When the AI answers too shallowly: eight techniques to make it think deeper

Prompting was long sold as a collection of clever phrasings. Somewhere there was supposed to be the one right sentence after which a model suddenly works more intelligently. In practice, strong prompting works differently. Its purpose is not the polite request but the fitting mode of thinking: the model should break the question apart, contradict an idea, search for weak spots and check its own conclusions once more.

In the last piece it was about bringing a finished answer up to working quality — editing it, cross-checking it, playing through failure in advance. This time I set in earlier. You give the model the direction before it writes the first word. Eight techniques, two groups. The first opens up a topic. The second secures a decision.

Two things up front. Context is half the battle — the prompts below are scaffolds you fill with your situation. Where I write “[topic],” your concrete question belongs, with goals, resources, constraints. And treat the templates as a starting point. Change them, add your own questions, drop what does not fit.

Techniques that open up a topic

The first four techniques serve understanding. You come to the AI with a topic and give it a format for thinking: how it fans the topic out, in which order it proceeds, where it digs deeper.

Vibe learning: drift, but with a rudder

A good entry into an unfamiliar topic, and one I often choose. You do not write a long study plan. You pick a starting point and guide the AI further with your questions. Say you want to understand the history of writing. The first prompt:

I want to understand the history of writing. Suggest a few entry points for me to choose from.

You take the point that appeals to you and add:

Option [choice]. We go through the topic chapter by chapter. After each chapter you suggest two things: deeper questions on the chapter and possible directions for the next step.

The gain lies in the control. At every moment you decide whether to go deeper or move on.

Socratic method: be questioned first, then understand

The Socratic mode is so common among AI users that dedicated learning modes build on it. The idea is simple. Instead of outputting material chapter by chapter, the model first asks you questions and adds to or corrects you afterward. A typical prompt:

Don’t answer right away. Help me penetrate [topic] Socratically: ask me one guiding question after another, expose hidden assumptions, demand examples and counter-readings. Only after that do you assemble the explanation.

The mode does not fit every purpose. If you need a quick fact, ask directly. And if you know nothing at all about a topic? Then you give nonsense back to every question and get annoyed. For review and for deeper penetration the method is strong.

First principles: break the topic down to the ground

From Socrates to Aristotle, more precisely: to his thinking in first principles, where every question goes back to its basic building blocks. With AI this is especially useful. Modern models like to reach for ready-made phrasings and other people’s assessments from the web and output them as their own answer. First principles forbid exactly that:

Explain [topic] from first principles. Do not lean on common phrases and familiar explanations, the internet only for numbers and facts. Name the basic building blocks and the fixed boundaries of the topic first, then rebuild the explanation from there step by step.

That works far beyond learning. Take a question from my day-to-day: why does a brand not appear in AI answers? Broken down from first principles, it splits into three sub-questions. Is the brand recognizable as an entity of its own? Do independent third-party sources confirm it? Does it stand in the same context as its product category? From these building blocks comes a finding instead of a guess.

Decomposition (least-to-most): big task, small steps

A common mistake in dealing with AI: handing over the task in too large a chunk. Then the model answers too generally or skips important intermediate steps. Least-to-most solves this by having the AI first cut the task into a sequence of smaller subtasks and then work through them from top to bottom.

Explain [topic] by the least-to-most principle. Show the decomposition plan first, then work through the steps one after another, so each builds on the previous one.

If you want more control, have each step output as its own chapter and step in between with follow-up questions. Which variant fits depends on the model. Current top models like Claude Opus 4.8 or GPT-5.5 often pour out the whole text at once, smaller models guide you better in small pieces. The same grip helps with planning. A project for GEO, that is Generative Engine Optimization, runs more cleanly if the AI breaks it down first: measure, prioritize, execute — and for each block names what to do and which risks lurk.

Techniques that secure a decision

Penetrating a topic is one half. The other: building no botch when acting. The next four techniques do not explain, they check. You come with a finished idea, a plan or a position — and have the AI search deliberately for holes, calculate consequences and take other people’s view.

