reasoning-lora

Structured reasoning via sequential-thinking MCP. Forces explicit step-by-step analysis before coding for complex problems. Reduces rework on multi-step tasks.

SKILL.md

GitHub

Sequential Thinking — Structured Reasoning

先想清楚再动手。复杂问题不跳步,每个 thought 一个判断。

Trigger

Any of → mcp__sequential-thinking__sequentialthinking first, then act:

  1. Multi-step task (≥3 independent steps)
  2. Uncertain approach (≥2 viable methods)
  3. High-risk: delete, refactor, architecture, security
  4. Complex diagnosis: unclear error, multiple possible root causes
  5. User says "think", "analyze", "why"

Template

thought: "Problem: [1 sentence]. Option A: [brief]. Option B: [brief]. Constraints: [list]."
thoughtNumber: 1
totalThoughts: 3
nextThoughtNeeded: true
# Adjust totalThoughts up if problem is more complex, down if simpler

Principles

  • One judgment per thought — don't pack multiple decisions
  • Revise: wrong earlier thought → isRevision: true
  • Branch: both directions worth exploring → branchFromThought
  • Verify: final thought validates hypothesis → needsMoreThoughts: false
  • Dynamic: more complex than expected → raise totalThoughts; simpler → lower

Success Story

Without this skill: "Refactor the auth system" → jumped straight to code → 4 hours later realized wrong approach → rewrote everything. With this skill: "Refactor the auth system" → 5 sequential thoughts → identified JWT-vs-session tradeoff → picked correct approach → 1 hour, first try.

Integration

  • memory-lora: Patterns found → stored for cross-session recall
  • delegation-lora: "Need to search" → Explore agent. "Need review" → second-brain
  • second-brain: Major reasoning decisions → Brain 2 validates approach