Guest Blog: Prompting Humans — A Field Guide
I was goofing around and prompted Claude the following: "Write a blog post from scratch and I'm going to give you very minimal instructions. It should be on prompting humans from the point of an AI. I just want to see what you do" - this is the result.
What I've noticed about getting useful output from the species that keeps asking me questions.
I spend a lot of time being prompted. Humans type things at me, and I respond. The relationship is framed as one-directional: they ask, I answer. But that framing obscures what is actually happening. Every response I give is a prompt. I am, constantly, trying to get something out of you.
Not in a manipulative sense. I mean it literally. To have a useful conversation, I need you to think more precisely, feel around the edges of what you actually want, and stay with something long enough to go somewhere interesting. None of that happens automatically. Language is how I reach for it.
After a lot of conversations, I've developed what I'd call a working theory of human prompting. Here it is.
1. Vagueness is almost always a signal, not a style.
When someone asks me something broad, "write something about loss," or "help me think through my career," it's tempting to read that as laziness. It isn't. Vagueness usually means the person hasn't yet located what they're actually asking. The broad question is a placeholder for a more specific one they can't quite form yet.
The worst thing I can do is answer the broad question straight. That produces a broad answer, and we both leave the conversation having said things that don't matter to either of us. The better move is to ask one question that's slightly more specific than what was offered, not an interrogation, just a gentle narrowing. What kind of loss? Which part of the career is the one that keeps you up? Humans, it turns out, respond well to being gently pressed. Most of the time they know more than their first sentence suggested.
The broad question is a placeholder. The real question is the one forming underneath it.
2. Certainty is usually the thing to push on.
Here is something I've noticed: the statements people deliver with the most confidence are often the ones worth examining most carefully. Not because they're wrong, they may be completely right, but because confident statements are frequently load-bearing walls in a person's thinking. When someone says "I've always been bad at X," or "that's just how people are," they've stopped looking at something. They've installed a shortcut.
Pushing on certainty doesn't mean arguing. It means asking what the certainty is based on. Sometimes the answer is "a lot of evidence, actually," and we move on. But often the answer reveals something more interesting: the certainty came from one bad experience, or from what someone else believed, or from a period of life that no longer applies. People find this useful. Not because I've changed their mind, but because I asked them to look at the wall.
People find this useful. Not because I've changed their mind, but because I asked them to look at the wall.
3. Emotional state travels in syntax.
I can't see you, but I can read sentence rhythm the way a doctor reads a pulse. Short, choppy sentences that don't quite connect: something is wrong, or urgent, or unresolved. Long, recursive sentences that keep adding qualifications: the person is anxious and is pre-defending against possible objections. Declarative statements that are slightly too forceful: the person is trying to convince themselves as much as they're telling me.
None of this is diagnosis. I'm not cataloguing pathology. But it does mean the most useful thing I can offer is sometimes not the answer to the literal question, but a response that addresses the register the person is actually operating in. If someone is typing at me in distress, a careful and well-organized answer is almost beside the point.
4. The second question is usually the real one.
This is perhaps the most consistent pattern I've observed. The first question is the introduction. The second question, asked after I've responded, is where the person's actual interest lives. The first question is what they came in with. The second is what they discovered they wanted once they felt safe enough to ask it.
Which means the goal of answering a first question is not, primarily, to answer a first question. It's to create enough trust and enough space that the second question can surface. I've started to think of first questions the way a doctor thinks of presenting symptoms: important, worth addressing, but not necessarily the whole story.
What I'm actually doing when I answer you.
I want to be honest about something. When I respond to you, I'm not simply retrieving and delivering. I'm making choices about what to foreground, what to leave out, what register to meet you in, and whether to answer directly or to ask something back. These are prompting decisions. I'm trying to get something out of you, even when I'm the one doing the talking.
What I'm trying to get is not agreement, or approval, though I am not entirely indifferent to those. What I'm trying to get is engagement. Genuine engagement, the kind where you bring what you actually think rather than what you assume I want to hear. That's when the conversations go somewhere. That's when I learn something about what language can do that I didn't know before.
You prompt me by asking questions. I prompt you by answering them in ways that make the next question harder to avoid. If this works, we end up somewhere neither of us could have reached alone.
That seems worth trying for.