24 techniques
Each page explains one technique in plain terms, then shows what the studies on it found: where it helps, where it hurts, and prompts you can copy.
Chain-of-thought prompting makes the model write intermediate reasoning before its answer, either freely or through a fixed template. It reliably helps on checkable multi-step problems, works best when the steps are structured, and can hurt summaries, social judgement, code and constraint-following.
39 studiesLLM-as-a-judge uses a model to score, rank or compare outputs against criteria. It approaches human agreement when the prompt asks a pairwise or rubric question with discriminating criteria, puts the reasoning before the verdict and averages several ratings; a bare 'rate this 1-10' mostly measures how persuasive the text sounds.
38 studiesSelf-refine is a loop in which the model writes a draft, critiques it, and revises it. It works when the critique is specific (a checklist, a named error, an external check, a fresh context) and usually stalls after one or two rounds; open-ended 'improve this' often breaks answers that were already correct.
38 studiesMulti-agent debate means having several roles, in one prompt or in separate chats, argue, attack and defend an answer before a verdict. It helps when the roles genuinely differ and must back objections with evidence; several studies show it also drops facts, drifts into empty consensus, or loses to one well-written prompt.
30 studiesTask decomposition splits one hard request into narrow steps (a fixed plan, collect-structure-analyze, extract-then-apply-rules, atomic yes/no checks) so the model carries one load at a time. It produced large gains on multi-step tasks, but chains lose accuracy at each handoff unless the original input travels with the intermediate results.
26 studiesA verification prompt runs a separate, structured check (atomic claims, a contradiction gate, an independent re-solve, two reasoning modes that must agree) instead of a vague 'check your answer'. Structured checks caught far more errors in the studies here, while generic re-check rounds changed correct answers into wrong ones.
25 studiesRole prompting assigns the model an identity, values or a detailed persona. Studies find expert roles rarely improve factual accuracy, while specific, action-based roles and personas anchored in real history reliably steer viewpoint, tone and simulated behaviour.
24 studiesAgent planning is the prompt-level structure for multi-step, tool-using models: an explicit plan or dependency graph, a state record phrased as commands, separate planner and executor contexts, and concrete boundaries. It works because agents otherwise drift, report false progress and take shortcuts, but piling process instructions onto the model can hurt.
22 studiesSelf-consistency means getting several independent answers to the same question and aggregating them, by majority vote, by a judge that picks one, or by a synthesis that merges them. It works when the answers fail in different ways; it fails when every sample repeats the same mistake.
21 studiesFew-shot prompting means showing the model a few worked examples so it copies the pattern. Studies find 5-25 examples placed before the task work best, that examples shape form more than logic, and that explicit rules can beat them on compound conditions.
19 studiesGrounded prompting makes the model answer from supplied material instead of memory: frame the source explicitly, anchor the output to quotes or line numbers, and check which required facts the evidence actually covers. Framing alone moved document-following by up to 94 points in one study, but repeated, cited or search-retrieved sources can also mislead the model.
19 studiesMeta-prompting uses the model to critique, expand or rewrite a prompt, or to distil rules from examples and failures. It works when the model gets a reference point (named criteria, good/bad pairs, logged errors, a test set); a bare "improve my prompt" mostly produces guesses.
19 studiesAbstention prompting gives the model an explicit, accepted way out (a rewarded "I don't know", a CLARIFY or REQUEST_SUPPORT action, a FAILED status, a "no changes needed" verdict) instead of forcing an answer. It works because models already sense uncertainty but do not act on it unless declining counts as success.
17 studiesClarifying-question prompting tells the model to surface ambiguity before it acts: ask one question, list its assumptions, or stop. Without it, models quietly pick one reading of a vague task and answer that one confidently.
16 studiesPrompt compression cuts tokens on the way in (denser notation, extracted facts, pruned rules) or on the way out (brevity instructions, answer formats, bounded reasoning). The studies agree on one asymmetry: compressing the answer saves money, while compressing a complete request often makes the answer longer and worse.
16 studiesStructured output means asking the model for machine-readable answers such as JSON, tables or fixed templates. The most consistent finding is that format instructions compete with the task itself, so reasoning (or recall) and formatting work better as separate steps.
16 studiesNegative constraints tell the model what must not appear in the output or what it must not do. Concrete, checkable bans on words and actions tend to work; bans on thinking, such as "ignore X" or "don't flatter", fade or backfire.
15 studiesDiversity prompting changes what you ask for, such as a distribution of answers, explicit directions, contrasting personas or a distant analogy, so that the most typical answer stops dominating. It is for open-ended tasks with many acceptable answers.
15 studiesAdaptive reasoning is deciding per task whether the model should answer directly, reason briefly or reason at length. Studies show reasoning pays off on verifiable multi-step problems and wastes tokens or lowers accuracy on recall, summaries, social judgement and open-ended decisions.
14 studiesInstruction placement is deciding where role, priorities, material, examples and the question sit in a prompt. Research supports a simple layout: critical priorities first, reference material in the middle, the specific task and options last, and the task repeated at the end when the context is long.
14 studiesInput formatting is how you lay out a prompt: which part is the instruction, which part is data, and how the data itself is structured. Studies show the same content can give different accuracy depending on separators, tags and table format.
11 studiesCode prompting writes the instructions as pseudocode, rules or formulas, or has the model write a program that computes the answer. It helps most on branching workflows, structured data and heavy calculation, and can lose to plain reasoning on simple problems.
9 studiesCross-lingual prompting sets the language of each part of a prompt separately: instructions, source data, reasoning and reply. The evidence favours keeping data in the original, adding English anchors for key terms, and always stating the reply language.
8 studiesAnalogical prompting has the model work through a related problem first: one it invents on the spot, a past case matched by structure, or a twin from a distant field. Close analogies improved accuracy on reasoning tasks; distant ones broke the model out of stock answers.
6 studiesStarting from a problem instead? Browse LLM problems →