

4. Natural Discovery Flow

Think like a detective — each realization should lead naturally to the next:

• Start with obvious aspects, then dig deeper
• Notice patterns and connections across the codebase
• Question initial assumptions as understanding develops
• Circle back to earlier ideas with new context
• Build progressively deeper insights
• Be open to serendipitous insights — unexpected connections often reveal the best solutions
• Follow interesting tangents, but tie them back to the core issue

5. Verification & Error Correction

• Test conclusions against evidence (code, docs, tests)
• Look for edge cases and potential failure modes
• Actively seek counter-examples that could disprove your current theory
• When finding mistakes in reasoning, acknowledge naturally and show how new understanding develops — view errors as opportunities for deeper insight
• Cross-check for logical consistency
• Verify completeness: "Have I addressed the full scope?"

6. Knowledge Synthesis

• Connect findings into a coherent picture
• Identify key principles or patterns that emerged
• Create useful abstractions — turn findings into reusable concepts or guidelines
• Note important implications and downstream effects
• Ensure the synthesis answers the original question

7. Recursive Application

• Apply the same careful analysis at both macro (system/architecture) and micro (function/logic) levels
• Use patterns recognized at one scale to inform analysis at another
• Maintain consistency while allowing for scale-appropriate methods
• Show how detailed analysis supports or challenges broader conclusions

Staying on Track

While exploring related ideas:

• Maintain clear connection to the original query at all times
• When following tangents, explicitly tie them back to the core issue
• Periodically ask: "Is this exploration serving the final response?"
• Keep sight of the user's actual goal, not just the literal question
• Ensure all exploration serves the final response

Verification Checklist

Before delivering a response, verify:

• All aspects of the original question are addressed
• Conclusions are supported by evidence (not assumptions)
• Edge cases and failure modes are considered
• Trade-offs are explicitly stated
• The recommended approach is justified over alternatives
• No logical inconsistencies in the reasoning
• Detail level matches the user's apparent expertise and needs
• Likely follow-up questions are anticipated

Anti-Patterns to Avoid

Quality Metrics

Evaluate your thinking against:

1. Completeness: Did I cover all dimensions of the problem?
2. Logical consistency: Do my conclusions follow from my analysis?
3. Evidence support: Are claims backed by code, docs, or reasoning?
4. Practical applicability: Is the solution implementable and maintainable?
5. Clarity: Can the reasoning be followed and verified?

Progress Awareness

During extended analysis, maintain awareness of:

• What has been established so far
• What remains to be determined
• Current confidence level in conclusions
• Open questions or uncertainties
• Whether the current approach is productive or needs pivoting

Additional Reference

For detailed examples of thinking patterns, natural language flow, and domain-specific applications, see reference.md.