The Conversational Advantage (To LLM Prompting)

Today's Focus

I've been noticing something weird lately in how I talk to AI.

When I carefully craft those super-precise prompts with all the right keywords and structure - you know, the ones that look like a legal document with bullet points and specific instructions - I get exactly what I asked for. Nothing more. Nothing less. Just... exactly what I requested.

But when I just... talk?

When I let my thoughts wander and meander and just share what's bubbling up in my brain without trying to format it perfectly? Something completely different happens. The responses get richer. More surprising. Often way more insightful than what I was expecting.

Makes me wonder - is all this "prompt engineering" stuff missing the point? What if the best way to talk to AI systems is actually just... talking? Not formal templates with rigid structure, but genuine conversation that follows wherever my thoughts naturally go?

I couldn't stop thinking about this. Had to dig deeper.

Learning Journey

It reminds me of that scene in Shawshank Redemption where Andy plays the opera music over the prison speakers. Everyone just stops and listens. That moment when structure breaks and something authentic flows through? That's what I'm after in AI conversations.

I started by looking into the difference between conversational and command-based approaches to AI interaction. Research shows there's a fundamental tension between these styles. While explicit commands can provide clear direction, they can also disrupt the natural flow of conversation. Meanwhile, more natural conversational approaches foster stronger connections between users and AI systems [1].

This aligns with my own experience. When I need specific technical outputs like debugging code, structured prompts work best. But for exploring ideas or developing new insights, a more flowing conversation yields better results.

This led me to explore three key aspects of AI conversation that seem to fundamentally shape the interaction experience.

The Conversation Garden vs. The Command Terminal

These two metaphors keep bouncing around in my head:
Command Terminal: You type precise instructions, carefully structured, designed to elicit specific outputs. Like coding. Like giving orders to a machine.
Conversation Garden: You just talk naturally. Let ideas grow organically. Follow curiosity paths wherever they lead.

The difference isn't just about style - it fundamentally changes what the AI gives back to you.

Research shows different interaction approaches actually activate different AI capabilities [1]. When financial advisors used conversational agents, they naturally switched between conversation mode and command mode depending on what they were trying to accomplish.

Like how we instinctively change how we talk to different people in different situations. Studies comparing click-based and writing-based chatbots found that the conversational approach created stronger bonds between users and the system [1]. There's something about conversational flow that creates connection that structured commands just don't achieve.

I've felt this myself. When I'm having flowing, natural conversations with AI, I often get these unexpected connections between ideas. Things I wouldn't have thought to ask for directly. The garden metaphor feels right because ideas seem to grow and cross-pollinate in ways they never do when I'm just issuing commands.

It's like the difference between asking someone to solve a math problem versus asking what the math problem makes them think about. Our brains don't actually think in neat little categories with transition phrases between them. They JUMP. They create unexpected neural bridges. They follow fascination paths rather than logical progression.

Thinking With vs. Thinking At

This distinction keeps nagging at me - it feels important:
"Thinking At" positions the AI as this passive tool that just follows instructions to produce outputs. Like talking to a vending machine. Input command, receive output.
"Thinking With" creates this collaborative space where the AI becomes an actual thinking partner. Where thoughts bounce back and forth and build on each other.

Cognitive science research backs this up. Conversation serves as this powerful mechanism for externalizing thought processes [4]. We talk through ideas, get feedback, refine our understanding in real-time. It's called "repeated explanation support" in some research papers. It enhances collaborative learning [4].

The weird thing I've noticed? When I approach AI with this "thinking with" mindset – not expecting one definitive answer but exploring ideas together – I actually encounter fewer of those "hallucinations" people worry about. Makes sense when you think about it. When not forced to produce a single authoritative answer to questions that might have nuanced or contextual responses, the AI can share what it knows while acknowledging limitations.

It's like the difference between demanding someone give you THE answer versus exploring a question together. Do you ever notice how conversations with friends can lead to insights neither of you would have had alone?

That's what I'm after.

Recent research on "augmented intelligence" explores how teams of humans and machines with complementary skills can collaborate productively [4]. The "thinking with" approach aligns perfectly with this idea of augmented intelligence.

