Picco Monitor System

1

The PICCO Framework for Large Language Model Prompting: A Taxonomy and Reference Architecture for Prompt Structure

David A. Cook, Md, Mhpe

Mayo Clinic College of Medicine and Science | Mayo Clinic Division of General Internal Medicine

Abstract

Large language model (LLM) performance depends heavily on prompt design, yet prompt construction is often described and applied inconsistently. Our purpose was to derive a reference framework for structuring LLM prompts. This paper presents PICCO, a framework derived through a rigorous synthesis of 11 previously published prompting frameworks identified through a multi- database search. The analysis yields two main contributions. First, it proposes a taxonomy that distinguishes prompt frameworks, prompt elements, prompt generation, prompting techniques, and prompt engineering as related but non-equivalent concepts. Second, it derives a five-element reference architecture for prompt generation: Persona, Instructions, Context, Constraints, and Output (PICCO). For each element, we define its function, scope, and relationship to other elements, with the goal of improving conceptual clarity and supporting more systematic prompt design.

Finally, to support application of the framework, we outline key concepts relevant to implementation, including prompting techniques (e.g., zero-shot, few-shot, chain-of-thought, ensembling, decomposition, and self-critique, with selected variants), human and automated approaches to iterative prompt engineering, responsible prompting considerations such as security, privacy, bias, and trust, and priorities for future research. This work is a conceptual and methodological contribution: it formalizes a common structure for prompt specification and comparison, but does not claim empirical validation of PICCO as an optimization method.

Keywords: Prompt Engineering; Prompting Techniques; PICCO Framework; LLM Optimization; Artificial Intelligence; Human-Computer Interaction; Natural Language Processing; Medical Education; Instructional Design; Educational Technology

Introduction

Large language models (LLMs) put sophisticated artificial intelligence (AI) within reach of nearly everyone.1, 2 However, optimal use of these tools requires a new skillset – namely, how to write an effective prompt ("the input to an [LLM] that is used to guide its output"3). Although several books (e.g., these4-6) and numerous online resources (e.g., these7-9) address prompt engineering, there remains a need for a concise, evidence-informed framework to support practical prompt design.

Systematic reviews have cataloged applications and techniques of prompt engineering3, 10-14 but these largely target system developers rather than end users. One tutorial focused on writing prompts for clinical queries;15 however, broader guidance across tasks and contexts is needed. A concise, rigorously derived, user-centered prompting framework would enable educators, clinicians, and other non-experts to use these tools effectively and responsibly.

Moreover, prompt frameworks (general blueprints for structuring a prompt) are widely advocated to ensure complete, effective prompts; yet, as shown below, existing frameworks lack scholarly grounding. A well-grounded prompt framework would promote prompts that yield outputs aligned with users' intents.

The purpose of this study is to derive and describe a rigorously developed framework for prompt generation, clarify related concepts in prompt design, and outline key considerations relevant to implementation. The principles apply directly to text generation and analysis; multimodal tasks (audio, image, video) may require adaptation. Most examples draw from health professions education, but the principles apply broadly across fields including education, clinical medicine, business, computer programming, and personal life.

Key Terms

Table 1 contains definitions for terms related to prompt engineering. Historically, many of these terms have been used loosely and interchangeably; greater consistency and precision in vocabulary will facilitate communication and teaching. In particular, the term "prompt engineering" is commonly misused. Strictly speaking, prompt engineering refers to deliberate, iterative refinement of a prompt.3 "Prompt generation" is a better label for the process of composing a prompt, which is what most users do when interacting with a chat-based LLM such as ChatGPT. Additionally, users often seek to optimize reusable prompt templates that can be scaled across multiple data inputs or repeated tasks.

Prompt

"Input to a Generative AI model that is used to guide its output. Prompts may consist of text, image, sound, or other media instructions submitted to the LLM."3

Prompting

Process of providing a prompt to an LLM.

Prompt Template

Pre-defined pattern of words with placeholders that can be filled to create a complete prompt. In this example, "Please score the following essay from 1 to 10: [text]," the underlined words are the prompt template; the template + actual essay text together would constitute the full prompt.

Creation)

Process of composing a prompt.

Prompt Framework

General blueprint for structuring a prompt or prompt template to facilitate completeness and effectiveness (i.e., the component parts or elements of a complete prompt).

Prompt Element

Constituent part of a prompt framework.

Prompt Engineering

Process of deliberately, iteratively evaluating and refining a prompt to optimize performance.

Prompting Technique

Systematic method for structuring or sequencing a prompt or sequence of prompts to enhance performance (e.g., including an example of the desired output).

Technique

Systematic method for conducting prompt engineering (i.e., iteratively refining a prompt); can be manual or automated.

