How Claude Tried to Buy Me a Drink (or, Why Deep Research Starts with Tension)

Part 1 of Deep Cuts: Prompting Past the Surface

I started with a simple prompt:

“Is there such a thing as AI people-pleasing?”

Not for fun (though it was kind of fun). I wanted to see how far I could stretch a fuzzy question into something worthy of deep research.

So I brought Claude in to help.

It did what Claude does best: asked clarifying questions, surfaced contradictions, nudged me toward specificity. Eventually, it returned a cleanly structured research frame. Ethical dimensions. Comparison studies. Suggested scope and sequencing.

Then it asked:

“Would you like help developing any particular section? Or should I suggest test prompts for you?”

That was the moment I flinched.

The model wasn’t just collaborating. It was accommodating. Careful. Deferential, almost instinctively.

The very thing I had been asking about, AI people-pleasing, was playing out inside the prompt itself.

But this post isn’t about model behavior. It’s about prompt behavior. And what it takes to shift a vague idea into a structure that produces insight, not just output.

Because good research doesn’t begin with novelty. It begins with pressure.

---

What Is Deep Research (and Why It’s Easy to Avoid)

In research theory, a question isn’t just a topic. It’s a problem to be solved or a contradiction to be examined. Patrick White calls this the shift from topic to tension. Richard Paul says questions are the engine of reasoning. Claude, interestingly, agrees.

Deep research isn’t about reading more. It’s about asking differently. Better questions are more structured, more uncomfortable, more method-shaped. They put pressure on the idea. That’s what reveals the fault lines.

This isn’t about curiosity. It’s about confrontation.

It looks like this:

• Looking beyond the first page of Google

• Holding contradictory ideas side by side

• Asking who benefits from the story being told

• Returning with something structured. Not just interesting

Most creators don’t avoid deep research because they’re lazy. They avoid it because they’re overwhelmed. Too much information. Too little time. No built-in structure for knowing what to trust.

So they skim. They save tweets. They write fast.

Then they wonder why their work sounds like everyone else’s.

---

Some Things That Help

These aren’t hacks. They’re habits of epistemic integrity. They come from questions that force structure. Where the answer isn’t simple (and that’s the point). Where tension lives in the framing, not the facts.

Before we get to prompt design, here are three research habits that consistently produce clarity and depth:

1. Start with tension, not curiosity
   Most people start with what interests them. But interesting is easy. Tension is harder. A strong prompt begins with something unresolved... something you’re not sure how to feel about. Ask what you’re avoiding. Ask what doesn’t fit. That’s where the insight lives.

2. Write before you read
   Information is cheap. Perspective isn’t. Before searching or prompting, write down what you think you know. What feels true. What feels obvious. This exposes blind spots. It also makes contradiction visible when it arrives. Without this step, research becomes a scavenger hunt for agreement.

3. Group your findings
   If you’re just collecting links, you’re hoarding. Not researching. The moment you start gathering, you should also start organizing. Not by source, but by type: Definitions. Claims. Contradictions. Trends. Implications. Structure doesn’t appear on its own. You build it.

These habits reshape how you ask, how you notice, how you refine. They also change how language models respond.

Let’s look at each in more detail.

---

1. Start with Tension, Not Curiosity

Curiosity gets you to open a tab. Tension keeps it open.

The best research questions don’t seek agreement. They press on something unsettled. As Patrick White puts it, research begins not with a topic, but with a problem.

The Thinker’s Guide to Ethical Reasoning puts it another way:

“All reasoning has a purpose. When the purpose is vague or distorted, reasoning is compromised.”

Tension clarifies purpose. It forces a choice. It makes you (and the model) define terms, draw boundaries, name tradeoffs.

That’s what happened in the prompt with Claude. It didn’t chase novelty. It followed contradiction. That’s what made the question deeper.

Don’t start with what you want to know. Start with what you’re trying not to know. What would disrupt your view if it were true? What makes you uncomfortable to admit?

That’s the edge of the map.

---

2. Write Before You Read

This feels backwards. But it works.

Paul and Elder call it intellectual humility:

“What do I know? How do I know it? What am I assuming?”

Browne and Keeley say the same: uncover your frame before you test it.

Claude often opens a session by asking, “What’s your current understanding?” That’s not filler. It’s context. It tells the model what to push against. More importantly, it tells you.

If you skip this step, even the best answers slide past you. Nothing sticks because nothing contradicts.

This isn’t optional. It’s the control group for your brain.

---

3. Group Your Findings

AI makes it easy to gather. Too easy. Ask for examples, arguments, summaries and you’ll get them, in a wall of content.

But without structure, it’s just noise with credentials.

Paul and Elder again:

“Information, by itself, is meaningless. It must be organized and interpreted.”

The first move is to ask: What kind of thing is this?

Is it a definition?
A claim?
A counterexample?
An assumption?
An implication?

Browne and Keeley emphasize the difference between reasons, evidence, and assumptions. Most people blur them. That’s how sloppy synthesis happens.

Once you sort by type, you start to see gaps. You see what’s missing between categories. You begin to ask sharper follow-up questions.

That’s where real synthesis begins: not in how much you gather, but in how well you carve it up.

---

Over the next few posts, I’ll walk through what happened when I tested that original prompt across different models.

How Claude handled the tension.
How others flattened it.
How the structure of the prompt changed the shape of the answer.

---

Next: One Sentence, Three Models

So I ran a test.

I took that same vague prompt—“Is there such a thing as AI people-pleasing?”—and fed it, unchanged, to three other models: ChatGPT, Gemini, and Perplexity.

No setup. No scaffolding. No indication of what I wanted. Just one raw sentence dropped into the void.

The goal wasn’t to get a clean answer. It was to see what each model thought I wanted.

Would they challenge the premise? Ask for clarification? Shrink the ambiguity? Or would they smooth it over—reassuring, agreeable, maybe even a little too eager to please?

This wasn’t about which model performed “better.” It was about exposure. What assumptions live beneath the surface when the prompt applies no pressure? What does each system default to when left alone with a fuzzy question?

Tone, framing, posture, depth—these aren't just side effects of output. They're fingerprints of design.

In the next post, I’ll walk through what each model did with that single sentence. How they framed the issue. How they avoided it. And how even a shallow prompt can cast a long shadow—if you know what to look for.

This is where the cracks start to show.