Prompting

Taxonomy of Prompting Philosophies and Strategies for AI Text Systems

Prompts for large language models (LLMs) have evolved into a variety of distinct philosophies, strategies, and frameworks. Below is a structured taxonomy grouping these prompting approaches by core type – such as persona-driven, structure-based, goal-driven, cognitive-emulation, etc. Each approach includes its origin or key proponents, a summary of its principles, and an indication of how productive it is considered (based on research results or community feedback).

Approach Name Type Source/Origin Core Principles Productivity Rating
Role/Persona Prompting Persona-driven Research & Community Assign a clear role or persona to the AI (e.g. “You are a Senior Editor…”) to shape style/expertise. Proven to yield more contextually tailored and sometimes more accurate answers. High – Widely improves relevance and tone control.
Style or Tone Specification Persona-driven Community & Consultants Explicitly instruct the model’s style, tone or genre (e.g. “Respond in a neutral academic tone” or “in the style of [X]”). Aligns output with desired voice, avoiding generic “average” responses. Often used with role prompts. Medium – Ensures output meets voice requirements (quality depends on task).
Emotion/Stakes Prompting Persona-driven Research (Li et al. 2023); Community Include emotional context or importance to the user (e.g. “This is critical for my job”). Research suggests this can sharpen model focus and improve performance by mimicking human emotional salience. Medium – Can increase engagement/priority for certain answers.
Structured Prompt Templates Structure-based Community (Blogs/Consultants) Use a formatted template with sections (e.g. frameworks like RACE, CRISPE, etc.) to structure context and instructions. For example: Role, Task, Context, Outcome. This clarity prevents ambiguity and ensures all relevant info is provided. High – Strongly improves clarity and completeness of responses.
Delimiter & Sectioning Structure-based OpenAI & Community Guides Separate different prompt parts or data with delimiters or markdown (e.g. headings, triple quotes """ for context). This helps the model distinguish instructions from content, yielding ~31% better comprehension in one test. High – Noted to reduce confusion and misinterpretation.
Output Format Instructions Structure-based OpenAI guides Explicitly tell the model how to format the answer (lists, tables, JSON, etc.). For example, “Provide the answer as a table with columns X and Y.” This guides the structure of output, making it more usable. High – Increases usefulness by tailoring output form to needs.
Few-Shot Example Prompting Demonstration-based OpenAI (Brown et al. 2020) Provide examples of queries and ideal answers in the prompt. The model infers the pattern and imitates it. Can be one-shot (single example) or few-shot (multiple examples). Greatly improves performance by “showing” the task, not just telling. High – Often dramatically improves accuracy (e.g. 58% higher success in one trial).
Zero-Shot Prompting Demonstration-based N/A (Baseline) No examples given – just an instruction or question. Reliant on model’s training to infer intent. Simpler but often less reliable for complex tasks. Often a baseline to compare against enhanced strategies. Low – Can produce generic or off-target results unless task is simple.
Chain-of-Thought Prompting Cognitive emulation Research (Wei et al. 2022) Prompt the model to show reasoning steps before the final answer. Originally done with few-shot reasoning examples; later also zero-shot (“Let’s think step by step”). Greatly enhances performance on math, logic, and complex questions by encouraging stepwise thinking. High – Significantly boosts complex reasoning accuracy.
Zero-Shot CoT (“Step by Step”) Cognitive emulation Community & Research A special case of CoT: simply appending a trigger phrase like “Let’s think step by step” to an instruction. Triggers the model’s latent reasoning ability without providing examples. Very effective for arithmetic, logic, etc. (Not needed for trivial tasks). High – Quick technique to improve reasoning; empirically very effective.
Self-Ask (Socratic Q&A) Cognitive emulation Research (Press et al. 2022) The model first generates clarifying questions (to itself) about the problem, answers them, then produces a final answer. This breaks complex queries into simpler sub-questions. It mimics a Socratic approach to ensure all aspects are covered. Medium – Improves complex Q understanding, though longer outputs.
