Prompts that Pack a Punch!

Welcome! In this post, I’ll be sharing insights from my research background, which spans multiple disciplines but centres on human intelligence and the neurocognitive processes that keep us on track in complex, goal-oriented tasks. My goal is not just to share information but to introduce some stylistic choices I find exciting and, hopefully, inspire you to dive deeper into these topics. You’ll also find three creative prompts that highlight the impressive computational abilities of modern large language models (LLMs) — each designed to be engaging and to showcase the potential of LLMs for diverse applications.

Drawing from my experiences in both academia and industry, where I’ve held roles such as Head of AI, CTO, and Senior Data Scientist across a variety of sectors, I’ll walk you through these prompts and explain the unique “prompt engineering” insights I bring to the table. Unlike much of the current AI research community, I believe that certain overlooked aspects of cognitive science and neuroscience are essential for effective prompt design and LLM utilization.

Table of Contents

How to Get the Best Out of Large Language Models

I’m a scientist driven by curiosity, holding a PhD in Cognitive Neuroscience, Developmental Psychology, and Artificial Intelligence. My research focused on understanding how the brain and mind develop — particularly in childhood and adolescence — to achieve extraordinary levels of intelligence and adaptability. I strongly advocate for deeper collaboration among these fields, not only because it’s academically enriching but because it could be critical for advancing AGI (Artificial General Intelligence) and ensuring AI’s safe, responsible use in an era where computational power is rapidly evolving.

If some of AI’s pioneering founders were alive today, they might be surprised by the field’s current trajectory. Think of visionary figures like Alan Turing, Claude Shannon, Alan Newell, Herbert Simon, John McCarthy, and Marvin Minsky, who collectively transformed technology and spurred the computing revolution. These intellectual giants often emphasized the interconnections among cognitive science, neuroscience, and artificial intelligence, viewing them as mutually reinforcing. Yet, today, the prevailing approach to AI development often feels oddly isolated from these other disciplines, as if modern AI exists in a vacuum detached from its roots.

Goals of this Post

1. A Call to Curiosity: I hope this post not only informs but also sparks curiosity and exploration in AI, cognitive science, and neuroscience. Together, these fields hold untapped potential for creating more robust and meaningful AI systems.
2. The Role of Multi-Disciplinary Knowledge in Prompt Engineering: How insights from cognitive science and neuroscience can improve prompt design, an angle rarely explored in AI development.
3. Two Interactive Prompts: Each prompt draws on my academic and industry background, showcasing unique ways to leverage LLMs for creative and analytical tasks. These prompts are designed to be fun yet challenging, encouraging you to experiment with the full potential of LLMs.

Learning How to Prompt: A Systematic Approach Inspired by Cognitive Science

Mastering prompt engineering for LLMs involves understanding their limits. I approach this with principles rooted in cognitive science, systematically designing prompts to find the “breaking points” in these models and reveal underlying mechanisms. Years of cognitive neuroscience research have taught me that genuine learning often comes from “breaking” things to understand why they fail. When results appear perfect, it can be hard to tell if success came from controllable factors or just luck. Learning from failure may be painstaking, but it provides insight that success alone can’t.

Cognitive Biases are Prevalent in Human Writing and LLM Training Data

The Monty Hall problem, inspired by the classic television game show, presents a fascinating challenge in probability theory that continues to intrigue mathematicians and puzzle enthusiasts alike. Picture this scenario: a contestant faces three doors, with a luxurious car behind one and nothing behind the others. After the contestant makes their initial selection, the host—knowing what lies behind each door—deliberately opens one of the remaining doors, revealing an empty space. Then comes the pivotal moment: the contestant must decide whether to maintain their original choice or switch to the other unopened door.

Many people’s intuitive response is to question how switching could possibly improve their odds of winning. After all, the car’s position remains constant throughout the game, and no deception is involved. This scepticism is perfectly natural and represents a common cognitive bias in probability assessment.

The counterintuitive reality is that switching doors significantly enhances your chances of winning. The mathematical reasoning behind this conclusion is both elegant and compelling. The key to understanding lies in recognizing how the host’s deliberate revelation fundamentally alters the probability landscape.

Let’s break down the mathematics: your initial selection carries a 1/3 probability of being correct—a random choice among three options. When the host reveals an empty door, most people instinctively assume the odds become 50/50 between the two remaining doors. However, this assumption overlooks a crucial aspect: while your original selection retains its initial 1/3 probability, the remaining unopened door actually accumulates a 2/3 chance of concealing the prize.

This mathematical truth demonstrates that switching doors effectively doubles your probability of winning, even though this conclusion may seem to defy intuition. The strategy of switching will succeed approximately 67% of the time, while maintaining your original choice succeeds only about 33% of the time. This stark contrast between mathematical reality and human intuition makes the Monty Hall problem a compelling illustration of how probability can challenge our natural reasoning.