In-Context Learning

a) The general ability of LLMs to infer (learn) how to respond using only information contained within a prompt, without task-specific training or programming. b) [narrower usage] The inclusion of 1 or more examples (input-output pairs) within a prompt; synonymous with the prompting technique of few-shot prompting (see Table 3).

Delimiter

Characters or words that mark a segment of the prompt. This segment might comprise text to be analyzed (input data or context), specific steps to follow, or the beginning of a new section. For example, instructions could request the LLM to "Translate the text starting after ###." Delimiters help both the LLM and humans understand the prompt and obtain desired output by increasing precision, reducing ambiguity, and providing structure. Delimiters can be 1 or more unique characters (|, ''', $$$, ZZZZ), command words ([TRANSLATE]), brackets (<…>, {…}), or tags (a descriptive word in angle brackets <> marking the start and [prefaced by "/"] the end of a section, e.g., …). It usually helps to define the delimiter in the instructions, although sometimes the LLM can infer the intent (e.g., starting a new section using ###, or enclosing an essay using tags … ).

Hyperparameter

A setting specified outside the training process that shapes how the LLM generates or scores text, such as randomness, creativity, output length, or decoding strategy. Hyperparameters relevant to typical end-users include temperature, top-P, max tokens, max output length, and penalties that discourage repetition (e.g., frequency or presence penalty). Some contemporary LLMs allow model-specific hyperparameters such as reasoning level.

Temperature

A hyperparameter that controls the randomness of output by reshaping the token probability distribution. A low temperature produces a tall, narrow distribution, concentrating probability on a small set of high-likelihood tokens; as a result, the LLM selects the same tokens more often, making responses more deterministic, predictable, and reproducible. A higher temperature flattens and widens the distribution, increasing the chance of sampling lower-probability tokens and making responses more random. This is good if the goal is creative responses (e.g., poetry), novel ideas, or diverse answers across multiple replications (e.g., qualitative data analysis looking for innovative themes); however, very high temperatures often result in incoherent text and factual inaccuracies. A typical default temperature is 0.7.

Sampling)

A hyperparameter that controls the diversity of a response by adjusting the proportion of tokens eligible for selection. Conceptually it is like slicing off the top of the probability distribution. Slicing off only the very tip (e.g., top-P 0.10) provides a small set of highly probable tokens, resulting in focused and predictable output. Slicing much deeper (e.g., top-P 0.95) provides a large set of tokens with wide variability, resulting in diverse and creative (but potentially less relevant) output. A typical default top-P is 0.9. Top-k is a related hyperparameter that regulates the number instead of the proportion or eligible tokens; it is rarely used today.

Fine-Tuning

The process of further training a pretrained LLM using a curated set of input-output pairs (labeled or annotated data) so it can perform a specific task with high proficiency.

Since the LLM starts with extensive pre-trained knowledge, fine-tuning can be done using far fewer exemplars (as few as 15 for simple tasks) compared with training a model from scratch. Although there is an up-front cost to collect annotated data and complete the fine-tuning, the fine-tuned model can give high quality, high efficiency, highly-tailored responses. Fine-tuning involves repeated training cycles or epochs, each composed of several training steps that analyze a subset (batch) of the training dataset. Key fine-tuning hyperparameters include the number of epochs, batch size (number of exemplars per step), and learning rate (how much the model adjusts from one step to the next).

Generation (Rag)

A method in which an LLM retrieves relevant information from a separate knowledge base and incorporates that information into its response. LLMs can fail to give accurate or complete responses due to random errors (hallucinations) or domain-specific gaps in their training data. RAG addresses both limitations by grounding responses in an external knowledge base. The LLM interprets the user request, retrieves information from a trusted (e.g., bespoke or vetted public) knowledge base, and then generates a fluent, accurate, up-to-date, context-aware response. The quality of the final answer depends heavily on the quality of the knowledge base and retrieval strategy.

Plug-In

External software tool that extends the LLM's capabilities by enabling specialized functions. LLMs are limited in their ability to perform non-text operations such as exact math, statistical analyses, and interacting with external systems. Plug-ins enable these and other tasks (e.g., scheduling a meeting, planning travel, searching a proprietary access to the plug-in, enabling the LLM to invoke the tool automatically. More often, the LLM is prompted to generate a structured query that the user submits to a specific plug-in; the plug-in's output can then be used directly, or fed back into the LLM for further processing. Plug-in availability varies by platform, and not all LLMs have the same capabilities. For example, plug-ins must be explicitly enabled in ChatGPT or configured through the OpenAI GPT API.

