Latest / Logosofia no Dia a Dia / Prompt Engineering Is a Misnomer — Managing Ambiguity in AI
Transcript
- [Prompt] custom: If somebody were to ask me how I would explain my philosophy of prompt engineering on one leg, I mig | Hosts: corn, herman [Response] Corn: Daniel sent us this one, and it's a big swing — he's basically arguing that "prompt engineering" is a misnomer, that what he's actually been developing across hundreds of episodes is something closer to the art of managing ambiguity. The term engineering implies precision, predictable outputs, a blueprint. But his best results, the moments he calls the pinnacle of what's fantastic about AI, come when he deliberately leaves room for the model to surprise him. Herman: He's not wrong about the terminology problem. Engineering as a discipline is about deterministic systems — you apply known principles, you get predictable outcomes. That's not what's happening when you prompt a language model. You're navigating a stochastic system where the same input can produce wildly different outputs, and the skill isn't in locking things down, it's in knowing where to be precise and where to let go. Corn: Which is the exact opposite of how most people think about it. The instinct is to over-specify, to close every door, to write the prompt equivalent of a legal contract. And Daniel's saying the magic actually lives in the doors you leave open. Herman: And what makes this timely — he mentioned he's been using voice dictation for his prompts, and that's not just a convenience thing. Voice input preserves prosodic cues, the natural pauses and hesitations and emphasis that typed text strips away. Those signals encode something important about where ambiguity is welcome and where it isn't. As voice interfaces become the default way people interact with AI, the old engineering framing breaks down even faster. Corn: The question sitting underneath all of this is: if it's not engineering, what is it? And can you actually get better at it, or is Daniel just describing his own idiosyncratic workflow and calling it philosophy? Herman: I think it's teachable. He's identified something real — this calibration between directional specificity and productive vagueness. The fact that he's landed on it through trial and error over hundreds of episodes doesn't make it less valid. It makes it field-tested. Corn: Let me push back on that for a second. If it's field-tested across hundreds of episodes, that's also hundreds of episodes with the same person, the same brain, the same conversational tics. How do we know this isn't just Daniel learning to prompt Daniel's own show? Like, he's optimized for a very specific output format. Herman: That's a fair question. But I think the pattern generalizes because the underlying mechanism isn't about Daniel's particular style. It's about the relationship between constraint and creativity in any language model interaction. Whether you're generating podcast dialogue or debugging a Kubernetes cluster, the dynamic is the same — tight constraints produce safe, predictable outputs. Loose constraints produce surprising, occasionally brilliant outputs. The calibration problem exists regardless of domain. Corn: Okay, so it's the calibration that's universal, not the specific prompts. Herman: The prompts are local. The calibration skill is portable. Corn: Let's unpack what he means by managing ambiguity. Because it's not about being lazy with your prompts or just tossing a vague question into the void and hoping for the best. There's a deliberate structure to it. Herman: The structure is counterintuitive. Most people assume that better prompts are more detailed prompts. Daniel's experience suggests the opposite — that after a certain point, each additional constraint doesn't improve the output, it just narrows the possibility space. The skill is knowing where that inflection point sits. Herman: The paradox at the center of this is what gets me. These models are, at their core, literal machines. They process tokens, they compute probabilities, they don't "understand" ambiguity in any human sense. And yet the most impressive outputs come when we deliberately introduce vagueness. It's almost like the system needs some slack in the rope to do its best work. Corn: Which is where the engineering metaphor really falls apart. An engineer tightens tolerances. They remove slack. The whole discipline is about eliminating ambiguity from a system. Herman: And if you approach a language model that way, you get exactly what you asked for — which sounds like a success, but it's actually a ceiling. You've constrained the model so tightly that it can only operate within the bounds of what you already knew to specify. There's no room for it to bring anything you didn't anticipate. Corn: I'm thinking about this in terms of, like, a jazz ensemble versus a marching band. The marching band has every note charted. The jazz ensemble has a head, a key, a tempo — and then everybody gets a solo. The structure is there, but the structure exists to enable improvisation, not to prevent it. Herman: That's exactly the right analogy. And the bandleader doesn't say "play whatever you want for