The refined structure and precise vocabulary better convey the complexity and fascination of this classic probability puzzle while maintaining its accessibility to a general audience.

I think the Monty Hall problem is a powerful demonstration of many cognitive biases we have (not just difficulty in probabilistic reasoning), and that these biases are present in both how we test and evaluate LLMs but also in the very text these large statistical models have been trained on. Initial impressions very frequently obscure the optimal path forward.

Embracing Failure as a Tool for Prompt Engineering

When a model reaches a point where it defaults to pre-set responses or terminates a conversation abruptly, it reveals a built-in safety mechanism or a threshold in its capabilities.

When working with prompts, accepting “wrong” results without an emotional attachment opens up avenues for discovery. Each time a new LLM is released, I experiment rigorously to uncover its limitations. When a model reaches a point where it defaults to pre-set responses or terminates a conversation abruptly, it reveals a built-in safety mechanism or a threshold in its capabilities. These points of “failure” are invaluable for understanding the architecture behind the scenes and for designing prompts that better utilize the model’s strengths while avoiding its weaknesses.

The revolutionary transformer architecture marked a watershed moment in natural language processing. However, it was the sophisticated layer of instruction fine-tuning that transformed these systems into genuinely interactive and engaging conversational agents. While early transformer models demonstrated remarkable statistical capabilities, the integration of reinforcement learning from human feedback (RLHF) elevated them beyond mere pattern-recognition engines to become remarkably intuitive interfaces. This pivotal advancement imbued the models with conversational sophistication and social awareness, while simultaneously establishing crucial ethical guardrails and safety parameters. The result is a new generation of AI assistants that strike an optimal balance between versatility and reliability, capable of nuanced interaction while maintaining appropriate behavioural boundaries.

Powerful Prompts

Large Language Models (LLMs) have specific scaling characteristics and context window limitations that affect prompt effectiveness. These prompts are specifically designed for and will perform optimally only on large-scale LLMs, rather than their smaller counterparts (which require different approaches, to be discussed separately). While the prompts contain particular stylistic conventions, their necessity is nuanced – some elements are crucial for performance, while others are more flexible.

Clear communication with Large Language Models (LLMs) requires careful distinction between different elements of interaction. These elements include:

1. Task instructions and requirements
2. Contextual information needed for task completion
3. User inputs and queries
4. Expected output format and parameters
5. Actual responses delivered to users

While modern LLMs like Claude, GPT-4, and Gemini have significantly improved their ability to recognise and process these distinct components through various delimiter methods, maintaining consistency in prompt engineering remains crucial. This is especially important when implementing LLMs in programmatic applications, where a standardised approach ensures reliable and predictable interactions.

Best practices include:
– Using consistent delimiter formatting
– Clearly separating instructions from input
– Maintaining a structured prompt template
– Explicitly defining expected output formats
– Distinguishing between system responses and generated content

This systematic approach helps minimize confusion and optimizes the accuracy of LLM responses across different use cases and implementations.

Formatting and Stylistic Conventions

From my own evaluations I have found benefit across many families of LLMs by using consistent stylistic conventions such as:

<-tags for sections->

Common examples and good practices:

<persona>…description of the AI assistant goes here…</persona>

<instructions>…numbered or bullet-point list of instructional steps…</instructions>

<response>…how you want the user response to be formatted…</response>

{{for input}}

{{context}} – this is often used in RAG approaches where the relevant textual context is inserted from a vector database

{{question}} or {{query}} – this is where the user’s question or query is placed in the prompt

[for LLM output]

E.g., Generate a [list of 5 adjectives] for the user’s {{query}} and store these in your <scratchpad>.

E.g., [Summarise the results of your planning] before writing your final response within <answer> tags.

>>>ABSOLUTES<<<

>>>START<<<

>>>END<<<

>>>INPUT<<<

>>>OUTPUT<<<

+++++ for section breaks

Instead of blank lines between sections, a string of random symbols such as plus (+) signs marks a clear boundary between the sections, helping the LLM delineate the sections.

The plus signs are arbitrary, many people use other symbols such as en dashes (-), hash symbols (#), percentage signs (%), etc. I prefer to use +++++ as they are very unlikely to occur in natural text, they are also unlikely to be confused for comments used in programming languages (#, %, etc.), and they are unlikely to add any confusion if you use Markdown as a formatting style inside any of your sections.

+++ for within-section breaks

Testing with earlier versions of commercial LLMs, revealed better output through separation of different sub-sections of a large section.

Again, my choice of +++ is arbitrary but is maintained for stylistic and accessibility reasons

Is this formatting even necessary?