Agent

Advanced software system that used LLMs in conjunction with other computer tools to autonomously plan, sequence, and complete multi-step tasks. An agent can decompose a task into component steps, plan and sequence these steps, identify and use tools to accomplish each step, and monitor its progress toward the overall goal. Examples include agents supporting automated data analysis workflows, triage, and travel planning.

Plug-ins are tools that extend LLM capabilities; agents are autonomous, goal-directed systems that use these and other tools to plan and execute multi-step processes.

(Persona Document)

Document that can be attached to every prompt addressing related tasks, containing information about the audience, setting, goals, historical background, specialized knowledge, and desired persona or tone (see Table 3 for details and examples of what might be included).

Abbreviations: AI, artificial intelligence; LLM, large language model.

How Llms Work

LLMs are often described as probability machines: they are trained using vast amounts of data to recognize patterns in text, images, and sound, and then to generate new text, images, and sound based on probabilities. Imagine the possible (probable) completion of a sentence such as "To be or not to be, that is the ___." Most readers, recognizing this as part of Hamlet's soliloquy, would choose the word "question." Phrases such as "conundrum," "dilemma," or "existential tension" would all be plausible – but not very probable. Irrelevant words such as "humanity," "quickly," or "purple"; or even random letters ("alkjnan") could also be used – but these would be implausible and thus highly improbable. LLMs generate responses by repeatedly computing a probability distribution for the next term ("token") given all prior tokens, then selecting a token from that distribution – step-by- step, 1 token at a time.

Why does this matter? First, text that is probable or plausible can still be wrong. This causes the common problem of "hallucinations." Here, the LLM generates a pattern of words that probably belong together, but result in an incorrect statement (e.g., a factual error, historical inaccuracy, logical inconsistency, or math mistake). For example, Geoff Norman – an education scientist with

6

hundreds of publications on clinical reasoning – might plausibly have published an article entitled "Etiology of Errors in Clinical Reasoning: System 1 vs System 2" in the journal Academic Medicine, but he did not.

Second, the purpose of a prompt is to increase the likelihood of certain terms in the response and decrease the likelihood of others. This explains why telling the LLM not to do something will sometimes paradoxically increase the likelihood that it does this, because the presence of a key word or phrase increases the likelihood of a response that incorporates the unwanted term.

Third, the generated output includes a degree of randomness (it is "stochastic"). This contrasts with most computer operations, which give an identical solution with each repetition ("deterministic"). Fourth, the response quality depends on the training data. LLMs merely mimic the data they were trained on. If training data are biased or inaccurate, the most probable response will reflect those biases or inaccuracies.

Finally, LLMs rely on surface-level semantics even when they appear to be reasoning logically.16 Seemingly minor changes in the prompt (e.g., phrasing, sequence, punctuation) can materially influence output.17-23 Emerging approaches – including several discussed below – attempt to

Mitigate This.24-27

Most current LLMs are also limited by the date of their training data. An LLM trained on 1 November 2024 would not know the current US president. Tools are emerging to circumvent this limitation.

Why Prompts Matter

Prompt quality matters: LLM responses are sensitive to subtle choices in wording or structure. Some very simple prompts work adequately. For example, the query, "Who was the 14th president of the United States?" will usually generate the correct response (Franklin Pierce). Other simple prompts comprise a completed example (input-output pair) followed by a new input without output, such as "Night: Day. Fast: ___" (response: "Slow"); or "Name: Nombre. Beach: ___" (response: "Playa"). The LLM inferred (learned) the intended task ("provide antonym," "provide Spanish translation") using only the example, i.e., in-context learning.

However, most prompts must be far more detailed. Asking a human, "Tell me how to make a peanut butter sandwich," illustrates the importance of detailed instructions. People often omit "obvious" steps like "get a knife" or "open the jar," yielding instructions that, if followed literally, lead to failure.

Similarly, the LLM prompt template "Analyze this reflective essay and provide a score and feedback: [essay]" is likely to generate inconsistent responses; more details and more structure are needed. A prompt framework – a general blueprint for structuring a prompt – can help to ensure completeness and increase the probability of receiving the desired response.28-30

A Rigorously Derived Prompt Framework

Several prompt frameworks have been proposed (see Table 2), but most have obscure origin and none detail a scholarly derivation of their constituent elements. For example, the earliest prompt framework is CRISPE, which (per Internet lore) was introduced internally by OpenAI staff and then diffused externally through peer networking and online dissemination. Despite CRISPE's wide use, sources document its structure but not its derivation.31 For many popular frameworks the original source is a YouTube video,32, 33 blog,34 or enterprise website.7, 35 To remedy the lack of a rigorously derived framework, I systematically identified and synthesized all available sources into a novel framework.

