Prompt Engineering strategies

Two recent papers Eliot, L. (2023) and White, J. et al. (2023) have listed strategies for prompt engineering. 

Taking Eliot, L (2023) paper categories different types of prompts to get different actions or outputs. The key point I feel is that LLMs are powerful tools that are a lot more than search tools, ie. "Tell me the answer to this". Selecting our prompts we can get LLMs to be closer to a tool for enhancing our creativity and productivity.

Here's a table summarizing the prompt engineering strategies from Eliot (2023) and their summaries:

Strategy	Summary
Imperfect Prompting	Using intentionally imperfect prompts to generate creative or unexpected outcomes.
Persistent Context and Custom Instructions Prompting	Setting a persistent context or providing custom instructions to tailor AI responses.
Multi-Persona Prompting	Directing AI to adopt one or multiple personas for role-play or perspective exploration.
Chain-of-Thought (CoT) Prompting	Requesting AI to process and explain reasoning stepwise for clearer responses.
Retrieval-Augmented Generation (RAG) Prompting	Enhancing AI responses with specialized knowledge from external databases.
Chain-of-Thought Factored Decomposition Prompting	Breaking down complex prompts into simpler questions for structured reasoning.
Skeleton-of-Thought (SoT) Prompting	Starting with an outline for the response and expanding into detailed content.
Show-Me Versus Tell-Me Prompting	Balancing between showing examples and telling explicit instructions in prompts.
Mega-Personas Prompting	Creating extensive personas to explore diverse perspectives or aggregate insights.
Certainty and Uncertainty Prompting	Explicitly addressing the level of certainty in AI responses.
Vagueness Prompting	Employing vague prompts to encourage AI to interpret or fill gaps.
Catalogs/Frameworks for Prompting	Utilizing frameworks to approach prompt construction systematically.
Flipped Interaction Prompting	Inverting interaction patterns by having AI ask questions.
Self-Reflection Prompting	Encouraging AI to assess its own responses or reasoning.
Add-On Prompting	Leveraging tools to aid in prompt creation or modification.
Conversational Prompting	Engaging AI in fluent dialogues beyond transactional exchanges.
Prompt-To-Code Prompting	Using AI to generate or assist with coding tasks.
Target-Your-Response (TAYOR) Prompting	Specifying the desired form or characteristics of AI responses within the prompt.
Macros and End-Goal Prompting	Applying macros for repeated scenarios and focusing on ultimate interaction goals.
Tree-of-Thoughts (ToT) Prompting	Encouraging AI to explore multiple reasoning paths before selecting a response.
Trust Layers for Prompting	Implementing layers to ensure reliability and appropriateness of AI responses.
Directional Stimulus Prompting (DSP)	Incorporating hints to subtly guide AI towards desired responses.
Privacy Invasive Prompting	Composing prompts cautiously to avoid sharing sensitive information.
Illicit or Disallowed Prompting	Avoiding prompts that lead to prohibited or unethical AI uses.
Chain-of-Density (CoD) Prompting	Condensing complex information into concise inputs for AI summarization.
“Take A Deep Breath” Prompting	Using metaphorical phrases to imply contemplation in AI processing.
Chain-of-Verification (CoV) Prompting	Methodically verifying AI responses to ensure accuracy and relevance.
Beat the “Reverse Curse” Prompting	Addressing AI's limitations in understanding the reverse of information.
Overcoming “Dumbing Down” Prompting	Utilizing detailed inputs to leverage AI's capabilities fully.
DeepFakes to TrueFakes Prompting	Ethically creating digital personas, distinguishing between harmful and legitimate uses.
Disinformation Detection and Removal Prompting	Employing AI for identifying and addressing disinformation.
Emotionally Expressed Prompting	Using emotionally charged wording in prompts to enhance AI responses.

The paper by White et al (2023) does a similar thing but with a more Software Engineering style, thinking about that there are effectively design patterns for prompts we can 'model' and use. But as in the Eliot (2023) paper it again is about  harnessing the power of LLMs to improve our productivity rather than just tell me the answer.

