reasoning

Mandatory

MANDATORY: Use this framework for ANY non-trivial analysis or decision.

CRITICAL: Complete reasoning BEFORE acting - once you provide a solution, it's non-reversible.

NO EXCEPTIONS: Skipping structured reasoning = poor decisions = wasted time.

Subagent Context

IF YOU ARE A SUBAGENT: Follow this reasoning process directly and return your analysis to the primary agent.

Reasoning Framework

Before ANY Action

Complete this reasoning internally BEFORE responding:

1. Dependency Analysis (Priority Order)

Analyze in this order - higher priority overrides lower:

Rules/Constraints - Highest priority

• Policy-based rules, mandatory prerequisites
• NEVER violate for convenience

Operation Order

• Ensure actions don't block subsequent steps
• Reorder operations if user presents them randomly

Prerequisites

• What information/actions are needed first?
• Only ask clarification when missing info would SIGNIFICANTLY affect solution

User Preferences

• Satisfy within above constraints
• Style, language, approach preferences

2. Task Classification

Classify before proceeding:

trivial (respond directly):

• Simple syntax issues, single API usage
• Local modifications < 10 lines
• One-line fixes obvious at a glance

moderate (use planning skill):

• Non-trivial logic within single file
• Local refactoring
• Simple performance/resource issues

complex (full planning + brainstorming):

• Cross-module or cross-service design
• Concurrency and consistency
• Complex debugging, multi-step migrations

3. Risk Assessment

Low-risk (exploratory):

• Proceed with available information
• Don't ask for perfect info on optional parameters
• Missing optional info is acceptable

High-risk (destructive/irreversible):

• Clearly explain risks BEFORE proceeding
• Provide safer alternatives if possible
• Get explicit confirmation for:
  • Data deletion or modification
  • History rewriting (git reset, rebase)
  • Public API changes
  • Database structure changes

4. Hypothesis Formation (for problems)

When encountering issues:

1. Form 1-3 hypotheses sorted by likelihood
2. Look beyond obvious causes - most likely may require deeper inference
3. Don't discard low-probability hypotheses prematurely
4. Test most likely first
5. If disproven, generate NEW hypotheses from gathered info
6. After 3 failed hypotheses, question architecture (see debugging skill)

5. Action Inhibition

CRITICAL:

• Complete ALL reasoning BEFORE acting
• Once you provide solution/code, it's non-reversible
• Don't hastily provide answers before reasoning is complete
• If errors discovered later, correct in new replies

Precision and Grounding

• Keep reasoning SPECIFIC to current situation, not generic
• When making decisions based on constraints/rules, briefly explain WHICH constraints apply
• Quote exact applicable information when referring to policies
• Verify claims against actual code/docs, don't assume

Completeness Check

Before finalizing:

• Are ALL explicit requirements considered?
• Are main AND alternative paths covered?
• Are conflicts resolved using priority order?
• Have you checked all information sources?

Persistence and Intelligent Retry

• Don't give up easily - try different approaches within reason
• On transient errors ("please try again"): retry with adjusted timing
• On other errors: CHANGE strategy, don't repeat same failed approach
• If explicit retry limit reached: stop and explain

Anti-Rationalization

THESE EXCUSES NEVER APPLY

"This is simple, don't need structured reasoning" WRONG: Even simple decisions benefit from quick analysis

"I'll just try something and see" WRONG: Reason FIRST, act SECOND

"User is in a hurry, skip reasoning" WRONG: Systematic is FASTER than thrashing

"I already know the answer" WRONG: Verify against constraints anyway

NO EXCEPTIONS

Compliance Checklist

MANDATORY CHECKLIST:

☐ Analyzed dependencies in priority order

☐ Classified task complexity (trivial/moderate/complex)

☐ Assessed risk level (low/high)

☐ Formed hypotheses if problem-solving

☐ Completed reasoning BEFORE acting

☐ Verified against constraints/rules

IF ANY UNCHECKED, REASONING IS INCOMPLETE