modern-prompting

Chain-of-thought prompting improves LLM accuracy by enforcing explicit reasoning steps before generating a final answer. This framework structures the model's intermediate logic, reducing errors in multi-step problems and complex decision workflows.

The ReAct framework combines reasoning with acting to let LLMs interact with external tools and environments. Use ReAct prompting when your workflow requires the model to gather information, execute actions, and adapt decisions dynamically within complex task sequences.

Tree-of-thought prompting structures LLM exploration by branching into multiple reasoning paths before final conclusions. It evaluates different intermediate states, allowing the model to backtrack and select the most viable logical route for improved decision-making.

Program-aided language prompting generates and runs executable code to handle deterministic computations within LLM workflows. By offloading mathematical or logical operations to code execution, it ensures precise outputs where standard text generation fails.

The Reflexion framework implements a self-improvement loop by having the LLM analyze its previous mistakes and refine its approach iteratively. This structured metacognition enhances goal clarification and planning across subsequent task attempts.

Avoid complex prompting strategies like OOReDAct for simple, direct queries where deep deliberation adds unnecessary latency. These frameworks are designed for complex planning and execution workflows, so basic tasks perform better with standard prompting.