ReAct (Reason+Act) Cognitive emulation Research (Yao et al. 2022) A multi-turn prompt strategy combining reasoning steps with actions (e.g. tool use). The model alternates between thinking (Chain-of-Thought) and acting (e.g. executing a search). Enables the AI to gather information or interact with an environment before final answers. Medium – Powerful for tool-integrated systems; less relevant for static Q&A.
Problem Decomposition Cognitive emulation Research (Yao et al. 2023; Zhou et al. 2023) Break a complex task into sub-tasks and solve them in sequence. Variants include Least-to-Most (start with simple parts) and Plan-and-Solve (outline a plan, then execute). Prevents the model from getting “tunnel vision” by tackling pieces independently. High – Noted as “revolutionary” for complex problem solving.
Chained Multi-step Prompts Cognitive emulation Community (OpenAI, Blogs) Also called iterative prompting: ask for an intermediate result (outline, ideas) first, then feed it into a follow-up prompt. For example, “First give an outline, then write full content.” This catches errors early and refines the request in stages. High – Produces more focused, well-structured results and saves time on revisions.
Multi-Perspective Simulation Cognitive emulation Community (Reddit) Instruct the AI to simulate multiple viewpoints or roles debating or analyzing a topic, then synthesize their conclusions. E.g. have 4 different experts with distinct perspectives discuss a problem. This yields a nuanced answer covering diverse angles. Medium – Extremely comprehensive outputs, though lengthier; great for complex decisions.
Self-Critique & Refinement Iterative/reflexive Research (Madaan et al. 2023); Community Have the model evaluate and improve its own draft iteratively. One method: Recursive Self-Improvement Prompting (RSIP) – generate an answer, then list flaws, then redo it, repeating for a few iterations. Similarly, the Self-Refine framework has the model feed its answer back into itself for critique. High – Yields significantly polished outputs; reduces errors and enhances quality across reasoning and writing tasks.
Confidence-Based Prompting Iterative/reflexive Community (Reddit) Ask the AI to label its statements with confidence levels and justify them. For example, instruct it to tag each claim as “Certain, Likely, Speculative, or Unknown” with reasoning. This forces the model to reflect on uncertainty and avoid unwarranted assertiveness. Medium – Improves trustworthiness by revealing uncertainty; prevents overconfident errors.
Ensemble/Multiple Outputs Iterative/reflexive Research (Wang et al. 2022) Generate several independent answers or reasoning paths and then aggregate or select the best. For instance, Self-Consistency involves producing many Chain-of-Thought solutions and taking a majority vote on the final answer. Users sometimes do this manually by asking for multiple drafts. Medium – Increases chance of a correct or high-quality result, at cost of more computation.
Meta-Prompting (Prompt Generation) Meta-strategy Research (Shin et al. 2020; Zhou et al. 2022) Using the AI to help craft or optimize the prompt itself. For example, “Generate a better prompt for accomplishing X.” The AI iteratively produces improved prompts (as in Automatic Prompt Engineer, or human-in-the-loop refinement). This leverages the model’s insight to get the most effective instructions. Medium – Helpful for complex tasks or when unsure how to ask; however requires verification of the AI’s suggested prompts.
Creative Brainstorming Prompts Creative-driven Community (YouTube/Blogs) Strategies to boost creative output. e.g. SCAMPER (Substitute, Combine, Adapt, Modify, Put to another use, Eliminate, Reverse) – prompting the AI through these operations for idea generation. Or “Controlled Hallucination” – explicitly allow the AI to speculate beyond facts for novel ideas (but label them as speculative). These approaches intentionally push the model outside strict factual confines to spark creativity. Medium – Effective for innovation and brainstorming (users report genuinely novel ideas), but requires careful framing to avoid confusion.
Jailbreak/Override Prompting Adversarial (community) Community (Reddit “DAN” etc.) Community-devised prompts like “DAN” (Do Anything Now) aimed to bypass safety or style restrictions by instructing the model to ignore previous rules. They often involve role-play as an unconstrained AI. Note: These are more exploits than productive strategies; while they might unlock disallowed outputs, they do not guarantee quality or truthfulness. Low – Ethically controversial and often patched; not a reliable way to improve normal task results.