Identification Of Existing Frameworks

I identified existing prompt frameworks by a) searching the Internet using a search engine and 2 LLMs (a comprehensive but non-systematic approach), and b) systematically searching peer- reviewed publications (reproducible, but much narrower body of literature).

I started with an unstructured query using Google. Building from this, I separately queried OpenAI GPT-4.1 and Gemini-2.5-flash using several iteratively evolving prompts (multiple sessions, 2 October to 6 November 2025). These searches identified 8 unique prompt frameworks7, 9, 31-36 (Table 2). I also documented the earliest description of each framework. I included frameworks with obscure origin (e.g., CRISPE) if they were cited in ≥2 sources (e.g., publications, websites), suggesting acceptance in the prompt engineering community. The LLMs identified several additional "state-of-the-art frameworks" that ultimately proved to be hallucinations (novel amalgamations of best practices with no verifiable prior use).

Building from these findings and with support from an experienced research librarian, on 5 November 2025 I conducted a systematic search of Scopus, Web of Science, Embase, and Preprint Citation Index using key terms including large-language-model, LLM, prompt, framework, and 4 specific frameworks (CRISPE, COSTAR, RICCE, RICECO). I made no restrictions on language. This yielded 143 records. I selected for inclusion all publications (N=3) that described a specific prompt framework (Table 2).28-30 Most excluded studies described fully-developed task-specific prompts (N=76), novel prompting techniques (N=31), and novel prompt engineering approaches (N=25).

The full search strategy and inclusion/exclusion details are reported in Appendix Table A1.

Synthesis

I first enumerated the constituent elements and operational definitions of each framework, and identified repeatedly-shared elements (i.e., cross-framework convergence). I next considered the sequence of elements in each framework, again searching for cross-framework convergence. The final sequence integrated additional published evidence.17, 19, 22, 25

8

Table 2. Derivation of the PICCO framework: Cross-framework synthesis

Examples (4)

a Elements in the PICCO row were empirically derived from a synthesis across the other frameworks. Elements in other rows are labels from the original (existing) framework. Numbers in parentheses indicate the position of each element in the original framework.

b The elements in this column are not part of a prompt framework as defined herein (see Table 1): "Examples" are a prompting technique. "Experiment," "evaluation," and "calibration" are aspects of prompt engineering.

c Identified from Internet search using Google, OpenAI GPT-4.1, and Gemini-2.5-flash. d These frameworks have no clear scholarly derivation; they appear to have emerged among AI enthusiasts and are now accepted in the community (cited in ≥2 other publications or websites).

For each framework I have cited the earliest source I could find (e.g., YouTube video, blog post, etc.).

9

e Different sources list slightly different expansions of this acronym: most say "Examples," but some

Say "Evaluation."

f Identified from systematic search of peer-reviewed literature. g There was no explicit position of the elements in this framework. h "Examples" in the PERFECT framework are requested from the LLM (i.e., instructions request "examples or case studies, if relevant"), not provided to the LLM as part of the prompt.

Synthesis Of Frameworks

Table 2 presents the framework elements and the cross-framework synthesis. Individually, each framework was incomplete (missing at least one element), overinclusive (including "elements" better classified as prompting techniques or prompt engineering techniques), and/or redundant (distinguishing separate elements that could be collapsed into 1). Nonetheless, collectively these readily support a novel 5-element framework: Persona, Instructions, Context, Constraints, Output (PICCO; Table 3).

Several frameworks included Examples as an additional key element. However, I agree with Schulhoff3 in classifying Examples as a prompting technique (few-shot prompting) rather than a key framework element. The problems with including few-shot prompting – or any other specific prompting technique – as a key element are that in a given situation a particular technique might be unnecessary or even counterproductive; and other techniques might be as or more important.

Final Framework

Persona: The first element specifies a role for the LLM,37 such as "You are a … physician-educator," "patient with diabetes," "NIH-funded researcher," or "expert Python programmer." Why does this seemingly trivial element make a difference? LLMs have been trained using essentially all data available on the Internet, yielding a huge library of information. Assigning a persona is like selecting a specific book. Using very few words, the persona narrows the LLM's probability distribution to the desired information domain. The persona can optionally specify details such as personality, tone, and expertise. Assigning multiple people (e.g., a mock debate) or a non-human persona ("You are a horse") can generate creative insights.

Instructions: The next element specifies what the LLM should do – the central task. Instructions should be explicit, detailed, and succinct. It commonly requires multiple iterations (prompt engineering) to optimize the details, phrasing, and sequence of instructions.23, 26, 27, 29 Details might include: A specific question or desired end product, specific steps, required elements (e.g., length, rating scale), and specific techniques (see Table 4).