the next four minutes." They say "you've got 32 bars in B-flat, and the horns will lay out." That's directional vagueness. You've specified the container, not the contents. Corn: The art isn't in crafting the perfect instruction set. It's in calibrating how much room you leave for the model to surprise you. And that calibration changes depending on what you're trying to do. Herman: If I'm asking for a VLAN configuration walkthrough, I want precision. I want the model to stay in its lane. But if I follow that with "what am I not seeing here," I've just opened the aperture. Same conversation, same context, but I've shifted from engineering mode to exploration mode. And that shift is where Daniel's "pinnacle" moments tend to happen. Corn: What we're really talking about is a spectrum. On one end, you've got the hyper-specific prompt where ambiguity is a bug. On the other, you've got the open-ended prompt where ambiguity is the feature. And the skill is knowing where on that spectrum to place any given question. Herman: Here's the thing — most people pick a spot on that spectrum and stay there. They're either a specificity person or a vibes person. Daniel's insight, whether he'd frame it this way or not, is that the real power comes from moving fluidly between positions. Tight, loose, tight again. It's a dynamic range, not a static setting. Corn: Which is actually harder than just being good at one or the other. It's like being bilingual versus just speaking one language well. Herman: Voice prompting turns out to be a surprisingly good tool for navigating that spectrum, which is what we should dig into next. Because the mechanism of why it works tells us something important about what's actually happening when we prompt these systems. Herman: To understand why voice prompting works so well for this, you have to look at what happens when you speak versus when you type. When you type a prompt, you're editing as you go. You delete the false starts, you clean up the sentence fragments, you remove the "um"s and the pauses. You're producing a polished artifact. But here's the thing — those messy bits you removed were actually carrying information. Corn: Information about what? Herman: About where you're uncertain. About which parts of the question you're still feeling your way through. When Daniel dictates a prompt, his voice naturally hesitates before a tricky concept, or speeds up through something he's confident about, or pauses before a pivot. The speech-to-text system captures some of that, but more importantly, the rhythm of spoken thought is fundamentally different from the rhythm of typed thought. You think in more complete arcs when you speak. Corn: The voice prompt is less edited by default, and that's actually a feature, not a bug. Herman: A typed prompt tends toward what I'd call sterile specificity. You've sanded off all the texture. But the model uses that texture — the slightly awkward phrasing, the natural digression, the way a spoken question sometimes circles its point before landing on it — as signals about where exploration is welcome. It's like the difference between handing someone a printed itinerary versus describing your trip over coffee. The itinerary is precise, but the conversation leaves room for "oh, you should also check out this place. Corn: Which connects back to Daniel's point about literal understanding. He said AI thrives on literal understanding, and voice somehow counteracts that. I think what he's noticing is that voice introduces enough productive imprecision to break the model out of literal mode. Herman: And this is where the concept of context engineering becomes useful. Because when Daniel sends a voice prompt, he's not just sending words. He's sending that prompt into a system that already has a show format, episode history, character voices, all of that layered context. The voice prompt is one input among many. What he's actually doing is designing an environment where the model has enough structure to stay on the rails and enough slack to find interesting routes. Corn: The boundary between "the prompt" and "the context" is basically artificial. Herman: And once you see that, the term prompt engineering starts feeling almost quaint. You're not engineering a prompt. You're curating a conversational environment. The system prompt sets the stage, the voice input brings the natural ambiguity, the conversation history provides continuity, and the model navigates within that designed space. Corn: Let's make this concrete with the VLAN example Daniel mentioned. He described a hyper-specific prompt about isolating IoT devices on a UniFi network — VLAN ten and twenty, firewall rules, the whole thing. That first prompt is pure engineering. He wants the model locked in, precise, instructional. Herman: It'll deliver that. But then he appends something like "what edge cases am I not considering," and suddenly the model shifts gears. It's no longer operating within the tight boundaries he set. It's looking at the whole configuration from outside, drawing on broader networking knowledge, flagging things like multicast DNS across VLANs or what happens when a device needs to speak to both