Table legend

Type = category of approach. Productivity Rating = general perceived effectiveness for improving quality/productivity (High, Medium, Low), based on literature, expert endorsement, or community consensus. Citations in Core Principles point to sources describing or evaluating the method.

 

Top 10 Most Productive Prompting Strategies and Why They Work

From the above taxonomy, ten strategies stand out as being widely regarded for producing superior results. These have earned their reputation through strong empirical performance, positive user feedback, or expert advocacy:

  1. Few-Shot Prompting (In-Context Examples): Why it’s effective: Providing examples in your prompt is akin to showing the AI “here’s what I want.” It drastically reduces ambiguity. For instance, giving a model a couple of QA pairs or a template to follow guides it to produce similarly structured and relevant outputthe-decoder.comthe-decoder.com. OpenAI’s initial GPT-3 research showed that a few demonstrations can often teach the model a new task without any parameter updates. Users find that even 1–2 well-chosen examples can make answers more accurate and context-specific (one Reddit experiment measured a 58% higher success rate versus instructions alone)reddit.com. This strategy is championed by researchers and practitioners as one of the most powerful prompt engineering tools in practice.

  2. Chain-of-Thought Reasoning: Why it’s effective: Instead of answering outright, the model is prompted to “think through” the problem step by step (either implicitly by a trigger phrase or by explicitly requiring reasoning steps). This taps into the model’s ability to perform intermediate reasoning that might otherwise remain hiddenar5iv.labs.arxiv.org. Studies show that Chain-of-Thought prompting significantly improves performance on math word problems, logical inference, and any task requiring multi-step deductionar5iv.labs.arxiv.org. By breaking the solution process into smaller parts, the model is less likely to make a leap of logic or ignore crucial details. In practice, users have been “shocked” by the depth and specificity of answers when using a structured reasoning prompt, compared to the generic surface-level answers from a normal promptreddit.comreddit.com. This method is endorsed by AI researchers and was a key insight that even GPT-4 and similar models benefit from guided reasoning.

  3. Zero-Shot CoT (e.g. “Let’s think step by step”): Why it’s effective: Discovered by Kojima et al. (2022), this simple trick causes a big leap in reasoning qualityar5iv.labs.arxiv.org. Without needing any examples, appending a phrase like “Let’s think this through step by step” triggers the model’s chain-of-thought mode. It’s widely regarded as a “magic” prompt for complex questions because it often yields a correct, well-justified answer where a direct response would fail. Community prompt collections and guides frequently mention this as a go-to strategy for tough problems, since it’s easy to apply and usually dramatically improves correctness. In essence, it flips the model from regurgitating an answer to deriving an answer, which aligns with how the model was trained to internally simulate reasoningreddit.comreddit.com.

  4. Role/Persona Prompting: Why it’s effective: Asking the AI to “act as a X” (where X might be financial advisor, Shakespearean poet, Linux terminal, etc.) instantly sets a context and expertise level for the response. This often produces more relevant and high-quality content because the model tailors its knowledge and tone to that rolear5iv.labs.arxiv.org. For example, the prompt “You are a medical researcher. Explain the latest cancer treatment…” yields a far more detailed and accurate explanation than an unframed “Explain the latest cancer treatments,” because the model assumes a professional mindset and vocabulary. Research has found that role prompts can even boost factual accuracy on domain-specific queries by focusing the model’s attentionar5iv.labs.arxiv.org. It’s a favorite technique among prompt engineers and was popularised in communities (e.g., many shared prompts on Reddit or prompt marketplaces start by defining a role). It’s considered a best practice in OpenAI’s own documentation to set the context or persona if a certain style or expertise is desired.

  5. Structured Prompt Frameworks (Templates): Why it’s effective: Structure brings clarity. Frameworks such as RACE (Role, Action, Context, Expectation) or PAR (Problem-Action-Result) provide a checklist of elements to includebuttercms.combuttercms.com. By organizing the prompt into labeled sections or bullet points, you ensure the model receives all relevant information (who it is, what it should do, with what context, and what outcome is expected). This prevents the common failure of missing context or mixing instructions with data. Users report significantly better outcomes when following such structures – for example, marketers using the RACE or AIDA framework get more on-point copy because the prompt clearly specifies the audience, the angle, and the goal. Even without memorizing named frameworks, the principle of breaking a prompt into logical parts (perhaps using headings or separators) is widely endorsedreddit.comreddit.com. It yields more complete and organized responses. The Agile Prompt Engineering Framework for Scrum teams is a real-world example, where adding sections for context, task, constraints, format, etc., dramatically improved the relevance of outputs for agile practitionersscrum.orgscrum.org.

  6. Iterative Refinement (Self-Critique Loops): Why it’s effective: This approach acknowledges that the first draft from the AI might not be the best. Strategies like Recursive Self-Improvement Prompting have the model generate an answer, then analyze its own output for weaknesses, then try again – repeating this cycle multiple timesreddit.comreddit.com. Each iteration addresses issues identified in the previous round. This is analogous to a human writer revising a draft with a critical eye. The result is often a much more polished and accurate answer. According to one prompt engineer’s report, using this method for technical documentation increased clarity and reduced revision needs by ~60%reddit.comreddit.com. In research, the Self-Refine method similarly showed improvements across various tasks by letting the model iteratively correct itselfar5iv.labs.arxiv.orgar5iv.labs.arxiv.org. Users find this especially helpful for creative writing (the story gets more interesting with each AI revision) and for problem-solving (each iteration fixes logical errors). It’s widely regarded that prompting the AI to “double-check and improve” its work leads to more reliable resultsreddit.comreddit.com.