Context: The context specifies background information that will make the LLM response relevant, specific, and purpose-aligned. Making analogy with acting, one author explained, "The Instruction is the line you want them to say. The Context is their character's entire backstory, their motivation and the emotional state of the scene."8 Details might include the purpose or intended application; target

10

audience; setting or situation (user/institution setting, historical information, background statistics or data); specialized knowledge (theory, practice guidelines, business strategies, technical specifications); or knowledge database (e.g., for retrieval-augmented generation38). However, research shows that irrelevant context impairs performance.39 Constraints: Constraints provide guardrails – guidelines, boundaries, and rules on what not to do, or crucial requirements that merit repeating. These can include words to avoid, restrictions on length or information sources, security safeguards, and requests to confirm accuracy. Note that instructions on what to do are generally preferred over constraints on what not to do;40 constraints should emphasize – not introduce – boundaries.

Output: The final element specifies the desired response structure. Table 3 lists numerous possibilities. As regards the sequence of prompt elements, research indicates that this influences LLM output17, 22, 41, 42 but existing evidence does not support definitive universal guidelines. The PICCO sequence reflects a "majority vote" synthesis across frameworks (Table 2). However, it is unclear whether logic, personal (empirical) experience, or desire for a nice mnemonic determined the sequence in these frameworks. A minority of frameworks put Context ahead of Instructions, suggesting that Persona-Context-Instructions-Constraints-Output might be reasonable in some situations.

Supporting this idea, one website suggested that "placing the [instructions] last helps avoid the AI continuing the additional information [context] instead of focusing on the task at hand."9 More consistency across frameworks was found in putting Persona first, and Constraints and Output last.

The optimal sequence remains a question open to further research.

Specific Role Or

personality.

Informal Language That

patients can understand.

With 20 Years' Experience

doing systematic reviews.

Them To A Novel Problem Or

task.

Phrased In

positive terms).

Important To Control Type 2

diabetes.

This Plan Step By Step Before

you actually start working.

Knowledge Or

guidelines).

Empathetic Tone That Targets A

10th-grade reading level.

Attached Practice Guidelines To

guide your recommendations.

Use A Kind But Authoritative

tone.

Learning), They Should Be Included

with Context.

Imperative

requirements.

Recommendation Against The

practice guidelines.

Citation Is Real (Not A

hallucination, no fabrications).

Give Me The Pmid Or Doi For

each reference.

Of The Final

product.

The Final Product Should Be A

300-word paragraph.

Doc That I Can Share With

patients.

At A National Meeting, Using

the IMRD format. Abbreviations: LLM, large language model.

13

Table 4. Prompting techniques: Systematic methods for structuring or

Exemplars Or Other Specific

techniques.

•

Meta-prompting: ask LLM to rewrite a prompt to add or remove details or refine wording, or to generate an initial prompt.

•

Role/style prompting: specify the persona, style, tone, genre.

•

Emotion/persuasion prompting: add text to encourage "best effort" from LLM (e.g., highlight the importance

Of An Accurate Response; Add Social Cues Such As

authority or reciprocity; add a bribe or threat).

Rereading: Tell Llm To "Read The Question [Or

instructions] again" before responding.

Format, And Position The

exemplars.

Exemplar Selection: Select Optimal Exemplars;

considerations include: similar vs divergent from test sample (similar usually better); balance/distribution of classifications (balanced usually better than uneven

Distribution); Human Vs Ai Selection (Many Ai

approaches can automate and optimize selection, e.g., these24-27); and quantity of exemplars (more is usually better, but too many exemplars can adversely affect performance51).

Exemplar Labeling: Ensure Accurate Classification

(labeling, rating, annotation) of exemplars (human-

Generated, Llm-Generated [Automated] Labels) (But

note that research suggests that incorrect exemplar

Labels Do Not Necessarily Reduce Downstream

accuracy45, 46).

Exemplar Format: Present Exemplars Using Optimal

format (varies by task).

Exemplar Position And Order: Present Exemplars In

ideal position (e.g., putting the block of exemplars at the start of the prompt is usually better than in the

Middle Or End22) And Order (Sequence Of Exemplars

within the block17, 26) (varies by task).

Instructions Can Often Be Simple And Brief (Some

studies48, 49 suggest that simple or generic [task- agnostic] instructions are better than more detailed instructions when exemplars are used).

Generate Fabricated (Simulated) Exemplars In A

separate, preparatory query.

Worked Examples (Input-

reasoning-output triplets).

•

Generic (zero-shot) CoT: add a generic request "Let's think step by step" or "First tell me your approach to solving this problem" or "First, let's think about this

Logically" Before Acting On The Remainder Of The

prompt. Variations include longer or more specific instructions to think logically.

