Explicit step-by-step prompting can improve some arithmetic, commonsense, and symbolic reasoning tasks. The effect depends on the model and task, so compare it with a concise prompt on labeled examples.

Forces entity recall while separating thinking from writing. Aligns with how executives read.

Separates outline approval from drafting so structure can be checked before detail is generated.

Uses an external result, rubric finding, or human comment to guide one bounded critique and revision pass.

Draft an answer, plan targeted verification questions, answer those questions without relying on the draft, then revise the final answer.

Separates language reasoning from deterministic computation by having the model generate executable code and use the program result in its answer.

Each step has one job. Easier to debug and more repeatable across users and runs.

Define output schema upfront. Model extracts only what matches, reducing hallucination and enabling validation.

Interleaves reasoning and action. Model thinks, acts, observes result, then thinks again. Reduces hallucination by grounding in external observations.

Provide class definitions and optional examples. Model maps input to most appropriate class. Few-shot examples dramatically improve edge case handling.

Transform clustering into classification. First prompt LLM to generate potential labels, then assign items to labels. Combines semantic understanding with clustering.

Use LLM as a relevance judge after initial retrieval. Listwise ranking outperforms pointwise. Two-stage pipeline: fast retriever (100 docs) → LLM reranker (top 10).

Generate hypothetical answer document, then use its embedding for retrieval. Generate 5 docs and average embeddings. Works best when LLM has domain knowledge.

Samples multiple reasoning paths and selects the most frequent final answer. Temperature and sample count are evaluation controls, not universal settings.