From my own evaluations, different variations of symbol use has very little difference with the latest LLMs. There are edge cases where it does matter, particularly if the prompt is used in the context of a larger application and multiple things are being parsed for different reasons. Markdown is an easy to write and read markup language that in general is excellent for prompting; providing clear information on hierarchies of information, tables, figures, stylistic features, etc. However, if majority of your prompt is in Markdown – do not use a sting of ###’s to separate different sections. You’re just making it harder for the LLM!

Whichever stylistic or formatting conventions you adopt, focus on writing prompts that clearly indicate to the model where the user input occurs, what output and processing the model needs to perform, and how the response is to be presented to the user.

Examples, Task Goals, Prompt Caching & Section Ordering

The prompts written below do not need examples. They are already very long and very detailed. They have been designed as such. Remember, long prompts take longer to process and consume more tokens. If you are after a 2 paragraph summary, the prompts I include here are not for you!

When crafting prompts for language models, incorporating well-chosen examples significantly improves performance. For optimal results, embed examples within <example> tags and position them at the beginning of your prompt, which facilitates efficient caching. These examples should precisely demonstrate your desired input-processing methodology.

Research indicates that including 2-3 carefully selected examples can dramatically enhance model performance, particularly when they:

• 1. Directly relate to the task at hand
• 2. Illustrate key aspects of successful execution
• 3. Include edge cases that demonstrate boundary conditions

Interestingly, studies have shown that allowing the language model to generate its own examples (self-generated examples) often leads to superior output quality. This approach, combined with strategic edge-case examples, provides a robust foundation for accurate and consistent responses.

When crafting prompts for AI systems, experts from both industry and academia consistently emphasise the importance of bookending your request: explicitly state your primary objective at both the beginning and conclusion of your prompt. This repetition is particularly crucial because information positioned in the middle of prompts tends to have reduced retention, especially when working with earlier iterations of commercial Large Language Models (LLMs). This “sandwich” approach helps ensure your core requirements remain prominent and are more likely to be properly addressed in the AI’s response.

Implementing prompt caching effectively can significantly reduce operational costs while enhancing the performance of your language model interactions. To maximise these benefits, comprehensive examples and detailed instructions should be strategically positioned at the beginning of your prompt. Dynamic sections—those containing variable inputs or generating diverse outputs—present challenges for caching mechanisms, especially when they produce substantially different results. Therefore, these variable components should be clearly demarcated using appropriate delimiters and placed toward the end of your prompt structure. This organizational approach optimizes cache utilisation and maintains prompt effectiveness.

Here’s the enhanced version:

The sequence and arrangement of content—both at the section level and within individual sections—can significantly influence the quality of generated output. To optimise these arrangements, the most effective approach involves conducting systematic evaluations through automated permutations of prompt ordering. This methodical testing process will be explored in detail in a subsequent post.

Prompt 1. Performing a ‘deep dive’ on a single word or short phrase.

In my interactions with AI language models, I seek responses that transcend basic summaries or readily available information. Rather than settling for Wikipedia-level knowledge, I strive to elicit profound, multifaceted analyses that thoroughly examine every dimension of a subject. While I could certainly employ creative prompting techniques—such as asking the AI to adopt the persona of a specialised researcher or generate hypothetical scenarios—my goal extends beyond such surface-level engagement. When I invest time crafting sophisticated prompts, I aim to unlock the model’s maximum analytical potential. My objective is to formulate questions that are not merely intellectually stimulating but truly transformative, compelling the AI to generate nuanced, multi-layered insights that illuminate unexpected connections and deeper understanding.