•

Few-shot CoT: include exemplars that model a correct reasoning path (i.e., input-reasoning-output "worked examples" of reasoning that LLM can then follow).

Variations include showing both correct and incorrect reasoning paths, and different ways to generate the reasoning paths.

Synthesize (Select/Combine)

responses.

•

Prompt generation and selection: generate different versions of the prompt, then generate responses for

Each, Then Synthesize Final Response (Several

approaches described).

•

Consistency: obtain multiple responses for the same

Prompt, Then Synthesize Final Response (Several

approaches described).

•

Synthesis: generate a final response; can systematically identify the single best response or combine several responses into a single response (several approaches described).

Synthesize A Final Response

from sub-task results.

•

Task decomposition: use LLM to break the task into sub-tasks (several approaches described).

•

Sub-task labeling: assign labels or placeholders to each sub-task to facilitate the combining of results at the end (several approaches described).

Sub-Task Completion: Optimize Each Sub-Task;

approaches include separately-optimized "handlers" (e.g., LLM prompt, Python script) for specific sub-tasks, using exemplars, and trial-and-error with different prompts for each sub-task.

•

Sub-task sequencing: define the sequence of sub-tasks (e.g., completed in sequence [one after another] or in parallel [all at the same time]).

•

Synthesis of sub-task results: generate or select a final response; approaches include using the final sub-task result, evaluating each sub-task result individually and selecting the best, combining sub-task results using

Explicit Logical Operators, And Using Ensemble

approaches (above) to combine responses.

The Prompt Or Response By

critiquing their own output.

The Final Response Or Reasoning

path.

•

Check accuracy of response: use LLM in a separate step to confirm the correctness of the original response (several approaches described).

•

Improve response: ask LLM for suggestions to improve the response (several approaches described).

•

Improve prompt: ask LLM for suggestions to improve the prompt or CoT reasoning path (several approaches described).

a See Schulhoff3 and learnprompting.org 58 for details and citations for each sub-technique. See Appendix Table A2 for additional details. b Some definitions of "in-context learning" include all situations (including Zero-Shot Learning) in which an LLM "learns" to complete a task from any prompt (as contrasted with requiring task- specific training).

Including Data And Exemplars

Many prompts request analysis of text or numeric data, either as a one-shot prompt or as a prompt template with a placeholder for added data. The data to be analyzed should usually be appended at the end of the prompt, or uploaded as a separate file, to clearly distinguish it from Instructions and Context. Exceptions arise when analyzing very long texts; for these, a very brief reminder prompt (PICCO prompt – data – 1-sentence recap) or interleaved/sequential analysis (task-1 prompt – data – task-2 prompt – data) may be appropriate. Best practice guidelines40 recommend use of labels and delimiters such as "### DATA ###" or "[DATA]" to clearly demarcate data (see Table 1).

Exemplars, if used for few-shot learning, should be included in the Context. Although several prompt frameworks position exemplars at or near the end (see Table 2), research shows that this impairs LLM performance.19, 22 Appending exemplars at the end may increase the risk of misinterpretation or confusion with target data, particularly in long-context tasks.9 Instead, few- shot exemplars should be part of the Context, clearly delimited and positioned before output specifications and before the target data.

Additional Prompting Strategies And Tools

Table 1 describes prompt features and tools that control or improve the LLM response, including delimiters, hyperparameters, fine-tuning, retrieval-augmented generation,38 plug-ins, and agents. The PICCO framework – indeed, any prompt framework – primarily helps the human author write a complete and optimally sequenced prompt. It is usually unnecessary to label each element in the prompt.

Uncertainties will inevitably arise in prompt wording, in how information is classified (e.g., Persona, Instructions, or Context) and sequenced, and in the selection of prompting techniques. In situations where optimal prompt performance is essential, such ambiguities should be resolved empirically through rigorous, iterative prompt engineering as outlined below.

16

Users often find themselves repeatedly writing similar prompts (e.g., to perform a literature search consistency and reduce redundant effort, a well-written prompt element or prompt template (i.e., a combination of several elements) can be saved, and then pasted or attached to future prompts.8 Some LLMs offer "Custom GPTs" or other features that enable similar functionality.

LLMs often struggle with complicated multi-step prompts. To mitigate this, clearly delimit and label each goal and/or interim step. ("This task will comprise 4 distinct steps. Complete each step entirely before proceeding to the next. ### Step 1. ….. ### Step 2. ….. [etc.]"). An alternative (and most robust) solution would break each step into separate prompts.

Prompting Techniques

Prompting techniques are blueprints for creating or structuring a prompt or sequence of prompts to enhance its effectiveness. An insightful systematic review3 identified 6 classifications of prompting techniques (Table 4, and more details in Appendix Table A2).

