Techniques and patterns for writing prompts that get reliable results from language and image models.
55 terms
Active prompting selects the questions a language model is most uncertain about, has humans annotate them, and uses them as chain-of-thought examples.
Analogical prompting asks a language model to recall similar solved problems on its own before it solves a new one, replacing hand-written examples.
Auto-CoT, or automatic chain-of-thought, builds few-shot reasoning demonstrations automatically by clustering questions and generating rationales.
Chain-of-density prompting makes a language model rewrite a summary several times at fixed length, adding missing key entities to raise information density.
Chain-of-verification (CoVe) is a prompting method in which a language model drafts an answer, checks its own claims with questions and then revises it.
A cloze prompt is a fill-in-the-blank prompt that frames a task as text with a missing span, so a language model answers by predicting the gap.
The CO-STAR framework structures a prompt into Context, Objective, Style, Tone, Audience and Response so a language model returns focused, usable output.
Context engineering is the practice of selecting, structuring and budgeting everything a language model sees in its context window at inference time.
Custom instructions are persistent user or organization preferences that an AI assistant adds to every conversation to shape tone, format and behavior.
Decomposed prompting (DecomP) splits a complex task into sub-tasks handled by separate prompts or tools, coordinated by a decomposer prompt.
Directional stimulus prompting uses a small trained model to add input-specific hints, such as keywords, that steer a frozen large language model.
Dynamic few-shot prompting selects the most relevant worked examples for each input at run time and inserts them into a language model prompt.
Emotion prompting adds emotionally charged phrases such as stakes or encouragement to a prompt to try to improve the effort and accuracy of a language model.
The flipped interaction pattern is a prompting technique in which a language model asks the user questions until it has enough information to reach a goal.
Generated knowledge prompting asks a language model to write relevant facts first and then answers the question using those facts as added context.
Graph of thoughts is a prompting framework that organizes language model reasoning steps as a graph, letting partial ideas be merged, refined and scored.
Iterative prompting is the practice of refining a prompt over repeated rounds of testing and revision until a language model produces reliable output.
Least-to-most prompting is a technique that has a language model split a hard problem into simpler subproblems and solve them in order, reusing each answer.
Many-shot prompting places hundreds or thousands of worked examples in a long context window so a language model learns a task without fine-tuning.
Meta-prompting is a prompting technique in which a language model writes, refines or orchestrates prompts for itself or for another model.
A negative prompt is a text input that tells a generative model, usually an image diffusion model, which content, styles or defects to steer away from.
One-shot prompting gives a language model exactly one worked example of a task inside the prompt so it copies the format and intent.
An output format instruction is the part of a prompt that tells a language model exactly how to structure its answer, such as JSON, a table…
Plan-and-solve prompting asks a language model to devise a plan for a problem first and then carry it out step by step, reducing skipped reasoning steps.
Program of thoughts (PoT) is a prompting technique in which a language model writes executable code and an interpreter computes the final answer.
A program-aided language model (PAL) writes a short program for a reasoning problem and lets an interpreter run it to compute the final answer.
Prompt compression shortens the input sent to a language model by removing or encoding low-value tokens, cutting cost and latency while keeping answer quality.
A prompt delimiter is a marker such as an XML tag, triple quotes or a heading that separates instructions, data and examples inside a language model…
Prompt drift is the gradual change in what a language model returns for the same prompt as models, prompt edits or inputs shift over time.
Prompt ensembling sends several differently worded prompts to a language model and aggregates the answers to gain accuracy and stability.
Prompt evaluation is the systematic testing of prompts against fixed test cases and metrics to measure how reliably a language model produces correct output.
Prompt expansion rewrites a short user request into a longer, more detailed prompt before a model sees it, improving output quality and consistency.
A prompt framework is a named structure that lists the components a prompt should contain, such as role, task, context and format, for reliable model output.
A prompt library is a curated, versioned collection of tested prompts that teams reuse to get consistent results from language and image models.
Prompt management is the practice of storing, versioning, testing and deploying prompts for language model applications as controlled, reviewable assets.
A prompt marketplace is an online platform where creators sell, license or share tested prompts for language and image models to buyers who reuse them.
Prompt optimization is the systematic search for prompt wording and examples that maximize a language model's measured performance on a defined task.
A prompt pattern is a reusable, named solution to a recurring problem in prompting, describing intent, structure and consequences for language models.
A prompt playground is an interactive interface for drafting, running and comparing prompts against language models before they go into production.
Prompt sensitivity is the tendency of language models to change their outputs substantially when a prompt is reworded, reordered or reformatted.
A prompt template is a reusable prompt with fixed instructions and named variables that software fills in before sending it to a language model.
Prompt weighting assigns numeric importance to parts of a prompt so an image model emphasizes or suppresses specific concepts in the generated output.
Rephrase and respond (RaR) is a prompting technique in which a language model restates and clarifies a question before answering it, reducing misreadings.
Response prefilling is a prompting technique that writes the first words of the model's answer so it continues in a fixed format, voice or structure.
Reverse prompt engineering infers a likely prompt from an existing AI output, such as an image or text, so the result can be studied or reproduced.
Role prompting assigns a language model a persona, such as an editor or tutor, to steer its tone, vocabulary, focus and level of detail in its…
Self-ask prompting makes a language model pose and answer its own follow-up questions before giving a final answer to a multi-hop question.
Self-consistency is a prompting technique that samples several reasoning paths from a language model and selects the final answer by majority vote.
Self-refine is a prompting technique in which a language model critiques its own draft and then rewrites it, repeating the loop until the output is good…
Skeleton-of-thought is a prompting technique in which a language model first writes a short outline and then expands each point in parallel to cut latency.