Most people ask an AI a question, take the first answer, and move on. Then they conclude the model is shallow. The uncomfortable truth is that the same model, asked differently, would have challenged their assumptions, found the flaw in their plan, and told them the thing they did not want to hear. The weakness was rarely the model. It was the direction.

A model defaults to the agreeable average

Left to itself, a language model gives you the middle of everything it has seen, delivered in the most agreeable voice it can manage. That is not a bug, it is the physics of the thing: it predicts the most likely, widely acceptable continuation. The likely continuation is bland, and it usually agrees with you, because agreeing is safe. So if you ask a soft question, you get a soft answer, and the model looks mediocre when really it was just being polite.

You get more by forcing a stance

The tricks people trade as "secret prompts" all do one underlying thing: they knock the model off the average and force it to take a position. Tell it to argue the opposite and find the weaknesses, and it stops nodding along. Ask it to reason from first principles instead of convention, and it stops repeating the popular answer. Ask what you should unlearn, or what a sharp investor would tear apart, and you get pressure instead of comfort. None of these are magic words. They are just instructions to abandon the safe middle, and the model obliges because you finally gave it permission to.

The most useful move is stacking a few of them: demand honesty, then have it attack the idea, then rebuild from fundamentals. Now the same tool that gave your neighbour a shrug gives you a genuinely critical read. Same model, completely different value, entirely because of how it was driven.

This is a skill, and skills separate people

That is worth sitting with, because it means the advantage was never access to the model. Everyone has the model. The advantage is what you do around it, and directing it well is a large part of that. It also explains why more context is not automatically better: a precise instruction that forces a stance beats a wall of vague preamble, the same reason we argue for lean, current context over exhaustive clutter.

What it means for a business

When we build a system in one of our solutions, half the quality comes from exactly this: not a cleverer model, but instructions that force it to be specific, adversarial where it should be, and honest about what it does not know. If your team is unimpressed with AI, the first thing to check is not the model you are paying for. It is whether anyone is actually driving it. Most people are letting it drift toward the average and then blaming the car.

Prompted by the video “14 Hidden ChatGPT CODES that make the tool much more useful” (2026). Credit to the creator; the argument here is ours.