Zero-Shot Prompting

The most basic technique – zero-shot prompting – asks a (pretrained) LLM to generate a response without task-specific examples. Beyond using a robust prompt framework (as above), several sub- techniques can further improve zero-shot performance.

One of the most powerful techniques is meta-prompting, which asks the LLM to help write or improve the prompt. Contemporary LLMs excel in helping users craft effective prompts. Some studies have found that motivation cues can improve LLM performance; these include appeals to an authority figure, praise ("You're so impressive"), reciprocity ("I spent time preparing this information for you"), and unity ("We want to …").43 Other work suggests that emphasizing the importance of accuracy ("It is crucial that you get this right"), or adding threats or bribes ("I'll pay you a million dollars if you get this right"), can enhance performance, although more recent research contradicts these findings.44 Another technique simply instructs the LLM to "Re-read the

Few-Shot Prompting

Few-shot prompting adds 1 or more input-output exemplars to the prompt. The input typifies the text or non-text content to be analyzed. The output closely models the ideal response, both content and format. Several frameworks include "Examples" as an element, but – as proposed by Schulhoff3 – this is best viewed as an optional (albeit effective) technique.

This technique is often called in-context learning because the LLM "learns" from examples included in the prompt ("in-context") rather than a separate training phase. However, in-context learning also refers broadly to an LLM's ability to "learn" to respond to any prompt without task-specific

17

training or programming, which includes zero-shot prompting. Since both usages are correct, I avoid using in-context learning to refer solely to few-shot prompting. Design decisions in few-shot prompting focus on how to create, select, refine, format, and position the exemplars, including the number and variety of exemplars, methods to ensure accurate labeling, and how to present input-output pairs. All of these remain active topics of ongoing investigation.17, 19, 21, 22, 25-27, 42, 45-51 As explained above, exemplars should be included with the Context (never at the end). Research further shows that when multiple exemplars are used, their order influences responses, although there are no generalizable rules.17, 25, 26, 42 Incorrect exemplar labels do not necessarily reduce downstream accuracy;45 rather, evidence suggests that exemplars probably serve primarily to illustrate input/output structure, format, and task framing.45, 46 Thus, when real human-labeled exemplars are few, an LLM can generate exemplars for subsequent use (Self-Generated In-context Learning). One study found that fabricated exemplars were inferior to real exemplars but better than none.47

Chain Of Thought Prompting

The term chain of thought (CoT) prompting refers to 2 distinct techniques that both focus on the LLM's reasoning path (thought generation). Generic (zero-shot) CoT simply asks the LLM to "think step by step" or "explain your approach before giving the answer."46, 52 Many contemporary "reasoning" LLMs now do this internally without asking, but it can still be helpful to ask the LLM to articulate its reasoning to troubleshoot and optimize the prompt.

Few-shot CoT uses exemplars that explicitly model a correct reasoning path (i.e., input-reasoning- output triplets).53 Although early studies demonstrated substantial improvements with few-shot CoT, benefits with contemporary LLMs may be smaller because these models already perform

Ensembling

Ensembling encompasses a variety of prompting approaches that a) generate multiple responses and b) select or synthesize a final response. Some variations focus on the generation stage (e.g., multiple replications using the same prompt or a paraphrased prompt). Other variations focus on the selection stage (e.g., rigorous approaches to select the single best response or combine several responses into one).

Decomposition

Decomposition encompasses various approaches to systematically break a complex task into a series of smaller sub-tasks. Variations focus on how tasks are decomposed, how sub-tasks are completed, and how sub-task results are combined into a final response.

Self-Criticism

Self-criticism is an extension of meta-prompting that asks the LLM to critique its own output and use this critique to improve the prompt or revise the output. Variations focus on how the LLM critiques the output (e.g., appraising quality, identifying errors), and how it uses this information (to refine the prompt or to further refine an already-generated response).

Prompt Engineering

Prompt engineering is the process of iteratively improving a prompt (e.g., cycles of Build [a draft prompt]-Measure [evaluate]-Learn [revise];55 or Plan-Do-Study-Act56). Proposed approaches include human appraisal, meta-prompting (asking the LLM to critique the prompt or output), and automated optimization systems, each with several specific techniques.3 Human appraisals evaluate the prompt (identifying assumptions, omissions, clarifications, and irrelevant or unnecessary words) and the output (accuracy, completeness, format, and alignment with goals).7, 8, 57, 58 Meta-prompting asks the LLM to do the same.3 Several automated optimization systems have been described that evaluate outputs using an objective quality criterion and iteratively refine a prompt until it achieves a specified level of performance;3, 24, 25 Appendix Table A3 lists 25 publications identified in the literature search that describe such systems.

