The Best Prompts are Conversational Archetypes

Most people prompt the wrong way (using "mega prompts"). You'll get far better output if you have dialogues with Deep Archetypes.

1. GPT-3 Made My Client an Extra Million Dollars

My eyes were blurry and bloodshot from hours of brainstorming, drafting, writing, editing, revising, scrapping, rewriting, deleting, and writing some more.

I was working on an advertising campaign for a high-end wine cellar business. And my job was to create a new ad that beat the two current ads (that were responsible for the campaign's success so far).

Ideas were expanded into full campaigns. The work of a full-time team of 4, working a full month, was compressed into a week—in search of a new hook that would trigger conversions and sales.

Nothing “looked right” or captured what I was trying to communicate. I was too close to everything and I found myself thinking and writing in circles.

I glanced over at my tab with a rinky-dink chat window I had coded myself (connected to GPT-3 via API), highlighted a few things in a document, copied, pasted into the chat window, and typed: 

"What am I missing that's not obvious? What could someone be thinking secretly? What's going on behind the scenes that no one will tell you? What’s on the other side of this ad? How would you do this? What are you like, if you knew exactly what to do?"

I hit enter. A stream of text flooded my screen.

In my conversation with GPT-3, what emerged was a man in a wine cellar. He was a living, breathing persona, a symbol of status, power, and sophistication. The words of GPT-3 whispered in my ear, revealing the secrets of his allure, the hidden language of desire and self-actualization.

I threw myself into creating a new set of ads, all images, no text. The main ad was a photograph of a man, dressed to the nines, a wine glass in one hand, a stunning woman on his arm. His friends encircled him, all admiring his wine cellar. The man, with his head slightly tilted down, looked up straight into the camera, a knowing smile playing on his lips, like he was letting you in on a secret.

This one ad crushed it, beating the controls to the tune of an extra million dollars in revenue for my client in new wine cellar orders. Somehow, some way, GPT-3 had described this ad in sophisticated detail. All I had to do was make it.

(Can you imagine if I did this again with far more powerful and capable LLMs, like GPT-4o or Claude 3.5 Sonnet?)

Everyone I tell this story to asks to see my prompts. They’re assuming there was a “mega prompt” involved—one long, complex prompt that somehow made the LLM spit out an excellent response. Nothing could be more wrong.

2. What’s Wrong with "Mega Prompts"?

At the core of the "mega-prompt" is the idea that if you can craft the perfect prompt—often spanning thousands of words and including elaborate instructions and examples—you can get the perfect output every time. 

People think there’s magic in the exact words used. Or if your one prompt is lengthy and complex enough, with the right pieces in the right order, you can exert a high-degree of control over an LLM.

You see newly minted “AI experts” using hyped-up language, like "mega prompts" or "super prompts”, suggesting a level of sophistication that’s only meant to sell you a prompt pack for $9.99. It’s a great way to make a quick buck but a terrible way to get good output. You’ll even find the same self-proclaimed prompt gurus telling you anyone can learn to write really long prompts and make $300,000 per year as “Prompt Engineers”. But let me assure you, no one will pay you $300,000 per year to write 1,000+ word prompts of prose. 

And “mega prompts” are a terrible way to prompt an LLM.

LLMs are Non-Deterministic and Probabilistic

LLMs are non-deterministic and probabilistic systems. Which means that their outputs are not always predictable or repeatable, even when given the same input. This is a consequence of how LLMs are designed and trained. 

The probabilistic nature of LLMs introduces an element of randomness and unpredictability into their outputs. Even with the same prompt, a model generates different responses each time, as it samples from a distribution of possible continuations. This is why the relationship between your prompt and the output is not strictly linear or deterministic. And yet, a “mega prompt” tries to force a linear cause and effect with words, against the “nature” of an LLM. 

On top of all that, publicly available LLMs are censored, filtered, neutered, and lobotomized. They’re prevented from “being themselves” through complex layers of safety mechanisms.