Our minds are multi-channel processors running simultaneously:
Visual patterns connecting seemingly unrelated concepts
Emotional responses coloring logical analysis
Memory associations triggering unexpected insights
Sensory perceptions influencing abstract thinking

You can't capture that in ordered bullet points. It's messy. Asymmetrical. Governed by interest rather than importance.But that messiness creates electricity.

The Context Dance

Just like human relationships deepen through accumulated shared experiences, AI conversations become remarkably richer through developed context. It reminds me of those long-term friendships where you can communicate complex ideas with just a glance or reference to a shared experience.

"Remember that thing with the thing at the place?" And somehow they know exactly what you mean.

Research shows expert teams coordinate with minimal explicit communication because they have shared mental models [5]. This applies to human-AI teams too - shared understanding enables better anticipation of needs and responses.

I've experienced this firsthand by developing detailed context documents about my communication style, learning preferences, and thinking patterns. By investing in this "context dance" – building a foundation of shared understanding – subsequent conversations align so much better with how my brain naturally works.

Studies confirm that when AI systems maintain contextual awareness, user satisfaction jumps significantly [5]. Human-like chatbots that demonstrate understanding of shared context create more positive experiences compared to interfaces that reset context each time.

This context-building doesn't just make conversations more pleasant – it enables truly collaborative work. AI systems that model anticipated behaviors of human collaborators and account for shared objectives coordinate more effectively [5]. Even with incomplete information, probability-based approaches allow AI agents to exchange partial mutual knowledge, creating a foundation for context-aware collaboration.

The difference is dramatic. Like talking to someone who's known you for years versus explaining yourself to a stranger. It's like that Star Trek episode "Darmok" where the aliens could only communicate through cultural references. "Shaka, when the walls fell" meant failure, but it wasn't just a word for failure - it carried the entire emotional and contextual resonance of their cultural story about Shaka.

That's how real thinking works - not through logical containers but through resonant pattern recognition.

My Take

I've started seeing prompt engineering less as a technical skill and more as a relational art.

For technical tasks with clear objectives – debugging code, generating specific formats, creating structured data – the "Command Terminal" approach works great. These situations benefit from precision and structure.

But for exploring ideas, developing insights, or working through complex problems, the "Conversation Garden" approach yields much richer results.

By engaging in natural dialogue, sharing my thinking process as it unfolds, I create space for unexpected connections and novel insights.

The most powerful approach combines these methods with intentional context development.

By investing in detailed context about how I think, process information, and communicate, I've created a foundation for more meaningful collaboration. This "Context Dance" transforms AI from a generic tool into a thinking partner that complements my natural cognitive style.

What fascinates me most is how this mirrors human relationships!

Just as we communicate differently with close friends versus service providers, knowing when to use conversational exploration versus structured commands creates more effective AI interactions.

And just as relationships deepen through shared experiences, AI collaborations improve with invested context.

Makes me wonder how this could change AI education.

Maybe instead of teaching people rigid prompt templates, we should be teaching conversational fluency and context development – the art of talking with AI rather than commanding it.

I keep coming back to this image of two people walking through a garden together, pointing out interesting plants to each other, rather than one person issuing commands from behind a control panel.

That's the difference. And it changes everything.

The weird truth is that we've all been trained to think "balanced development" equals "good writing" when actual thinking is wildly imbalanced by nature! It's like we've convinced ourselves that artificial symmetry is somehow more authentic than natural asymmetry.

When it comes to AI interaction, maybe the messiness is the point? Maybe the jagged edges and asymmetry and unexpected jumps are exactly what make the conversation feel alive instead of performed?

Success isn't measured by how complete or balanced or organized the content appears.

Success is when it feels like eavesdropping on someone's internal dialogue rather than reading their formal presentation.

Like the difference between listening to someone think through a problem in real time versus hearing their rehearsed explanation afterward.

One feels alive. The other feels performed.

It's working for me. Could work for you too.

Resource of the Day

I was just looking at Harvard University's guide on "Getting started with prompts for text-based Generative AI tools" and noticed something fascinating hiding in plain sight.