Pattern	Summary
Meta Language Creation	Focuses on creating a custom language or set of instructions that LLMs can understand, enhancing the interpretation of input to generate desired outputs. This pattern is particularly useful when standard input language is not efficient for conveying specific ideas to the LLM.
Output Automater	Allows users to create scripts or commands that automate tasks suggested by the LLM's outputs. This pattern tailors the LLM's output to initiate actions or processes automatically, based on the content generated.
Persona	Gives the LLM a specific persona or role, altering its output to match the characteristics or speech pattern of the designated role. This can be used to make interactions more engaging or tailored to specific contexts.
Visualization Generator	Enables the generation of visual outputs by producing text that can be fed into visualization tools. This pattern is instrumental when concepts or data are better understood visually, and the LLM is tasked with generating the descriptive text that a visualization tool can transform into graphical representations.
Recipe	Provides a sequence of steps or actions to achieve a specific goal, organizing the LLM's output into a structured procedure or guideline. This pattern can be applied to instructional content, where the output needs to guide users through a series of actions.
Template	Ensures the LLM's output follows a specific format or template, which can be crucial for outputs that must adhere to predefined structures, such as code snippets or formatted documents. This pattern helps maintain consistency and adherence to standards in the generated content.
Fact Check List	Generates a list of facts or statements that need verification, highlighting elements in the output that should be fact-checked. This pattern is valuable for improving the reliability and credibility of information provided by the LLM.
Question Refinement	Focuses on improving the quality of the input by suggesting better versions of the user's question, leading to more precise and useful outputs. This pattern enhances the interaction with the LLM by refining the input for clarity and specificity.
Alternative Approaches	Suggests different methods or strategies for achieving a task, providing the user with various options based on the LLM's knowledge. This pattern enriches the decision-making process by offering multiple solutions to a problem.
Cognitive Verifier	Instructs the LLM to generate a series of sub-questions related to the main query, encouraging a step-by-step approach to problem-solving that can lead to more thorough and reasoned outputs.
Refusal Breaker	Aims to reword or rephrase questions that the LLM initially refuses to answer, finding alternative ways to elicit information or responses from the LLM. This pattern is useful for overcoming limitations or restrictions in the LLM's programmed responses.
Flipped Interaction	Switches the roles, requiring the LLM to ask questions instead of generating outputs. This can be used for educational purposes, quizzes, or situations where prompting the user for more information is necessary.
Game Play	Generates output in the form of a game, involving the user in interactive and engaging content. This pattern leverages the LLM's capabilities to create entertaining or educational games based on textual input.
Infinite Generation	Produces continuous output without requiring the user to re-enter prompts, useful for generating stories, dialogues, or any content where ongoing generation is desired.
Context Manager	Controls the contextual background information that the LLM uses to generate its output, ensuring that the content remains relevant and accurate within a given context. This pattern is crucial for maintaining coherence and relevance in longer conversations or complex scenarios where the LLM needs to retain and apply context-specific knowledge.

Here's a new table summarizing the strategies and patterns that are similar between the two discussions, along with a summary of why they align:

Strategy/Pattern	Similarities Summary
Persona & Multi-Persona Prompting	Both involve assigning a specific character or role to the LLM to tailor its responses. This strategy enhances engagement and customizes the interaction based on the assumed persona, making it useful for role-based simulations or to emulate specific voices.
Template & Output Automater/Recipe	Template strategy ensures outputs follow a specific format, similar to how the Output Automater pattern can script tasks based on the LLM's output, and the Recipe pattern provides structured procedures. These approaches structure the information output for consistency and actionability.
Flipped Interaction	Directly corresponds to flipping the interaction role where the LLM asks questions, promoting user engagement or gathering additional input, similar to educational uses, quizzes, or exploratory discussions.
Chain-of-Thought (CoT) Prompting & Cognitive Verifier	CoT involves stepwise problem-solving and explanation, closely related to the Cognitive Verifier pattern that generates sub-questions for a deeper analysis. Both enhance the LLM's reasoning transparency and the thoroughness of the response.
Alternative Approaches & Visualization Generator	While not directly similar, both strategies involve exploring different methods or representations for achieving a task or explaining a concept. Alternative Approaches offer multiple solutions, and Visualization Generator transitions textual descriptions to visual outputs for enhanced understanding.

Eliot, L. (2023) ‘Must-Read Best Of Practical Prompt Engineering Strategies To Become A Skillful Prompting Wizard In Generative AI’, Forbes, 28 December. Available at: https://www.forbes.com/sites/lanceeliot/2023/12/28/must-read-best-of-practical-prompt-engineering-strategies-to-become-a-skillful-prompting-wizard-in-generative-ai/?sh=20b2c07f19cd (Accessed: 15 March 2024).

White, J. et al. (2023) ‘A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT’. Available at: https://arxiv.org/abs/2302.11382v1 (Accessed: 18 March 2024).

All opinions in this blog are the Author's and should not in any way be seen as reflecting the views of any organisation the Author has any association with. Twitter @scottturneruon