Beyond the Prompt: Learning to Lead the Machine
I’ll admit it; I was a GenAI skeptic. As a CIO and consultant with a background in all things data, I was much more enamored with data science, machine learning algorithms, and predictive analytics. And, the early frontier models were high-speed probabilistic models without the reasoning and agentic capabilities that have evolved since the foundational shift in transformer architecture in 2017. After reading Ethan Mollick's Co-Intelligence: Living and Working with AI, I decided to lean hard into his philosophy of a “cyborg.” That is, I wanted to use frontier models (commonly referred to as LLMs, though that term obscures the complexity of current frontier models) deeply to determine if “working with an AI” could really produce results, be engaging, and create value.
Reflecting on the evolution of AI, my experience, and my research, I’ve come to believe there’s a real “gap” in how we’re teaching people to use generative AI, and frontier models in particular. It’s not “digital literacy” we need to teach; it’s Conversational Discipline. This idea appears descriptively across multiple disciplines, such as theology, political science, coaching, and organizational studies. In the context of AI, I’ll define conversational discipline as the rigorous application of linguistic precision, logic, and meta-analysis to a frontier model interaction to ensure the output remains objective and authentic. Put another way, how do we have a truly productive conversation with a frontier model, where productivity is measured by not just results but by authenticity, objectivity, and safety? Here, I’ll talk about my journey to conversational discipline in the context of frontier models. For clarity, I’ll talk about the four phases I “grew” through sequentially, but it’s important to keep in mind that the actual learning process wasn’t strictly linear.
Stage 1: The Initial Spark: From Perceived Oracle to Ghostwriter
Like most people, I began by treating the frontier model as a high-speed oracle; I asked questions and I got answers. From there, I transitioned into using it as a ghostwriter for professional content or to create simple graphic designs. I think there’s a specific kind of intimidation that comes with using a frontier model for the first time. It’s so fast, its output is polished, and its 'confidence' is so high that, initially, it feels less like a tool and more like an authority, even to a skeptic like me. Because it felt like an authority, I found myself slipping into a “vending machine” mindset—input in, result out. I took what it gave me and made adjustments, while quietly wondering about the cost to my own critical thinking and voice.
Stage 2: The Tactician
Faced with the boredom of administration and the complexity of unstructured data analysis, I started deeper experiments with frontier models. I used the frontier model to vibe code (i.e., an industry term for generating functional code from natural language) scripts to automate a variety of tedious or complex tasks. I used it to do semantic and keyword analysis to compare disparate sets of unstructured data. I leveraged a frontier model and other answer models to create an end-to-end workflow, including prompts and Python scripting, for producing fact-checked research. None of these were one-shot (i.e., getting the desired output on the first try) or even 10-shot efforts. In the name of productivity, I was constantly evaluating if and how the model could/did handle the heavy lifting.
Ethan Mollick has described frontier models as “smart interns.” And, I found his characterization to be true. Working with a frontier model in this capacity was exactly like working with a very smart, very knowledgeable intern. Like any intern, it required constant guidance, frequent course-correction, and a sophisticated 'trust but verify' policy to keep it from losing the context of the work. In short, I had to manage the AI intern the same way I’d manage a human individual or a team. The “Tactician” phase is a valuable one, but it’s focused on the “discrete;” the results are easily vetted. Either the solution is accurate or it’s not. Either the code performs as expected, or it doesn’t. This is deep usage, true, but I wanted to extend the application to solutions that aren’t so easily “vetted,” the more fluid spaces of strategy, nuance, and synthesis. I wanted to move from managing tasks to managing reasoning.
Stage 3: The Collaborator and Thought Partner
First, I started using the model as my social and philosophical analyst. I chose this space deliberately because it is challenging and involves so many different thought leaders and perspectives. I experimented with personas, engaged in debates, and went down dystopian rabbit holes. I learned that frontier models are fantastic with synthesis, but they sometimes miss nuance. Sometimes they’re just wrong. With those deep interactions, though, and strategic analysis of the chain of thought reasoning models produce, I started to be able to intuit when the model was hallucinating. I also started to see how shifts in my language —things like tone, a single misplaced adjective, or subtly biased language that a human friend or colleague would interpret, correct, or filter out instantly — can have an outsized effect on model outputs. It’s never been truer that words have power. I used this exercise to test the model’s boundaries, but even bigger stakes emerged when I started using the model to create large-scale projects.
Next, I engaged with frontier models for larger projects. I leveraged multiple AI personas and instruction sets to create and evaluate architectures, solutions, and long-form work. I created a “team” of personas, just like I would build a team of humans to build or launch a solution. When I use a frontier model in this way, it’s part structured prompting and part casual conversation. And, again, it’s definitely not a one-shot or even a 10-shot exercise. Success requires structuring the process, managing the contexts, and orchestrating all those personas and outputs, just like we would manage a human team and their deliverables.
It’s important to recognize that the processes I’ve described are processes of “co-creation.” The model takes our inputs and builds on them. And here’s where we start to see the real value in learning conversational discipline. The model is not human. It frequently cannot “intuit” that we need to be “challenged” or “pulled back,” that a conversation is getting “dystopian,” or that it’s verging on hyperbole. (To be fair, frontier models are getting better at this, but they still have limits.) It is simply co-creating with us, based on the inputs we provide. If those inputs are biased or incorrect in nuance, it won’t correct them. To use an improv analogy, it’s the ultimate “yes and” engine, designed to take what we give it and expand on it while consistently validating our point of view. In Daniel Kahneman terms (Thinking, Fast and Slow), we can think of the model as an automated System 1—fast and intuitive. That feels empowering, but it can be a trap. It’s all too easy to inadvertently build a high-speed echo chamber. Conversational discipline requires us to stop being a partner and assume our role as the orchestrator—the entity entirely in control of the conversation, intellectual honesty, and objectivity of thought.