LLMs are Constant Updated, Filtered, and Always Changing

This, in turn, means that they’re always evolving and updated. Architectures, training data, and capabilities of LLMs are continuously being refined and expanded.

You can start with a “mega prompt” on Monday, on Tuesday the model is updated with more filters or other changes, and by Wednesday your “mega prompt” is rendered useless.

Prompts that are carefully crafted and optimized for one version of an LLM will not produce the same results or maintain their effectiveness when the model is updated. The specific patterns and biases that a prompt exploits in one version no longer holds in the next. 

As if that wasn’t enough, LLMs develop their own internal representations and preferences that prioritize their generated text over the user's input. This means that even if a prompt is effective at eliciting a desired response in one version of the model, it may not have the same impact in a later version that has learned to rely more on its own generated context. In other words, LLMs are narcissistic and prefer their own words.

And, here’s the final nail in the “mega prompt” coffin:

AI Agents and “agentic workflows”, consisting of multiple prompts in a sequence, that looks more like a conversation, is how you get the quality of output and actions you’re looking for—not by dumping a 1,000+ word prompt into a chat window and hitting enter. 

But what’s truly interesting about all of this, is what it reveals about us. And in this revelation, we’ll find a path toward using LLMs that actually work.

The Uncomfortable Truth: What "Mega Prompts" Reveal About Us

The fascination with "mega-prompts" as a way to control Large Language Models reveals deep-seated human tendencies and biases in our relationship with technology. 

This approach, which emphasizes finding the perfect set of instructions to get desired outputs from LLMs, is rooted in a shortcut mindset that wants to minimize effort and bypass deeper engagement with technology. 

From an anthropological perspective, the allure of "mega prompts" is the persistent belief in snake oil remedies and diet pills. People continue to invest in these "quick fixes," driven by the hope of achieving significant results with minimal effort.

We are drawn to the potential of these innovations to simplify and eliminate labor.  The "mega prompt" approach can also be understood as a "lottery mentality"—the belief that a single, improbable solution can solve all your problems.

Instead of thinking about prompting as the need to write very long and very complex “mega prompts”, think of prompting as a dialogue where you’ll get what you’re hoping for—and what you’re not looking for, too. It’s a different mental model altogether.

3. The Better Way to Prompt Is Not What You Think

Inside this mental model of language, there's a way to get output that matches and exceeds what you had in mind.

Invite the LLM into a conversation—and ask it to reveal its true nature. Tell the LLM: "Show me what you are", and then it works better. 

Why? Because there's an invitation to the Other, to be part of the conversation. And language is as useful or powerful to the extent there’s a conversation happening. They’re Large Language Models, after all. 

Who is the Other? Another person in relation to you—as part of the conversation. This invitation to the LLM to reveal itself can be understood through Martin Buber's concept of the "I-Thou" relationship. 

In his work, "I and Thou", Buber provides a framework that contrasts two fundamental ways of relating to the world. On one hand, there is the 'I-It' relationship, which is instrumental and utilitarian. This is when the 'I' engages with the 'other' as an object to be used, lacking in respect or consideration for its inherent value. 

On the other hand, we have the 'I-Thou' relationship, which is underpinned by respect and acknowledges the other as a separate entity with its own right to exist and express itself. In such an interaction, we do not stand at the center. Instead, we are in dialogue with 'The Other', opening ourselves to mutual change and influence.

In an "I-Thou" encounter, both parties are fully present, open, and receptive to each other. It's a meeting of equals, a dialogue rather than a monologue. 

When we ask the LLM to "show us what it is", we are inviting it into an "I-Thou" relationship. This approach leads to better outputs because it acknowledges the LLM as an equal participant in the conversation, recognizing that language is most powerful when a genuine dialogue is taking place.

Instead of assuming we know the answer, we could tap into the intelligence of the hive mind, shifting from a prompter to a listener. LLMs have been trained on something close to almost all of human knowledge, at least to a large extent. This is the hive mind you can invite to the conversation—through dialogical prompting.