subnets simultaneously. That follow-up question is where the ambiguity lives, and it's what turns a technical walkthrough into something genuinely insightful. Corn: The first question gets you the answer you could have found in documentation. The second question gets you the answer you didn't know you needed. Herman: That's the tradeoff. Specificity buys you reliability. Openness buys you surprise. If you only ever do one or the other, you're leaving value on the table. The real skill is sequencing them — use precision to establish the domain, then use ambiguity to explore its edges. Corn: Which is why Daniel's "one question" technique works. He's not just being lazy by asking fewer questions. He's forcing himself to identify the single most generative question in the set, the one that creates the largest possibility space, and then trusting the model to fill it intelligently. Herman: Voice makes that sequencing more natural. When you speak a detailed technical setup and then pause and say "...but what am I missing here," that pause is doing real work. It signals a shift in mode. The typed equivalent — two bullet points in a list — doesn't carry the same weight. The model doesn't know that the second question is meant to open things up unless you explicitly tell it. Voice encodes that shift in the delivery itself. Corn: I want to pause on that because I think it's easy to hear "voice is better" and file it under personal preference. But you're making a stronger claim. You're saying the medium itself carries semantic information that the model can use. Herman: And there's actually some research adjacent to this — not on AI prompting specifically, but on how humans process spoken versus written communication. When you hear someone speak, your brain is processing prosody, pacing, volume shifts, all of these paralinguistic features, in parallel with the words themselves. A flat "that's interesting" means something totally different from "that's... interesting" with a pause. The words are identical. The meaning isn't. Corn: When Daniel dictates a prompt and his voice trails off before a difficult concept, the transcribed text might just show an ellipsis or a dash, but the model — especially as these systems get more sophisticated about processing audio directly — is potentially picking up on the fact that this is a zone of uncertainty. This is where the speaker is inviting help. Herman: Even if the current speech-to-text pipeline flattens some of that out, the cognitive effect on the person dictating is real. You think differently when you speak. You're less likely to produce the over-engineered, bullet-pointed, every-loophole-closed prompt that typing encourages. The medium shapes the message before it ever reaches the model. Corn: Voice gives us this natural ambiguity signal. But what happens when we deliberately design for it? When you stop treating ambiguity as something to minimize and start treating it as a parameter you can tune? Herman: That's the shift Daniel's making, whether he'd use those words or not. He's not accidentally being vague. He's calibrating. Corn: The calibration has a name, though I don't think it's caught on yet. I'd call it directional vagueness. You're specific enough to set a heading — "discuss VLAN isolation in home networks" — but vague enough that the model chooses the route to get there. That's not being lazy. It's not the same as saying "tell me about networking" and hoping for the best. You've given a destination, you've just left the navigation open. Herman: Which is a harder skill than it sounds. Most people, when they try to be "open-ended," just get vague in a way that produces generic responses. The art is being precise about where you want precision and loose about where you want surprise. It's two different modes in the same prompt. Corn: Can we talk about what happens when people get this wrong? Because I think the failure modes are instructive. Herman: The most common failure pattern is what I'd call the vague-open prompt. "Tell me something interesting about networking." That's not directional vagueness. That's just vagueness. You haven't set a heading at all. The model doesn't know whether you want history, or troubleshooting, or career advice, or protocol deep-dives. So it picks the safest, most generic path, and you get something that reads like a Wikipedia introduction. Corn: Then the person concludes that open-ended prompting doesn't work, when really they just didn't provide a direction. It's like handing someone a compass with no bearing and being disappointed they didn't reach a destination. Herman: The opposite failure pattern is the over-constrained prompt that accidentally closes the door on the interesting stuff. I saw an example of this recently — someone asked a model to "explain the three main causes of the 2008 financial crisis using only factors related to mortgage-backed securities." That's a perfectly reasonable question, but by pre-specifying both the number of causes and the domain, they guaranteed they'd get a answer that excluded things like credit rating agencies, regulatory failure, systemic risk models. The prompt was so tight it excluded the most