Prompting Strategies

Interleave reasoning traces with task-specific actions.

Decompose questions by asking and answering follow-up questions.

Connect multiple focused prompts in sequential workflows.

Assign an expert persona to access specialized knowledge.

Add emotional context to enhance AI response quality and engagement.

Control writing style, tone, and voice characteristics.

Let AI ask clarifying questions before responding.

Regenerate context to remove irrelevant information before answering.

Simulate different perspectives for theory-of-mind reasoning.

Rephrase the question before answering to improve comprehension.

Re-read the question to enhance understanding of complex problems.

Define AI behavior, personality, and constraints through persistent system-level instructions.

Ground AI responses in retrieved documents to reduce hallucination and deliver source-backed answers.

Design AI systems that autonomously plan, execute, and adapt tasks using tools and reasoning loops.

High-level structural templates that teach HOW to approach tasks rather than providing specific content.

Process multiple task instances in a single prompt for efficiency and cost reduction.

Simply repeating key instructions improves non-reasoning LLM performance through emphasis.

Use targeted hints and cues to steer LLM outputs toward desired directions.

Model generates relevant knowledge before answering to improve accuracy on reasoning tasks.

Generate hypothetical answer documents to improve retrieval without task-specific training.

Transform task inputs into question-answer format and aggregate multiple reformulated responses.

Use reflective Socratic questioning to guide AI through iterative reasoning and discovery.

Use multi-turn dialogue structure to guide reasoning through conversational exchange.

Infer task instructions from a few input-output demonstrations automatically.

Automatically generate and select optimal prompt instructions via LLM-driven search.

Use LLMs as optimizers where the optimization task is described in natural language.

Generate multiple semantically similar prompt variants and ensemble their responses.

Treat reasoning as planning using the LLM as both world model and agent with MCTS.

Synergize reasoning, acting, and planning via Monte Carlo Tree Search for agents.

Iteratively refine summaries to be progressively more entity-dense without increasing length.

Optimize prompts by reflecting on runs in language and evolving along a Pareto frontier.

Improve prompts from pairwise comparisons of the model's own outputs, no labels needed.

Cost-aware evolutionary optimization with racing and a length penalty for cheaper prompts.

Optimize prompts by token-level loss, locating weak spans by masking, no sampling or judge.

Evolve prompts through structured debates scored by Elo, optimizing subjective tasks.

Treat prompt selection as a dueling bandit over LLM-judge pairwise feedback, label-free.

AutoML search over prompting patterns and demos via successive halving, emits editable PDL.

Treat an agent's context as an evolving playbook refined by generate, reflect, and curate.

Separate a Planner writing high-level plans from an Executor carrying them out.

Recursive planning that reinserts the parent plan so agents stay anchored to the goal.

Curate the smallest high-signal set of tokens across an agent's whole loop.

Design tool and function definitions agents use well, from namespacing to lean responses.

Three-tier multi-agent orchestration: plan, design roles, execute, with tree search.

Structured step-by-step templates for generating correct, interpretable tool calls.

Agents share a workspace; whoever can contribute next is chosen from the board state.

A meta-team that generates, configures, and rectifies a target multi-agent system per task.

Backpropagate natural-language feedback through an AI system to optimize its parts.

Optimize multi-turn agentic systems in-the-flow with group refined policy optimization.

Program language models with typed signatures and automatic compilation instead of hand-crafted prompts.

Optimize multi-stage LM programs with trace-grounded instruction proposals and Bayesian search.

Step-by-step templates for clear, repeatable prompt structure.

Teach AI by demonstration with zero-shot, few-shot, and example selection.

Chain-of-thought methods for transparent, step-by-step reasoning.