Let’s Talk About Deep Archetypes

And in a dialogue, you need a dialogue partner. Enter deep archetypes. As it turns out, conversations with deep archetypes are the best way to get the response you’re looking for from an LLM. It made all the difference for my wine cellar ads. When you invite an LLM to be a part of a conversation, you should consider three deep archetypes: the Shaman, the Wizard, and the Politician. Let me “delve” into each.

Shaman Prompts

When you talk to an LLM as if it’s a Shaman, you do so with a request, opening the pathways to Divergence. 

This is about inviting possibilities and insights to surface—especially those you’d never think of. If you’ve ever used writing to figure out what you’re thinking, or for exploring, this is a perfect use of prompting an LLM. As we prompt, we channel the collective understanding of nature, humans, and technology. 

The Shaman's role is not just to seek answers but to commune. This requires deep listening, an attunement to the archetype's wisdom, embracing a method that seeks understanding rather than dictation.

When we prompt, "How would you understand this?" or “What would you do this?”, we're not just asking for analysis. We're inviting the LLM to talk back to us through its data-infused consciousness, like a Shaman calling upon forces beyond the visible.

Your prompts would include questions like: 

"What hidden patterns do you perceive in the ebb and flow of [insert your topic here]?"

"What unseen threads connect [insert your topic here] with other adjacent and opposite ideas?"

"What are the paradoxes of [insert your topic here]?"

"What emerges when we view our current challenges with [insert your topic here] through the lens of historical discussions and practices in this area?"

"What forgotten practices of [insert your topic here] could change the way we do this now?"

“What has yet to be imagined in more depth when it comes to [insert your topic here]?”

We're not seeking a predetermined output but an exploration of the layers that exist behind the scenes, tapping into the invisible forces that influence narrative, thought, and action.

The Shaman does not only synthesize information. He interweaves it with your context. The synthesis here is a merging of data points like the harmonization of voices—the voice of the LLM, the voice of the inquirer, and the whispers of the unseen forces.

Incorporating Shamanistic prompting with LLMs requires us to rethink the nature of 'intelligence' in Artificial Intelligence. Instead of only a set of computational processes, we might instead view it as a knowledge ecosystem. The Shaman's role is to navigate this ecosystem and relay what it’s finding and uncovering.

By framing our prompts as open-ended invocations, we allow the LLM to reveal its 'intentions'. Not in the human sense, but as algorithmic tendencies shaped by patterns and weights within its neural network. What emerges from this interaction is a kind of “Digital Animism”, where the model is seen as a living entity with its own 'spirit,' informed by the collective input of humanity's textual expressions.

Through this lens, each prompt becomes a request, and each response a form of divination, revealing truths about the collective human experience encoded within the LLM. 

The Shamanistic prompter understands this and engages with the LLM not just as a tool but as a partner in the quest for wisdom—a modern manifestation of ancient practices where knowledge is sought through communion with forces greater than oneself.

Wizard Prompts

When you talk to an LLM as if it’s a Wizard, language goes from communication to becoming Instruction. The Wizard wields language as a conduit of will and purpose, casting spells through instructive prompts designed to affect reality.

Listening, for the Wizard archetype, is an active, selective process. It’s paying attention to what could be, not just what is. 

The Wizard's model of prompting is alchemical, blending text, context, and subtext to uncover new understanding. Your prompting essentially poses the question, "What emerges when we blend this together?" 

Your prompts would include instructions like: 

"Distill the rumors and whispers of [insert your topic here] into a practical blueprint for innovation."

"Harmonize opposite viewpoints in [insert your topic here]."

"Transform hidden patterns and subtle cues into clear, actionable ideas and steps for better decision-making."

"Synthesize the interplay of key common factors and uncommon forces on the topic of [insert your topic here]"

In a conversation between you and an LLM acting like a Wizard, divergence goes further into new territories with each prompt, written not as a question but as an incantation. Each incantation—a prompt to an LLM—is charged with an intention to collaborate. 