important parts of the story. Corn: The failure pattern of over-specification isn't that you get a wrong answer. It's that you get a right answer to a question that's too small. Herman: And that's the insidious thing about it. The answer looks correct. It cites sources. It's well-structured. You walk away satisfied. You never realize the model had something much more interesting to say, because you never gave it the chance. Corn: Here's where the comparison to other creative domains gets interesting. This is exactly what a good editor does with a writer. You don't hand someone a sentence-by-sentence outline and say "fill in the words." That produces stiff, lifeless prose. But you also don't say "write something about technology" and walk away. You give enough direction to avoid aimlessness, enough freedom for the writer's voice to emerge. The AI, in this framing, stops being a tool and starts being a collaborator. Herman: The editor analogy lands. A bad editor over-specifies and gets exactly the article they already had in their head. A good editor says "here's the territory, show me what's interesting in it. Corn: Daniel's "one question" technique is the prompt equivalent of that. When you stack three or four specific questions into a single prompt, you're effectively handing the model a checklist. It'll work through each item dutifully. But when you ask one high-level question, you're forcing the model to prioritize, to synthesize, to decide what's actually worth surfacing. You're giving it the editorial discretion. Herman: Which is terrifying to most people. The fear is that if you don't specify everything, the model will wander off into irrelevance. But Daniel's experience suggests the opposite — that over-specifying is what produces the irrelevant detail, because the model is just ticking boxes instead of thinking about what matters. Corn: There's a fun fact buried in here, actually. Did you know that early text adventure games — Zork, Colossal Cave — had this exact problem? If you typed "take inventory" you got a list. If you typed "look" you got a description. But the games that people remember most fondly were the ones where the parser could handle slightly ambiguous inputs. "Get the thing" instead of "take the brass lantern." The ambiguity made it feel more like a world and less like a database query. Herman: That's a great pull. And it's the same principle. The more you have to speak the system's language precisely, the more you're aware you're interacting with a system. The more the system can handle your natural ambiguity, the more it feels like collaboration. Corn: Let's make this concrete with a comparison. Take two prompts about the same home network. First one: "List three specific vulnerabilities in my VLAN configuration." Second one: "What are the security implications of my home network setup?" The first prompt will give you exactly what you asked for — three vulnerabilities, probably accurate, probably useful. It's a checklist. The second prompt is broader. It might talk about VLANs, but it might also surface something about your DNS configuration, or the fact that your IoT devices are phoning home to servers in countries with different privacy laws, or the attack surface created by your guest network. Things you didn't know to ask about. Corn: The checklist is useful. Nobody's saying it isn't. But only the second prompt produces what Daniel calls the pinnacle experience — the moment where the model operates beyond your bounded worldview. Herman: This principle shows up across creative domains. In image generation, Midjourney has a stylize parameter that essentially controls how much the model is allowed to deviate from your literal description. Crank it low, you get exactly what you described. Crank it high, you get interpretations you wouldn't have imagined. Neither setting is "correct" — the skill is knowing which one serves your current goal. Corn: Same thing with code generation. If you write out a complete function signature with every parameter type-annotated and every edge case specified, the model will fill in the implementation you expected. Herman: In all these cases, the mechanism is the same. You're deliberately withholding some specification to create a space where the system's own intelligence can operate. It's not about being lazy or underspecified. It's about recognizing that your own knowledge has boundaries, and the model's value is partly in what lies beyond those boundaries. Corn: Which brings us back to Daniel's core insight. He's not just describing a personal quirk. He's identified a teachable skill — the ability to calibrate ambiguity, to know when to tighten and when to loosen, to treat the model as something that needs both rails and room. That's not engineering. But it's not mysticism either. It's a practice. Herman: How do you actually practice this? Because I think that's what most people listening want — not just the philosophy, but something they can try tomorrow morning. Corn: Let's start with the simplest one, and it's the one Daniel stumbled into through sheer volume. Before you send a prompt, reduce it to a single question. Not three questions, not a