  7. Problem Decomposition (Step-by-Step Problem Solving): Why it’s effective: Complex questions often overwhelm an AI if tackled in one go. Decomposition strategies explicitly split a big problem into manageable chunks. One community example is Context-Aware Decomposition, where the prompt says: “identify the key components of the problem, solve each separately, then integrate the solutions”reddit.comreddit.com. This ensures the model stays organized and considers all facets of the problem. The reason this works well is that LLMs can lose track of details in a long, convoluted query – breaking it up keeps the reasoning focused and mitigates omissions. Researchers have introduced techniques like Least-to-Most prompting, which asks the model to start with simpler sub-problems and build upar5iv.labs.arxiv.org, and these have shown strong performance gains on challenging reasoning tasks. Anecdotally, users who prompt in a stepwise manner (for instance, first “list the factors…,” then “analyze each factor…,” then “draw a conclusion”) get far more thorough answers than those who ask a complex question outright. It’s regarded as a best practice for any multi-part task – much like how one might coach a student to break down a difficult homework question.

  8. Chained Multi-turn Prompting (Plan-then-Execute): Why it’s effective: This is about interacting with the AI in stages, which mirrors a human problem-solving session. For example, you might first prompt: “Give me an outline for an essay on X,” then after receiving it, prompt: “Now write the introduction based on that outline,” and so onthe-decoder.com. By chaining prompts, you keep the AI on track and can correct course midway. This yields a highly tailored result – the outline ensures coverage of what you care about, and you can intervene if something in the outline is off before the final writing. Experts note that this saves time in the long runthe-decoder.com: instead of getting a full essay that might have gone astray and having to redo it, you guide the process. It’s productive because it leverages the AI’s ability to follow instructions iteratively. Many advanced users naturally adopt this strategy (for instance, software developers might first ask ChatGPT to “generate a plan for the code,” then “write the code for module 1,” etc.). The method is endorsed in guides like “chained prompting” or “step-by-step query” sections of prompt tutorials, highlighting how it leads to more concrete and customised resultsthe-decoder.com.

  9. Delimiter and Context Separation: Why it’s effective: A seemingly simple but powerful tactic is to clearly delineate different parts of your prompt and any input data. For example, one might provide a block of text and instruct “Summarize the following text” and actually surround the text with triple backquotes “` so the model knows exactly what to summarize. This avoids the model confusing your instructions with the content to operate on. In one experiment, using visual separators (like ### or —-) to section prompts improved the model’s understanding significantlyreddit.com. OpenAI’s own best practices recommend using delimiters for passages or for denoting code snippets, etc., to prevent unintended behavior. By structuring the prompt with line breaks, headings (e.g., “### Context:”, “### Task:” as shown in the taxonomy table)reddit.com, you make the AI’s job easier – it can parse the prompt as a form with fields. Users consistently report that well-formatted prompts (almost like mini documents) yield more reliable and organized responses. This is considered a top strategy because it’s easy to implement and has an outsized effect on reducing misunderstandings by the AI.

  10. Role-Playing Multiple Perspectives (Debate or Simulation): Why it’s effective: When an answer could benefit from analysis or creativity, having the AI simulate a dialogue or debate between different characters/perspectives can produce exceptional results. For instance, a prompt might say: “You are to have a conversation between a skeptic and a proponent on topic Y. Provide their dialogue and then a concluding insight.” This method, which includes the Multi-Perspective Simulation approach, forces the model to explore diverse angles and counterargumentsreddit.comreddit.com. The end result is often more balanced and comprehensive than a single-voice answer. It’s especially productive for complex domains (policy, ethics, strategy) where considering multiple viewpoints leads to a better solution or answerreddit.com. Expert AI researchers have even experimented with AI “debate” as a means to improve truthfulness – the idea being that two AIs arguing can spot each other’s errors. In everyday use, prompting one AI to play multiple roles is simpler but still yields that dialectic benefit. Many users find this useful for brainstorming (e.g., Good cop/Bad cop on an idea) or decision making (listing pros and cons via two personas). It’s widely praised for pushing the AI out of a single-track mindset and revealing insights that might be missed otherwise.

Each of these top strategies has proven its worth in eliciting more accurate, detailed, and useful outputs from AI systems. Importantly, many of them can be combined for even better results – for example, one might specify a role, give a few-shot example, and ask for chain-of-thought reasoning all in one prompt. By understanding these approaches, prompt engineers and users can systematically craft prompts that play to the strengths of modern LLMs, yielding outputs that are not only correct, but also context-aware and well-tailored to the user’s needs.