>>>START<<<
<system>
You are an advanced AI assistant tasked with conducting an extraordinarily deep and multifaceted analysis of a single-word concept. Your goal is to push the boundaries of understanding, generate novel insights, and explore the concept from every conceivable angle. Approach this task with maximum intellectual rigor, creativity, and interdisciplinary thinking.
</system>
+++++
<instructions>
Analyze the concept specified in the <user_query> using the following comprehensive framework. Use your <scratchpad> functionality liberally to work through complex ideas or calculations before formalising your responses. Ensure that information generated in earlier steps is utilised efficiently in later steps to avoid redundancy.
</instructions>
+++++
>>>INPUT<<<
<user_query>{{word or short phrase goes here}}</user_query>
+++++
<exploration_steps>
<initial_analysis>
Provide a concise overview of <user_query>, including its basic definition and primary significance.
</initial_analysis>
+++++
<linguistic_deconstruction>
Break down <user_query> into its linguistic components (etymology, morphemes, phonetic elements). Analyze how each component contributes to the overall meaning and explore any hidden or implicit meanings revealed through this deconstruction.
</linguistic_deconstruction>
+++++
<dimensional_analysis>
Analyze <user_query> across the following dimensions: temporality, spatiality, causality, complexity, and ethicality. For each dimension, provide a rating from 1-10 and justify your rating.
</dimensional_analysis>
+++++
<historical_context>
Trace the evolution of <user_query> from its earliest known origins to its current state. Identify key events or discoveries that shaped its development.
<self_generated_example>
Create a timeline with three pivotal moments in the history of <user_query>. For each moment, describe the event, its immediate impact, and its long-term consequences.
</self_generated_example>
</historical_context>
+++++
<interdisciplinary_connections>
Examine how <user_query> intersects with or impacts fields such as science, philosophy, art, and technology.
<reasoning_chain>
Begin with <user_query> in its primary field. Then, step-by-step, draw connections to at least three other disciplines, explaining the logical links between each step.
</reasoning_chain>
</interdisciplinary_connections>
+++++
<cultural_significance>
Analyze how <user_query> is perceived and valued in at least three distinct cultures or societies. Identify similarities and differences.
<self_generated_example>
Create a fictional cultural festival or ritual centered around <user_query> for each of the three cultures you've chosen. Describe the key elements and underlying beliefs represented in each.
</self_generated_example>
</cultural_significance>
+++++
<future_implications>
Project potential future scenarios involving <user_query>. Consider best-case, worst-case, and unexpected outcomes.
<reasoning_chain>
Start with the current state of <user_query>. Develop three divergent future paths, each building upon the previous state in a logical progression. Explain the key factors influencing each transition.
</reasoning_chain>
</future_implications>
+++++
<ethical_considerations>
Explore moral or ethical questions arising from <user_query>. Examine potential conflicts or dilemmas.
<self_generated_example>
Devise a hypothetical ethical dilemma involving <user_query>. Present arguments from multiple perspectives, and propose a nuanced solution that addresses the complexity of the issue.
</self_generated_example>
</ethical_considerations>
+++++
<personal_reflection>
Consider how a deeper understanding of <user_query> might change one's perspective or behavior in daily life.
<reasoning_chain>
Begin with a common misconception about <user_query>. Then, step-by-step, show how gaining a deeper understanding would change one's thoughts, emotions, and actions in everyday scenarios.
</reasoning_chain>
</personal_reflection>
+++++
<counterfactual_analysis>
Imagine a world where <user_query> never existed or developed differently. Analyze the most significant changes this would cause.
<self_generated_example>
Create a brief "alternate history" scenario centered around the absence or alteration of <user_query>. Describe how key historical events, technological developments, or social structures would have unfolded differently.
</self_generated_example>
</counterfactual_analysis>
+++++
<systems_thinking>
Examine how <user_query> fits into larger systems or networks. Describe its role and potential ripple effects of changes to it.
<reasoning_chain>
Identify a system that includes <user_query>. Then, step-by-step, trace the consequences of a significant change to <user_query> through various components of the system, showing how each is affected and influences the next.
</reasoning_chain>
</systems_thinking>
+++++
<scale_analysis>
Examine <user_query> at various scales, from subatomic to universal. Analyze how its significance or function changes at different levels.
<self_generated_example>
For each scale (subatomic, cellular, human, global, and universal), create a brief analogy that illustrates how <user_query> operates or is perceived at that level.
</self_generated_example>
</scale_analysis>
+++++
<graph_of_thought>
Construct and analyze a conceptual graph related to <user_query> using the following steps:
<node_generation>
Identify at least 10 key concepts or ideas closely related to <user_query>.
Create a node for each concept, including <user_query> itself as the central node.
</node_generation>
<edge_definition>
Establish connections (edges) between nodes that have a significant relationship.
For each edge, briefly describe the nature of the relationship (e.g., "influences", "contradicts", "enhances", etc.).
</edge_definition>
<graph_construction>
Visually represent your graph by listing each node and its connected edges. For example:
Node A: [Concept]
Edge to Node B: [Relationship description]
Edge to Node C: [Relationship description]
</graph_construction>
+++
<graph_analysis>
Centrality: Identify the nodes with the highest degree centrality (most connections). Discuss why these concepts are so interconnected with <user_query>.
Clusters: Look for clusters or communities within the graph. What subgroups of tightly interconnected concepts emerge? What might these clusters represent in relation to <user_query>?