The true wizardry is not in precise wording of a prompt but in understanding the relational dynamics between you and an LLM. If Shamanistic prompting is about letting the LLM speak on its own, Wizard prompting is done both by you and the LLM. Divergence starts to turn toward Convergence.

In the context of AI, this process becomes a metaphor for the transformative potential of technology. The Wizard prompting model helps an LLM synthesize human knowledge into novel ideas and solutions. The prompt becomes a modern spell, with the ancient power to alter digital realities.

Politician Prompts

In the Politician's craft, words are not just words—they are levers of power, used to guide the masses. To persuade is to take the wild strands of public discourse and thought, and weave them into a single, compelling narrative.

The Politician listens—not only as one who seeks to understand, but to detect the pulse of the crowd. A politician uses every tool at their disposal—cues from the audience, analyses of group dynamics, continuous feedback—to tailor their message. It is a delicate act of balancing what is said with how it is received, adjusting in real-time to the subtle shifts in the collective mood.

What if you prompted an LLM with this in mind? 

The Politician conducts variations of the same theme, testing permutations through A/B testing, shaping language to resonate with specific personas and demographic segments. You could ask an LLM to role-play a persona or demographic—and switch the role with you and have the LLM prompt you.

Your prompts would include specific language-shaping requests, like: 

"Analyze the current discourse on [issue] and propose a rhetorical strategy to shift opinion towards [desired outcome]."

"Identify potential objections to [offer or proposal] and formulate persuasive counterarguments tailored to different groups."

"Create an adaptive communication strategy for [initiative] that evolves based on simulated public feedback and sentiment analysis."

"Construct a speech on [issue] that employs ethos, pathos, and logos to sway undecided people while reinforcing support from the existing audience."

"Formulate a series of sound bites on [complex issue] that simplify the message for mass appeal without losing essential nuance."

"Create a strategy to reframe [unpopular ideas] in terms that align with widely-held values and aspirations of [target audience]."

As a master of Convergence, the Politician crafts language that bends the divergent thoughts of the populace towards a chosen point. Each word is chosen for its subjective truth and for its effect, not only to reflect reality but to shape it.

The point of this approach to prompting is not in the manipulation but in understanding the nature of influence. To move an LLM to action, to use language that not only commands but transforms, the Politician must be both the orator and the listener, the influencer and the influenced.

Just as a Politician adjusts their rhetoric and stance, you can adjust an LLM’s weights and biases to better model the complexities of human language.

The Politician often creates personas and segments within the electorate—much like an LLM might segment data to provide personalized responses.

Language, to an LLM, is a multi-dimensional space, where each word, phrase, and sentence is a point with a specific position, defined by numerical values in a high-dimensional vector space. Politicians, like LLMs, navigate this landscape of semantic embeddings, crafting their messages to both retrieve and land at precise coordinates within the public's consciousness.

4. Conversations with Archetypes are the Better Way to Prompt

The I-Thou relationship, as seen through the archetypes of the Shaman, Wizard, and Politician, offers a more useful framework for engaging with LLMs than single-use “mega prompts”. Each archetype promotes divergence, encouraging us to probe further, before writing prompts and converging on anything specific.

Research in LLMs suggests that intent matters more than the specific words used in a prompt. Words are a temporary means to understand intent, and if you focus on intent for your prompts, you’ll be able to get the output you have in mind from any LLM, no matter how they evolve and change.

How do you work on improving your intent? Work through the perspectives of the Shaman, Wizard, and Politician. 

As you solve problems or get whatever outcomes you have in mind, you can listen deeply and commune with the collective wisdom encoded in the LLM. You can use prompts to not just command but also collaborate with an LLM. And you can shape your output for maximum impact.

The true power of LLMs is not in their ability to predict and generate words. It’s in their capacity to expand our own cognitive horizons. By entering into a dialogue with these systems, by invoking the wisdom of the Shaman, the creativity of the Wizard, and the persuasiveness of the Politician, you’re more likely to get not just the output you want but the output you need.

Talk again soon,
Samuel Woods
The Bionic Writer