A model-agnostic reference for the prompting techniques that actually move the needle. Use it to structure prompts, control output, and reason through hard tasks.
Zero-shotAsk directly with no examples. Best for simple, well-known tasks.Few-shotInclude 2-5 input/output examples to show the desired pattern and format.Chain-of-thoughtAdd "Think step by step" to force intermediate reasoning before the answer.Role promptingAssign a persona: "You are a senior tax accountant." Sets tone and expertise.DecompositionBreak a complex task into ordered sub-tasks the model solves one at a time.Self-consistencyGenerate multiple reasoning paths, then take the majority answer.ReActInterleave Reasoning and Acting (tool calls) for agentic, multi-step tasks.DelimitersWrap inputs in triple quotes, ###, or XML tags so the model knows what is data.Instructions firstPut the task at the top, supporting context below. Models weight the start heavily.Be specificReplace "make it good" with measurable criteria (length, tone, audience).Positive framingSay what to do, not just what to avoid. "Use plain English" beats "Don't be technical."One ask per promptBundling unrelated requests degrades quality. Split into separate turns.Output formatSpecify exactly: "Return JSON with keys title, body, tags."Length controlSet hard limits: "Max 3 sentences" or "Exactly 5 bullet points."Tone & styleName the register: formal, casual, witty, academic, or "like Hemingway."Schema / templateProvide a fill-in-the-blank template the model must complete verbatim.Refusal handlingAdd "If you are unsure, say so" to reduce confident hallucinations.Step-back promptingAsk for the general principle first, then apply it to the specific case.Tree-of-thoughtExplore several branches, evaluate each, then expand the most promising.Plan then executeHave the model draft a plan, confirm it, then carry it out.Critique & reviseAsk the model to find flaws in its own answer, then produce a v2.Rubric scoringGive a scoring rubric so the model self-evaluates against your criteria.GroundingProvide source text and instruct: "Answer only from the context above."Cite sourcesRequire quotes or references for every factual claim."I don't know" licenseExplicitly allow the model to admit uncertainty instead of guessing.TemperatureLower it (0-0.3) for factual tasks, raise it (0.7-1.0) for creative ones.Verify with a second passRun a separate prompt that fact-checks the first output.Vague verbs"Analyze" and "optimize" mean nothing without criteria. Define them.Buried instructionsKey rules lost in long context get ignored. Repeat them at the end.Overloading contextMore text is not better. Irrelevant info dilutes attention.Leading examplesBiased few-shot examples make the model copy your mistakes.No success criteriaIf you can't tell good from bad output, neither can the model.Want ready-to-use prompts to go with this cheatsheet?
Browse 30,000+ Prompts