AI can polish a sentence. It can't give you a point.

A brief that says "write about allergy safety in restaurants" has handed over a topic. Every piece written from it will be interchangeable with every other piece, because there's no position anywhere in it. Nothing in the brief can be argued with, so nothing that comes out of it is worth reading.

Three questions separate a topic from an idea. What is believed here. Why does it matter. What part of it isn't obvious. If the third question has no answer, there's no piece yet. There's a subject heading.

Then the harder question, which does most of the work: what's being softened because saying it plainly feels risky, or too specific, or a little strange? That's usually where the actual idea is sitting, one draft away from being useful to somebody.

AllergyTable researches whether a restaurant can feed someone with a serious food allergy. The obvious thing for a tool like that to say is "safe." Safe is the word people search for. Safe is the word that converts.

Its spec says never to use it. Not once, anywhere in the product. Verdicts are graded instead, fine or ask or avoid, and each one carries its source and the date that source was retrieved. The positioning is research assistant, never medical authority.

That decision costs conversions, and it's also the entire product, because a safety tool willing to say "safe" when it doesn't know is worth less than nothing. The conviction wasn't a layer of polish applied to the copy afterwards. It was the product decision, and the copy simply followed it.

Writing that rule down felt risky. That feeling is worth stopping on rather than smoothing over, because it shows up at a very specific moment.

When a claim feels slightly too strong to publish, there are only two explanations available. Either it's wrong, or it has finally gotten specific enough to be useful. From the inside those two sensations are close to identical, which is why so many drafts get hedged at exactly the point where they were about to become worth reading.

The way to tell them apart is to ask whether the claim can be defended with something concrete: a number, a source, a decision somebody actually made. If it can, the discomfort is a signal to publish it, not to sand it down.

When an idea is still cloudy, talk about it for sixty seconds with no attempt at elegance. Then say the same thing in twenty. Then in ten.

The ten-second version is the claim. Everything else was scaffolding holding it up while it was still wet. AllergyTable's ten-second version took months of building to arrive at and runs four words: never say safe.

The order matters, and most people run it backwards. Starting with compression produces a tidy sentence about nothing, because the connections that make the idea work haven't surfaced yet. The sixty-second version is where the thinking happens. The ten-second version only reveals what the thinking was.

Comedians work a bit for months. The observation is usually right early on. The wording, the example and the running order take far longer to find. Most writing advice would throw the bit out because the first performance went badly.

So before discarding an idea, push it through five frames: a story, a concrete example, a contrast, the strongest counterargument, and the implication nobody sees coming. An idea that dies in all five was genuinely weak. An idea that dies in four and comes alive in one was fine the whole time, and it nearly got deleted.

AllergyTable's best feature came out of exactly that process. Cross-contamination is the thing people most need to know and it's almost never published anywhere online. Framed as a data problem it was a dead end for weeks. Framed as a question you ask while standing in the restaurant, it turned into a generated script for that specific kitchen: is the fryer shared with the coconut shrimp? The weakest input in the system became the strongest thing it produces, and only the fifth framing showed that.

AI is very good at everything downstream of a decision. It drafts, restructures, tightens, checks and formats, and it does all of that faster than any person.

What it can't do is supply the belief. It holds no stake, it hasn't shipped the thing, it has never watched a decision cost money, and it will produce a fluent, well-structured, entirely interchangeable paragraph on any topic handed to it. That isn't a limitation to prompt your way around. That's what the tool is.

So the scarce input hasn't moved. Somebody still has to decide what's true, what's worth saying, and what they're prepared to be wrong about in public. Do that first and the machine makes it better. Skip it and the machine just makes it longer.