Red team: the stress test for your idea

Red team comes from IT security. That is the name of the group that deliberately attacks a system to find weaknesses before someone else does. Before the launch of new models, Anthropic and OpenAI invite exactly such people. The same logic you lay on your own plans. You explicitly ask the AI to attack:

I’m planning [idea/plan]. Do a red team: find weaknesses, hidden risks and reasons this can fail. Don’t be polite — I want honest criticism.

The models have become more honest on their own; for an obviously weak idea they say so. Without a clear instruction they still tend toward balance: a bit of pro, a bit of con. Red team tips this balance. In the last piece the technique appeared only as a link in the chain, here it gets its own place. An example from my field: before a content project starts, I have the AI break the planned strategy apart. Where do I overestimate the effect? Which assumption does not hold?

Thought experiment (what-if): the consequences step by step

Thought experiments are ancient. Einstein imagined chasing a beam of light and came closer to special relativity. What he had to calculate in his head, the AI rolls out in a minute — with numbers, sources, counter-arguments. You give a hypothetical condition and have the consequence chain built:

Do a thought experiment. Suppose [hypothetical condition]. Trace the consequences step by step: what happens first, which second- and third-order effects follow, who wins and who loses. Pull current numbers from the web.

The last line is important. The AI market turns fast, and even strong models do not know the current state, their knowledge lags months behind. What-if fits well after red team: you take the biggest risk from the stress test and roll it out as a scenario. Suppose AI answers take over half the clicks on informational questions — what does that mean for your channel, and where do you notice it first?

Steelmanning: build the strongest counter-argument

Many know the devil’s advocate: someone deliberately holds an unpopular position to test the other side. Steelmanning flips it. You take a position you reject and have the AI build the strongest arguments for it. Not a straw man to blow over, but the version that is hard to beat.

Why does this help? Because we all live in bubbles. Whoever prefers remote work to the office has a feed full of arguments for remote. The AI pulls you out of there if you ask it to.

I’m convinced: [your position]. Build a steelman — the strongest arguments against my view. No caricature, but the version a smart, well-informed opponent would put forward. Back it with numbers from the web.

In marketing I apply this to my own beliefs. “More content brings more visibility” is one such sentence, popular and rarely checked. A steelman against it quickly shows why structure and authority often weigh more than sheer quantity.

Multi-perspective: the advice that does not flatter

Giving a model role an abstract biography (“you are the world’s best marketer with 20 years of experience”) does little. The model tries hard anyway, and the invented CV only clogs the context. Roles do work in one case: when there are several of them and they look at the same thing from different interests.

I’m planning [decision/product]. Assess it from three perspectives: 1) [role] — what it sees first; 2) [role] — which objections it has; 3) [role] — what matters most to it. Then a conclusion: where the perspectives align and where they rub.

The trick lies in the concrete roles. Instead of “be an expert,” you give a specific lens. A product page, for instance, is seen differently by three pairs of eyes: the regular reader who comes every week; the first-time visitor who does not know the brand; the competitor who already has a similar page live. Each lens brings different objections, and the model keeps them apart.

Eight links for your chain

All eight techniques share one thought: you do not ask the AI for an answer, you give it a way of thinking. Vibe learning and Socrates steer the learning. First principles and decomposition break the big down to the small. Red team and what-if put your plans under pressure. Steelman and multi-perspective show what you cannot or will not see alone.

Combine them. Freshly opened up a topic? Send the result through a red team. Received criticism? Roll out the main risk in a thought experiment. Before a decision, first a steelman against your own position, then the multi-perspective for the blind spots. The five techniques from the last piece — self-refine, chain-of-verification, pre-mortem, inversion, five whys — attach to the same chain and work just as well.

The longer the chain, the smaller the chance that something important escapes you. Not every task needs all eight; for something small, one link is enough. A prompt is not a magic formula. It is the direction of thinking you set. The better you understand how a model works, the less often you need ready-made formulas, and the more often you write your own.