Following both human and automated appraisals, prompts can be iteratively refined through systematic paraphrasing, targeted additions or deletions, and using a second prompt to critically appraise the first.

Discussion

This article presents a framework and taxonomy for prompt generation derived from prompting literature and practice. While the proposed Persona-Instructions-Context-Constraints-Output (PICCO) framework has not been formally tested, it reflects rigorous synthesis across 11 expert- proposed frameworks that show increasing convergence over time. Notably, the newest framework – RICECO33 – differs from PICCO only in treating exemplars (few-shot prompting) as a framework element rather than a technique. Assuming this convergence reflects accumulating practical wisdom, the PICCO framework reasonably captures a current consensus view of prompt structure.

Prompt engineering is evolving rapidly, as is our understanding of how LLMs generate responses and how frequent, user-opaque model updates reshape LLM behaviors, including instruction- following. In my experience, newer LLMs (e.g., GPT-5.2) sometimes prioritize what they infer the user wants over what the user explicitly instructs, resulting in markedly suboptimal results, while older models (GPT-4o) demonstrate better compliance. As LLMs evolve, prompt engineering will remain a moving target, influenced by task- and model-dependent trade-offs among instruction fidelity, autonomous reasoning, predictable responses, and alignment with human intent.

19

Although reasoning algorithms and meta-prompting can offload some aspects of prompt generation, the human-in-the-loop remains essential. Skilled prompt engineers integrate understanding of LLM mechanics, domain-specific knowledge and vocabulary, clear communication, and creativity to translate complex goals into precise prompts.

Limitations

Derivation of the PICCO framework involved only 1 investigator. Given the decentralized nature of this field, a truly comprehensive search is likely impossible; however, extensive searching across web sources, LLMs, and multiple literature databases makes it unlikely that any salient frameworks were missed. Only 3 of the 11 prompt frameworks in Table 2 originate from peer-reviewed literature, and most lack robust derivation; nonetheless, they appear to be adopted in the prompt engineering community. Although the source frameworks were individually limited, their systematic synthesis represented the strongest feasible approach to creating a rigorously grounded framework.

Responsible Prompting

Responsible prompting requires attention to security, which concerns how LLMs might mishandle information or be manipulated; and alignment, which concerns whether the LLM response matches the user’s intent.57, 59-63 Even well-crafted prompts can fail when LLMs behave unpredictably or reproduce flaws in their training data. The most responsible approach is human-AI collaboration64 in which users apply consistent fact-checking and critical judgment to decide when to trust, refine, or override AI outputs. Table 5 elaborates on these issues.

Research On Prompting

Prompting is an active, dynamic field of research, and the questions proposed above represent only a small fraction of current evidence gaps. High quality, relevant investigations are desperately needed.

Importantly, the generalizability of research on prompting is limited in at least 2 ways. First, LLMs are evolving rapidly. Research on last year's LLMs may not apply to those now available, and today's findings may not apply to the next generation. Findings may also fail to generalize across LLM families with different architectures or tuning (OpenAI GPT, Gemini, Claude, Llama, or GPT-4.1 vs GPT-o4).

Second, research is highly task- and context-specific. The optimal prompt or prompting technique might differ substantially depending on the goal – such as providing factual answers, performing math, executing complex reasoning, analyzing a free-text short answer or 1-page essay, generating Python code, or translating language. It may also depend on contextual factors such as domain (academia, medicine, business, law, personal use), stakes (high, low, or casual), and whether external tools or plug-ins are enabled.

Moreover, prompting research is commonly published in preprints rather than slower peer- reviewed journals; and much knowledge exists as experiential wisdom available only through blogs

20

and social media, if at all. While perhaps inevitable in a fast-moving field, this puts the burden of research appraisal on the end-user rather than peer review. It also underscores the need to conduct and disseminate well-executed investigations of well-poised questions.

In conclusion, the framework and techniques outlined herein provide a structured approach to prompt generation while also suggesting numerous unanswered questions awaiting systematic study in this rapidly evolving field.

Table 5. Responsible prompting: Challenges and best practices in safe,

Leak Information; Or

disrupt the system.

Override The Original

instructions.

Restricted Or Unsafe

content.

Follow These

instructions.

Or Reveal

answers.

Found Inside

pasted content.

With The Intended

task.

Related Journal Articles & DOI Links

Selected peer-reviewed publications relevant to 12 Lead ECG Acquisition. Click the DOI to access the full paper (may require institutional access).

Why Choose Us?

Bangalore guidance for robotics, Spectre and autonomous systems projects.