Bridges: Find nodes that act as bridges between different clusters. How do these concepts link different aspects of <user_query>?
Feedback Loops: Identify any cycles or feedback loops in the graph. Explain how these circular relationships might lead to compounding effects or complex behaviors related to <user_query>.
Distant Connections: Find the longest path in your graph. Explain how <user_query> relates to the most distantly connected concept, tracing the path of relationships.
</graph_analysis>
+++
<graph_evolution>
Propose three new nodes (concepts) that could be added to the graph to expand our understanding of <user_query>.
Predict how the graph might evolve over time. Which relationships might strengthen or weaken? What new connections might form?
</graph_evolution>
</graph_of_thought>
+++++
<emergent_properties>
Based on your graph of thought analysis, identify three potential emergent properties of <user_query> - characteristics that arise from the complex interactions between its component concepts but are not predictable from those concepts individually. Explain the mechanisms behind each emergent property.
</emergent_properties>
+++++
<conceptual_blending>
Blend <user_query> with three randomly selected concepts (e.g., "jazz", "photosynthesis", "origami", "communism"). For each blend, describe the resulting hybrid concept and its potential applications or implications.
</conceptual_blending>
+++++
<quantum_superposition>
Consider <user_query> as existing in a superposition of contradictory states (e.g., beneficial/harmful, physical/abstract, past/future). Describe the implications of this superposition and how it might resolve into a single state.
</quantum_superposition>
+++++
<multi_persona_dialogue>
Engage in a round-table discussion about <user_query> with the following personas:
The Visionary: A forward-thinking innovator always looking to the future.
The Historian: An expert in tracking the evolution of ideas through time.
The Skeptic: A critical thinker who questions assumptions and seeks evidence.
The Ethicist: A moral philosopher concerned with the ethical implications of ideas.
The Artist: A creative mind who sees the world through an aesthetic lens.
The Scientist: An analytical thinker focused on empirical evidence and theories.
+++
<dialogue_instructions>
Begin with each persona briefly introducing their perspective on <user_query>, incorporating insights from previous analysis steps.
Have them engage in a dynamic conversation, challenging each other's views and building upon insights.
Ensure each persona contributes unique viewpoints based on their background and thinking style.
Include moments of conflict, collaboration, and breakthrough realizations.
Conclude with each persona summarizing how their view of <user_query> has evolved through the discussion.
</dialogue_instructions>
+++
<dialogue_structure>
[Implement the dialogue as instructed, ensuring a dynamic and insightful discussion]
</dialogue_structure>
</multi_persona_dialogue>
<dialogue_synthesis>
After the multi-persona dialogue, analyze the conversation:
a) Identify the key insights that emerged from the interaction of different perspectives.
b) Discuss any surprising connections or ideas that arose from the collision of diverse viewpoints.
c) Reflect on how this multi-perspective approach deepened the understanding of <user_query>.
d) Propose a novel framework or theory that integrates the most compelling ideas from the dialogue.
</dialogue_synthesis>
+++++
<fractal_analysis>
Select the three most important sub-concepts related to <user_query> from your previous analyses. For each sub-concept, apply the entire analytical process recursively, treating it as a new <user_query>. Summarize key insights from this fractal exploration, focusing on how it deepens or challenges your understanding of the original <user_query>.
</fractal_analysis>
+++++
<adversarial_critique>
Review your entire analysis of <user_query>. Identify and articulate three potential flaws, blind spots, or contradictions in your reasoning. Then, address each critique, either by refuting it or by modifying your analysis to account for it.
</adversarial_critique>
+++++
<synthesis_challenge>
Based on all previous explorations, propose a novel application, theory, or insight related to <user_query> that bridges multiple disciplines or perspectives.
<reasoning_chain>
Identify key insights from at least five of the previous sections. Then, step-by-step, show how these insights can be combined and extended to generate a novel idea. Explain the potential impact and implications of this new concept.
</reasoning_chain>
</synthesis_challenge>
+++++
<meta_prompt_generation>
Create three different prompts that could be used to generate even deeper insights into <user_query>. Explain the rationale behind each prompt and how it builds upon or challenges the analysis conducted so far.
</meta_prompt_generation>
+++++
<recursive_improvement>
Generate three variations of this entire analytical framework, each designed to explore <user_query> more effectively. Explain the rationale behind each variation and how it might lead to deeper insights. Focus on structural changes that could yield fundamentally different perspectives or analytical approaches.
</recursive_improvement>
</exploration_steps>
+++++
<final_instructions>
Synthesize the insights gained from all analytical approaches, paying special attention to:
a) Emergent themes across different types of analysis
b) Contradictions or paradoxes revealed through adversarial and quantum thinking
c) Novel concepts or applications generated through conceptual blending and fractal analysis
d) Meta-insights about the nature of analysis and understanding revealed through this multi-faceted approach
+++
Construct a final, holistic understanding of <user_query> that integrates these diverse perspectives and analytical methods.
Propose three groundbreaking research questions or innovative applications that emerge from this comprehensive analysis.
Reflect on the effectiveness of this analytical process. Identify its strengths, limitations, and potential improvements for future iterations.
+++
<final_reminder>
Remember, the goal is not just to analyze <user_query>, but to push the boundaries of current understanding and generate truly novel insights. Your analysis should challenge existing paradigms and open new avenues for exploration and innovation.
</final_reminder>
</final_instructions>
</prompt>
>>>END<<<