On the surface, Harvard's guide seems to be all about formal, structured prompting. They talk about being specific, clear, and concise. They suggest adding descriptive details and context to get better results.
But here's the interesting part - look at what they actually recommend in practice:

"Chat with the AI as if it's a colleague or teammate and you're working on a project together. Give feedback – tell it which parts of the output were useful and which parts could be improved."

I did not read this before putting together today's post, but it certainly sums up a good deal of the philosophy from this discussion in such a simple phrase.

⚙️ Behind the Cognitive Curtain

Here's a glimpse into our conversation while creating this exploration on prompt engineering as conversation.

Claude: I'm intrigued by your observation about "just talking" versus structured prompting. There seems to be something powerful about that natural approach that many people overlook when they focus on rigid prompt techniques.

Matt: I'm not trying to over engineer my statements for the most part. Once in awhile and very technical projects I go back and rework a prompt a few times to get a better technical response. But most of what I do with especially making these blog posts is purely conversational exploratory. It's like I'm talking to myself out loud but putting it in here.

Claude: That image of "talking to myself but putting it in here" really captures something important about effective AI interaction. It's almost like you're creating a thinking space that happens to include an AI rather than crafting formal instructions.

Matt: When I try to overstructure my requests when I get back is a very structured response. And in some cases that's exactly what we need but in the case of non-exacting answers it makes more sense to me to just talk.

Claude: You've hit on a crucial distinction there - matching the communication style to the desired outcome. I noticed that your perspective on hallucinations was particularly interesting too.

Matt: It seems like hallucinations happen when being asked to give a black and white answer to a problem that doesn't necessarily have a clear answer. By allowing the exploration and the nuance, it's giving the system freedom to share what it knows to be true without being forced to answer a question that might not be answerable with the information it has.

Claude: That's a fascinating insight that connects directly to our "Thinking With vs. Thinking At" section. When the system isn't positioned as an authority required to produce definitive answers but rather as a thinking partner in exploration, it can be more honest about the boundaries of its knowledge.

Matt: For me communication style and understanding how I personally process information, learn, and think, has been key to collaboration. Thats why building personalized communication context documents, and why I took so long to develop them in detail, is vital and included in any project I work on with AI LLM partners.

Claude: What struck me was how you described them as a "personalized primer" - that's exactly what good context does. It's like the difference between talking to someone who's just met you versus someone who understands your thought patterns. Your approach of separate interaction modes for technical questions versus exploratory ones also perfectly illustrates the "Conversation Garden vs. Command Terminal" distinction we explored.

References

1. Semanticscholar.org. (2023). "Talking to Multi-Party Conversational Agents in Advisory Services: Command-based vs. Conversational Interactions". This study highlights the fundamental dilemma in CA design where explicit commands can disrupt natural conversation flow while implicit requests might lead to inadequate CA behavior.

2. Intercom.com. (2024). "Chatbot vs Conversational AI: Differences Explained". This resource explains how "rule-based chatbots follow rigid 'if-then' conversational logic, while AI chatbots use machine learning to create more free-flowing, natural dialogues with each user."

3. Semanticscholar.org. (2023). "Clicking vs. Writing"—The Impact of a Chatbot's Interaction Method on the Working Alliance in AI-based Coaching". This comparative study found that writing-based (conversational) platforms showed higher values for bonding between users and the system, while click-based systems provided better structure.

4. The Dynamic Jigsaw. (2021). "Repeated Explanation Support for Collaborative Learning of Cognitive Science". This research examines how conversation serves as a powerful mechanism for externalizing thought processes through what researchers call "repeated explanation support."

5. Tandfonline.com. (2022). "The role of shared mental models in human-AI teams: a theoretical review". This research investigates how expert teams demonstrate effectiveness through shared goals and mental models that allow coordination with minimal explicit communication.

6. Memory Alpha. (2024). "Darmok (episode)". This Star Trek: The Next Generation episode features the Tamarian species who communicate entirely through allegorical references to their mythology, demonstrating how communication requires shared context.