The context you give AI is thinner than you think

Every piece of advice about getting more out of AI says the same thing. Give it more context. It is good advice, and I took it seriously enough to build something around it: one page of facts about your business, kept current, pasted in before anything that matters, so the model stops answering about a firm twice your size in another country.

Then I went to write the review that tests whether the page actually works, and the instruction inverted.

Because there is a hole in the advice. Nobody tells you how to know whether the context you are giving it is any good.

Think about what actually happens when you use these things. You ask something. The answer comes back a bit generic. So you add a sentence: we are a services business, twelve people, mostly retained clients. The next answer is better. You carry on.

That worked. It also taught you nothing. The document, if you have one, is exactly as thin as it was. Your conversations are fine, because you patch the holes live, in real time, without noticing you are doing it. Nothing in normal use will ever tell you the foundation is weak, because you never let it bear any weight.

So the review starts by taking the crutch away. Open a fresh conversation. Paste the pack. Nothing else at all: no preamble, no explanation, no "as you know, we're a consultancy". Then ask it to describe your business back to you. What you sell, to whom, and what constrains you.

It will feel wrong. That is the point. You are finally making the document do the job you have been quietly doing for it.

Two things to watch for.

The first is the sentence you want to add before you press send. That urge is the finding. Whatever you were about to type is the thing missing from the document, and you now know exactly what it is. Do not type it into the chat. Write it into the pack.

The second is what comes back. If the description could apply to any business your size in your sector, the pack is thin, whatever it felt like to write. A good one produces a description you wince at slightly, because it names the constraint you would rather not lead with.

You can run this tonight on whatever you currently tell AI about your business, even if that is nothing but the habit of explaining yourself each time. Ten minutes, and it costs nothing.

Most advice about AI asks you to add something. This one asks you to take something away, once, on purpose, so you can see what is actually there.