thought-based-reasoning_skill

Overview

This skill is a practical guide to thought-based reasoning techniques for large language models. It collects Chain-of-Thought and related prompting patterns (Zero-shot CoT, Self-Consistency, Tree of Thoughts, Least-to-Most, ReAct, PAL, Reflexion) with templates, decision matrices, and implementation notes to improve accuracy on complex reasoning tasks. It targets engineers and prompt designers who need reliable, research-backed strategies to handle multi-step logic, math, and exploration problems.

How this skill works

The skill inspects the problem type and recommends one or more reasoning strategies, providing prompt templates and procedural steps for generation, sampling, evaluation, and extraction. It also explains orchestration patterns: sampling multiple chains (self-consistency), structured search and backtracking (Tree of Thoughts), decomposition and sequential solving (Least-to-Most), and code-backed solutions (PAL). For tool-enabled agents it outlines ReAct-style action traces and evaluation loops.

When to use it

• Multi-step arithmetic, symbolic or logic problems where plain prompts fail
• High-stakes tasks needing confidence via multiple reasoning paths
• Problems that benefit from decomposition or sequential subproblem solving
• Exploratory or creative tasks requiring backtracking and search
• When external tools or retrieval must be integrated into reasoning

Best practices

• Match technique to task: PAL for precise computation, ToT for hard search tasks, ReAct when tools are required
• Use staged workflows: separate reasoning extraction from final answer extraction to reduce hallucination
• Apply Self-Consistency with diverse samples and majority voting for reliability
• Keep prompts modular: small templates for thought generation, evaluation, and actions
• Monitor compute vs. benefit: ToT and Self-Consistency increase cost—tune beam/shot counts

Example use cases

• Solving multi-step word math problems with PAL or CoT plus code execution
• Multi-hop question answering that needs web search using ReAct traces
• Complex puzzle solving or planning using Tree of Thoughts with beam search
• Compositional reasoning tasks using Least-to-Most decomposition and sequential solves
• Improving model agreement on high-risk decisions via Self-Consistency sampling