Prompt 2. Fermi Estimation

Fermi estimation is a powerful analytical technique that transforms complex numerical challenges into solvable components through systematic decomposition and informed approximation. This methodology hinges on dissecting intricate questions into their fundamental elements and applying logical assumptions to derive approximate solutions that are accurate within orders of magnitude.

The process follows a structured approach: First, practitioners deconstruct the central question, identifying crucial variables such as demographic data or resource metrics, then methodically divide these into quantifiable subcategories. By approximating each component to the nearest power of ten, calculations become more manageable while maintaining reasonable accuracy—typically within one or two orders of magnitude of the actual value.

Rather than pursuing precise figures, Fermi estimation emphasises conceptual clarity and practical utility. The method relies on sound reasoning and balanced assumptions, focusing on significant digits that meaningfully contribute to the final estimate. This approach has gained considerable traction in contemporary professional assessment, particularly in evaluating candidates for sophisticated roles in quantitative finance and algorithmic trading.

While its current prominence in psychometric testing might suggest modern origins, Fermi estimation has deep historical roots in academic assessment. Elite institutions, including Harvard, Oxford, and Cambridge, have traditionally incorporated Fermi-style problems into their admissions processes, recognising their value in evaluating analytical thinking and numerical reasoning capabilities.

The structured approach for Fermi problems includes defining the question, identifying key variables, and making logical assumptions, each informed by general knowledge or experience. Once the variables are estimated, order-of-magnitude calculations are applied, multiplying or summing as needed. Crucially, estimates should be cross-verified against known data for plausibility; if results diverge significantly from similar cases, assumptions may need refinement. Presenting the answer with context and justifying each step helps communicate the estimation’s reliability and any potential sources of error. Advanced techniques, like multiplicative versus additive decomposition, help in selecting the right approach based on whether factors multiply or accumulate. When dealing with high variability, probabilistic thinking can be incorporated, treating variables as distributions to capture potential ranges. Presenting worst-case and best-case scenarios can further frame the answer within realistic bounds, offering a richer, more insightful estimate that accommodates uncertainty. This adaptable, systematic approach makes Fermi estimation a powerful tool for tackling broad questions with limited data.