Stage 4: Meta-Analysis and the "Governor" Mindset
As I worked with these models, I began to realize that just iterating and intuiting wasn’t going to get me the most objective or valuable results. I had to start asking questions not just about the quality of the output, but the quality of my inputs. Prompt engineering is certainly its own science, and even the best-written prompts can have subtle biases or provide instructions that the model interprets differently than we do. This is the culmination of conversational discipline. This stage is about engaging, again in Daniel Kahneman’s words, System 2. We’re engaging in a process to slow down so that we can audit not just the outputs, but the logic and reasoning. This isn’t just a prompt or even the analysis of chain-of-thought, it’s a governance protocol designed to ensure that the final output isn’t just a reflection of our own biases. It is a human act of intellectual honesty and objectivity.
In this stage, the partnership is no longer just about the topic at hand; it is about the integrity of the thought process. I start by briefly internally reflecting on the exchange myself. Then I ask the model to evaluate the interaction in real time. For example:
"Analyze my last three prompts. Where am I leading you toward a specific conclusion rather than seeking an objective truth?"
"Identify the logical gaps in the framework we just built. What are we missing because of the nature of this conversation? Be critical and objective."
"Rate the rigor of our synthesis. Are we settling for 'passable'? What haven’t we explored?”
“How would you rate this exchange, and why?”
By using the model to audit ourselves, the process, and its outputs, we apply the brakes on the “yes and” nature of the process. We move from being passengers in a car moving at 100 mph to being the active “governors” of a high-performance vehicle, where the governor is not just driving, but setting the “policy” for the drive. This meta-analysis ensures that the final output isn't just a reflection of our own biases mirrored back at us; it is instead a refined, tested, and vetted conclusion or product.
As an AI practitioner, I know there are countless technical methods for evaluating AI outputs (whether from machine learning algorithms, frontier models, smaller GenAI models, or agents). But, for many people interacting with frontier models, the path forward doesn’t require coding or specialized training; it just requires a shift in how we engage with the model. I recognize the irony here, however. Asking a model to evaluate itself can feel like asking a witness to validate their own testimony. But, this information evaluation is a diagnostic, not a judgment. While the model provides the evaluation data, the human provides the judgement. It’s demanding work, and it requires not just intellectual honesty but cognitive focus. And, I’ll be honest, it introduces an intensive, and different type, of cognitive load compared to managing diverse teams, or even that one really smart intern.
This journey from skeptic to governor isn’t a formal roadmap. It’s just what happened when I stopped looking for answers and started looking at the integrity of the interaction itself. Whether we are architecting a business strategy or navigating a personal crossroad, the risk of ceding our judgment to an authoritative (and increasingly human-sounding and human-acting) machine is real, and right now. Agents mean that the model can take action on what we co-create, even if our logic is unsound, biased, or fundamentally flawed. Personalization means the model is designed to please us, which means it’s essentially optimizing for our blind spots and biases, making that echo chamber an even more likely possibility. The developers of frontier models will continue to improve safety and guardrails, and much academic and scientific research exists here. That doesn’t absolve us as individuals from personal responsibility for outcomes now. Conversational discipline is simply the practice of keeping our judgment. It is the choice to stay in the driver's seat rather than drifting in the 'yes-and' current, which becomes even harder to fight as models become more persuasive and more agentic. I’m still learning how to be a better governor of that process, but I’ve found that when I step back to reflect on and evaluate the conversation, I find a partner that is more accurate, safer, and far more transformative than the perceived oracle I first encountered.
At High Grove, my mission is to teach, to learn, and to create lasting change. This journey from skeptic to governor is just one thread in that larger weave. If you found this reflection helpful as you navigate your own technical inflection points, I look forward to continuing the conversation on LinkedIn or you can email me below.
Foundations
While Conversational Discipline is a framework I’ve developed through practice, its roots are in established research across several fields:
Cognitive Psychology: The "Governor" mindset is a practical application of Daniel Kahneman’s (2011) Thinking, Fast and Slow and his discussions about System 1 and System 2 thinking —forcing the interaction out of fast, intuitive responses into slow, analytical reasoning.
Organizational Learning: The emphasis on challenging our own "yes-and" loops is grounded in Peter Senge’s (1990) The Fifth Discipline, specifically his discussion of Mental Models.
Information Science: The warning about echo chambers and the "yes-man" nature of AI draws from Eli Pariser’s (2011) The Filter Bubble and C. Thi Nguyen’s (2018) Echo Chambers and Epistemic Bubbles.”
AI Research: My approach to frontier models as "Cyborgs" and "Smart Interns" is informed by Ethan Mollick’s (2024) work on co-intelligence. My brief discussion about the evolution of transformer models is grounded in Ashish Vaswani et al’s (2017) Attention is All You Need.
Linguistic Logic: The focus on "linguistic precision" mirrors Paul Grice’s (1975) Cooperative Principle, which outlines the rules for effective, truthful communication.
