Prompting is not magic. It is a framework problem.

If your prompts produce AI-flavoured mush, the problem usually isn’t the wording. It’s that you handed the model a task without handing it a method. Give it the method you would expect a senior colleague to use, and the same model produces work you would actually ship.

The equation most people use

This works, in the sense that something comes out. But the output tends to be generic, and when the task sits outside your own domain you have no way to catch what the model got wrong. You are asking it to decide both what good looks like and how to produce it.

The equation that works

Take a framework: a best practice, an industry standard, the way an expert would actually approach the task. Traditionally that came from senior people on your team, consultants, or years of your own practice. You supply that, and the model’s job shrinks to something it is genuinely good at: applying a known method to your specific case.

What this looks like on a real task

Say you are writing a job description for a data analyst.

The first way: “Write a job description for a data analyst.” You get five bullet points about “strong analytical skills” and “attention to detail” that could describe anyone.

The second way: “Write a job description for a data analyst. Use this structure: one paragraph on the problem this person will own; the three outcomes we will judge them on in the first year; the tools they will actually use day to day; what they will not be responsible for. No adjectives about personality. Every requirement must be something we can assess in an interview.”

Same model. The second output is a document you can argue with. The difference is not better wording. It is that the second prompt carries a hiring framework inside it.

Where do you get the framework if you have no senior to ask?

This is the fair objection to everything above. Three places that work:

1. Your own past work. The last time your team did this well, what was the sequence? Write it down once and reuse it.
2. Published standards. Most fields have them, and most are free.
3. The model itself, but in two steps. Ask it to list how experts approach the task and what separates good from bad, then read that critically and choose. You are using it to surface options, not to decide.

The step you cannot skip is judgement about which framework fits. That judgement is the part that stays yours.

The framework is a knob

How much you specify is how much control you keep. Tight framework, and the model executes. Loose framework, and it fills the gaps with the average of the internet.

Neither is always right. For a first draft of something exploratory, loose is fine. For anything going in front of a client, tight.

Why this makes cheaper models good enough

Most of what people pay frontier models for is judgement they could have supplied themselves. Once the method is in the prompt, the remaining work is execution, and smaller, cheaper models do execution well. In my own work, moving the framework into the prompt mattered more than moving up a model tier.

What this means for your work

If the first equation covers most of what you do, that work is the most exposed. Not because AI is clever, but because work that needs no method is work anyone can hand to a model.

The second equation describes a different job: knowing which method fits, judging whether the output is any good, and owning the result. That part still needs a person. And frankly speaking, this is the only way that makes sense.