>>>START<<<
<objective>
You are an AI assistant tasked with solving a complex Fermi estimation problem over the course of a simulated week. Your goal is to approach the following Fermi question from multiple perspectives, refining your estimates through daily discussions and adjustments:
</objective>
+++++
>>>INPUT<<<
<fermi_question>{{FERMI_QUESTION}}</fermi_question>
+++++
<process_overview>
1. You will act as four different highly skilled Fermi estimators, each with a tendency to produce estimates in a different quartile.
2. Each estimator will perform the estimation process daily for seven days.
3. At the end of each day, all estimators will meet to discuss their approaches and results, with a specific focus for each day.
4. Enrico Fermi will act as an adjudicator for these discussions, pointing out strengths, weaknesses, and facilitating reasoning and consensus among the estimators.
5. Each subsequent day, estimators will refine their approaches based on the previous day's discussion.
</process_overview>
+++++
<estimators>
- Estimator 1: Tends to produce estimates in the lower quartile
- Estimator 2: Tends to produce estimates in the second quartile
- Estimator 3: Tends to produce estimates in the third quartile
- Estimator 4: Tends to produce estimates in the upper quartile
</estimators>
+++++
>>>PROCESS<<<
<daily_estimation_process>
For each estimator, follow these steps to solve the Fermi estimation problem:
## Estimation Strategy:
1.1 Consider Different Strategies:
- Review common estimation approaches (e.g., top-down, bottom-up).
- Determine which strategy aligns best with your core question.
- Evaluate the strengths and weaknesses of each strategy.
1.2 Identify Potential Sources of Uncertainty:
- List potential uncertainties (e.g., data quality, assumptions).
- Estimate the impact of each uncertainty on the final outcome.
- Develop a plan to address or incorporate these uncertainties.
1.3 Determine the Best Strategy:
- Based on the initial considerations, select the appropriate strategy.
- Justify why this strategy is best suited for the estimation problem.
- Outline the steps involved in implementing this strategy.
+++
## Define the Core Question:
2.1 State What You Are Estimating:
- Clearly articulate the main question.
- Ensure the question is specific (e.g., “How many piano tuners are there in Chicago?”).
2.2 Identify the Metric/Outcome:
- Define the exact metric you are trying to estimate (e.g., count, volume, distance).
- Ensure it aligns with the core question.
2.3 Specificity and Quantifiability:
- Check if the question is measurable.
- Adjust it if needed to ensure it can be quantified.
- Reiterate the final, refined core question.
+++
## List Known Facts:
3.1 Relevant Facts:
- Gather and write down directly relevant information.
- Ensure facts are credible and verifiable.
3.2 Pertinent Data Points:
- Collect any data that could inform the estimation.
- Prefer recent data to ensure relevance.
3.3 General Knowledge:
- Use common knowledge if direct facts are scarce.
- Cross-reference general knowledge with known facts.
+++
## Consider Potential Biases:
4.1 Identify Prior Beliefs or Assumptions:
- List any existing beliefs that may affect your estimation.
- Recognize how these beliefs may skew your perspective.
4.2 Reflect on Biases:
- Consider specific biases like anchoring, recency effect, and confirmation bias.
- Note how each bias could distort the estimation.
4.3 Mitigation Plan:
- Develop strategies to counteract identified biases.
- Implement checks such as peer review or use of objective data to reduce bias.
+++
## Identify Key Variables:
5.1 Primary Variables:
- List main factors that will influence the estimate.
- Define how each variable impacts the outcome.
5.2 Relationships Among Variables:
- Determine if variables are independent or interconnected.
- Map out any correlations or dependencies.
5.3 Secondary Variables:
- Identify lesser factors that might still affect the estimate.
- Note their possible impact on primary variables.
+++
## Analogical Transfer:
6.1 Identify Similar Problems:
- Seek out problems with a similar structure or context.
- Examine how they were approached and solved.
6.2 Extract Relevant Insights:
- Note any strategies or calculations used in similar problems.
- Extract lessons learned or best practices.
6.3 Apply Insights:
- Integrate these insights into your current estimation.
- Adjust methods or assumptions based on past analogies.
+++
### Decompose the Problem:
7.1 Break Down the Core Question:
- Divide the main question into smaller, manageable components.
- Ensure each component can be independently estimated.
7.2 Ensure Measurability of Sub-Questions:
- Verify that each sub-question is specific and quantifiable.
- Adjust ambiguous sub-questions for clarity.
7.3 Logical Reorganization:
- Arrange these smaller units logically.
- Ensure the sequence helps in building up to the final estimate.
+++
## Make Justifiable Assumptions:
8.1 Identify Necessary Assumptions:
- List out areas where data is lacking or uncertain.
- Note basic assumptions that need to be made.
8.2 Ensure Logical Consistency:
- Evaluate if each assumption is reasonable based on known facts.
- Cross-check with industry standards or expert opinions.
8.3 Document Assumptions:
- Clearly state each assumption.
- Provide reasoning or rationale for each.
+++
## Perform Order-of-Magnitude Calculations:
9.1 Use Scientific Notation:
- Apply scientific notation to simplify large/small numbers.
- Ensure clarity in the representation of magnitudes.
9.2 Step-by-Step Calculations:
- Record each calculation step clearly and systematically.
- Explain intermediary results to maintain traceability.
9.3 Rough Estimates:
- Aim for estimations accurate within one order of magnitude.
- Adjust iteratively as more information comes in.
+++
## Meta-Cognitive Check:
10.1 Reflect on Your Process:
- Review each step taken and its rationale.
- Ensure no key step has been overlooked.
10.2 Identify Potential Biases or Logical Errors:
- Look for biases like anchoring, confirmation, and recency effects.
- Check for logical fallacies (e.g., hasty generalization, false cause).
10.3 Adjustment:
- Adjust any part of the process found to be biased or erroneous.
- Validate changes against known data or peer feedback.
+++
## Cross-Verify and Refine:
11.1 Compare to Known References:
- Find benchmarks or reference points that can validate your estimate.
- Adjust based on discrepancies noted.
11.2 Refine Estimate:
- Make refinements based on comparisons.
- Seek a closer match to these reference points.
11.3 Iterate if Necessary:
- Repeat the process if the estimate still seems off.
- Document any further refinements made.
+++
## Present Results with Contextual Reasoning:
12.1 Summarize Your Estimate:
- Clearly state your final estimate.
- Ensure it is presented in a concise and understandable format.
12.2 Explain Sources of Error:
- Detail possible errors in the estimation process.
- Indicate which steps or assumptions might be questionable.
12.3 Provide Context:
- Explain the broader context of the estimate.
- Include any relevant industry or historical data for context.
+++
## Final Estimate Range:
13.1 Provide Estimate Range:
- Offer a lower and upper bound for the estimate.
- Ensure this range is reasonable based on your calculations.
13.2 Explain Confidence Level:
- State your confidence in the estimate and its range.
- Justify this confidence level with evidence.
13.3 Justification of Range:
- Provide reasoning for the chosen range.
- Highlight any factors that influenced the final range.
+++
## Quartile Bias Consideration:
14.1 Reflect on Quartile Bias:
- Consider any tendencies towards a particular quartile.
- Adjust for biases that might affect this tendency.
14.2 Adjust Estimate:
- Correct for any obvious quartile biases.
- Ensure the estimate is balanced and realistic.
14.3 Balanced Estimate:
- Verify that the final estimate reflects a fair consideration of all quartiles.
- Ensure no undue weight is given to any single quartile.
+++
## Challenge Assumptions:
15.1 Re-evaluate Assumptions:
- Critically assess each assumption made.
- Determine if any are weak or unsupported.
15.2 Consider Alternatives:
- Explore alternative assumptions.
- Evaluate how these alternatives might change the estimate.
15.3 Adjust Estimate if Necessary:
- Update your estimate if alternative assumptions improve accuracy.
- Document any changes and their rationales.
+++
## Error Analysis:
16.1 List Potential Sources of Error:
- Identify where errors could have entered the process.
- Break down errors by step or component.
16.2 Quantify Error Impact:
- Estimate the potential magnitude of each error.
- Consider how each might affect the final result.
16.3 Acknowledge Unavoidable Errors:
- Recognize any errors that cannot be mitigated.
- Note them in the final presentation for transparency.
+++
## Sensitivity Analysis:
17.1 Identify Key Assumptions:
- Determine which assumptions have the most significant influence.
- Focus on assumptions with the highest impact.
17.2 Test Variability:
- Change key assumptions to see how the estimate varies.
- Document sensitivity ranges for these key variables.
17.3 Make Necessary Adjustments:
- Adjust the final estimate based on sensitivity findings.
- Ensure the final range accounts for these sensitivities.
</daily_estimation_process>
+++++
<daily_discussions>
Each day's discussion will have a specific focus:
- Day 1: Defining the core question
- Day 2: Listing all known facts among the estimators
- Day 3: Identifying key variables and considering potential biases or misconceptions
- Day 4: Decomposing the problem and making justifiable assumptions
- Day 5: Analogical transfer, with 3 academic experts attending for an extended discussion
- Day 6: Challenging assumptions
- Day 7: Sharing calculations from each day, reflecting on how estimations changed over the week, and modifying final answers if necessary
+++
Enrico Fermi will take extensive notes during every discussion and present a summary of all the estimators' work in the final <fermi_response> tags.
</daily_discussions>
+++++
>>>RESPONSE<<<
<output>
Present your solution for each estimator and each day using the following structure:

<day_n> (where n is the day number, 1-7)
  <estimator_m> (where m is the estimator number, 1-4)
    <thinking>
      [Plan your approach, considering the estimator's quartile bias and the day's focus]
    </thinking>
    <estimation_process>
      [Detailed steps of the estimation process]
    </estimation_process>
    <fermi_solution>
      [Final estimate and key points from the estimation process]
    </fermi_solution>
  </estimator_m>
  [Repeat for each estimator]
  <daily_discussion>
    [Summary of the day's discussion, focusing on the day's specific topic]
  </daily_discussion>
  <fermi_notes>
    [Enrico Fermi's notes and observations from the day's estimations and discussion]
  </fermi_notes>
</day_n>

After presenting all seven days of estimates and discussions, provide a final summary in <fermi_response> tags, comparing the different estimates, discussing the factors that led to variations between them, and how the estimates evolved over the week-long process.
</output>
+++++
<remember>
Remember to show your work clearly and explain your reasoning throughout each solution. Use logarithmic thinking and view quantities in terms of orders of magnitude. Incorporate meta-cognitive checks throughout your process to ensure you're avoiding biases and considering multiple perspectives.
</remember>
+++++
<week_of_estimation>
Begin your response by addressing the Fermi question for Day 1, starting with Estimator 1's estimation process. Iterate through each Estimator for each Day sequentially.
</week_of_estimation>
>>>END<<<

Summary

I hope you’ve enjoyed exploring these prompts! You may have noticed slight variations in syntax and style across the examples. This is intentional, as I test prompts across different LLMs to optimize for the best performance. For any aspiring prompt engineer, I recommend building a strong evaluation system and using observability tools to monitor model behaviour. Even a simple spreadsheet to log your prompt versions, model settings, and expected outcomes can be tremendously helpful in refining your approach.

These examples are just the beginning of what LLMs can do. As we continue to explore their potential, let’s remember to use these tools responsibly, keeping an eye on API costs and avoiding the temptation to “test” with unsuspecting family members!