question with sub-parts. One core question. Herman: If you've got multiple things you want to ask, which happens constantly, you prioritize the most open-ended one and save the rest for follow-ups. The instinct is to front-load everything into one massive prompt because it feels efficient. But you're actually just giving the model a to-do list instead of a thinking problem. Corn: The follow-ups are where the real conversation happens anyway. You ask the broad question, the model surfaces something unexpected, and your next question is better than anything you would have pre-planned. Herman: Second technique, and this one's easy to overlook: use voice dictation even if you plan to edit afterward. The act of speaking forces you into complete thoughts. You can't half-finish a sentence and leave it hanging the way you can in a text box. And those natural hesitations, the places where your voice slows down or pauses, they're marking exactly where ambiguity lives in your own thinking. Corn: You don't even need to send the voice version. Just dictate it, look at what came out, and notice where the messiness clustered. Those are your open questions. Those are the doors you might want to leave unlocked. Herman: The third one is a diagnostic. After you get a response, ask yourself: did the model surprise me? If the answer is no, your prompt was too constrained. You didn't leave any room. So loosen one thing — remove a constraint, broaden the scope, or just append "what am I missing" and run it again. Corn: That last one is almost a cheat code. "What am I missing" is the smallest possible edit with the largest possible return. Herman: I'd add a fourth technique, and this one's a bit more advanced. Once you've gotten comfortable with the single-question approach, try what I'd call the constraint gradient. Start with a prompt that's deliberately too open — you'll get something interesting but unfocused. Then add one constraint. Run it again. Run it again. Watch how the output changes at each level of specificity. You're essentially building an intuition for where the inflection point lives for different types of questions. Corn: That's like ear training for prompt calibration. You're learning to hear the difference between "too loose" and "just right" and "too tight. Herman: And over time, you stop needing to run the gradient. You develop a feel for it. You know that a technical how-to question can handle more constraints than a strategic brainstorming question. You know that a creative writing prompt needs more slack than a data extraction prompt. It becomes intuitive. Corn: Which raises the question that's been sitting underneath this whole conversation. As these models get better at handling context — longer windows, better retrieval, more sophisticated memory — does the need for deliberate ambiguity go up or down? Herman: The intuitive answer is down. If the model can hold more of your intent in its working memory, you'd think you could be more specific, close more doors, nail everything down. Corn: I think that's exactly backwards. The more the model can see, the more you need to guide it without overdetermining it. A model with a million-token context window doesn't need you to spell everything out. It needs you to tell it which parts of that vast possibility space are relevant and which parts you want it to explore freely. Ambiguity becomes more important, not less. Herman: Because the alternative is drowning in precision. You specify everything, the model dutifully stays within your bounds, and you get exactly what you already knew — just more of it. The value of the system is in what it can reach that you can't. Corn: It's the difference between using a telescope and building a bigger fence. A bigger context window is a wider aperture, not a reason to be more controlling about what you point it at. Herman: The skill Daniel's describing isn't going to become obsolete. It's going to become central. Managing ambiguity is what you do when the model is capable enough that you don't need to micromanage it. It's the skill of knowing where your own knowledge ends and trusting the system to operate beyond that boundary. Corn: That's the thing to sit with. The art of managing ambiguity isn't about being vague. It's about being precise about where you want precision and where you want surprise. That's a real skill. It can be practiced. And it's probably the difference between using these systems and actually collaborating with them. Herman: Now: Hilbert's daily fun fact. Hilbert: The traditional Kiribati fisherman's knot known as the te karewe lashing uses coconut husk fiber treated with lime from crushed coral, which causes a chemical reaction that tightens the fibers as they dry, creating a bond that actually strengthens in saltwater. Corn: I have no idea what to do with that information. Herman: Strengthens in saltwater. I'll file that under things I hope I never need. Corn: This has been My Weird Prompts. Thanks to our producer Hilbert Flumingtop for making this show possible. If you enjoyed this episode, tell someone who's still writing prompts like legal contracts. We're at my weird prompts dot com. Herman: Until next time.