How to Prompt AI When You Don't Know the Right Words

Describe the shape of the thing instead of naming it. That is the whole technique. When you do not know the term for what you want, the instinct is to search for the word first and then prompt, which wastes the step where AI is genuinely better than a search engine: it can go from a clumsy description to the right vocabulary in one turn, and then use that vocabulary properly.

The failure mode is asking a vague question, getting a vague answer, and concluding the model is not very good. What actually happened is that you asked it to guess your intent and it guessed the most common one.

Ask for the name before you ask for the answer

Split it into two turns. The first turn buys you the words. The second uses them.

I do not know the correct term for this, so I am going to describe it.

In a spreadsheet, I want the thing where the top row stays visible
while I scroll down the rest of the sheet. I have seen it called a
few things and I cannot remember which.

Give me the standard term, two or three near-synonyms it might be
called in different software, and the term for the related thing I
might actually mean instead.

That last clause is the part worth stealing. "The related thing I might actually mean instead" catches the common case where your description is closer to a neighbouring concept than the one you named. It is how you find out you wanted "split panes" rather than "freeze panes" before you spend twenty minutes on the wrong one.

Once you have the term, ask the real question with it. The answer quality changes noticeably, because the model is now retrieving against a precise concept rather than a paraphrase.

Give it the surrounding facts, not the label

When you cannot name a thing, you can almost always describe the situation it sits in. That context does the disambiguation for you.

Weak, because it forces a guess:

How do I fix the thing where my website looks wrong on mobile?

Strong, because there is nothing left to guess:

On my site, on a phone, the text runs off the right edge of the
screen and I can scroll sideways. On a laptop it looks correct.
It started after I added a table of prices. I do not know what
this problem is called.

Tell me what it is called, the two most likely causes given that
it started with a table, and how to check which one it is.

The second prompt contains four facts: the symptom, the condition it appears under, the condition it does not, and when it started. None require vocabulary. Together they narrow the space more than any single technical term would have.

This is the same principle as writing a good bug report, and for the same reason: the person receiving it cannot see your screen. Our notes on writing a bug report for an AI coding agent apply almost unchanged.

Use a wrong word deliberately, and say so

If you half-remember a term, use it and flag it as unreliable rather than omitting it. A wrong-but-adjacent word is genuinely useful signal, because it locates the neighbourhood.

I think this is called "normalising" the data but I am not
confident that is the right word.

What I mean: I have a column of prices where some say "12.50",
some say "£12.50" and some say "12,50". I want them all in the
same format so I can add them up.

If "normalising" is wrong, tell me the right term, then answer
the question.

Saying "I am not confident" changes the response. Without it, models tend to accept your framing and answer the question you literally asked, which sends you further down the wrong path. It is a small instance of a general problem covered in how to stop AI from being too agreeable.

Ask for the vocabulary map, not just one word

For a whole area you are new to, one term is not enough. Ask for the small set of terms and how they relate.

I am setting up email for my business and I do not know the
vocabulary. I keep seeing SPF, DKIM, DMARC, MX and "warming up".

Give me a table: term, what it actually does, whether I need to
care about it in my first week, and what breaks if I ignore it.
Order the table by what I should deal with first.

The ordering request is what makes this useful rather than a glossary dump. It converts a list of unfamiliar words into a sequence of actions, and it surfaces which ones you can safely defer.

When you cannot even describe it

Sometimes you do not have the words for the shape either. Two things work.

Compare it to something. "It should work like the undo in Google Docs, but for my form" carries an enormous amount of specification in nine words. Analogies to well-known software are dense and models read them accurately.

Show it. A screenshot with a circle drawn on it removes the vocabulary problem entirely. Prompting AI with a screenshot covers what to include and what to crop out.

The check that keeps this honest

Because you started without the vocabulary, you are not well placed to spot a confident wrong answer. Add one line to the second turn:

Before answering, restate what you think I am asking in one
sentence. If my description is ambiguous between two things,
say so and ask rather than picking one.

This costs a sentence and catches the misreadings early, while they are cheap. It is a specific case of getting AI to ask clarifying questions, which is worth turning into a habit rather than a one-off.

FAQ

Is it better to search for the term first?

Usually not, when you cannot name it. Search rewards knowing the word; a model can work backwards from a description. Search becomes the better tool once you have the term and want authoritative documentation.

Will using the wrong word make the answer worse?

Only if you present it as certain. Flagging it as a guess turns a wrong word from a misleading instruction into useful evidence about what you might mean.

Why do I get a generic answer to a vague question?

Because a vague question has many valid answers and the model returns the most common one. Adding the situation, the constraint and what you have already tried removes the ambiguity. This is the same mechanism behind getting non-generic answers out of AI.

Does this work for a specialist field with real jargon?

Yes, and it is where the two-turn approach pays most. Describe the situation in plain language, ask for the field's term, then ask your real question using it. The vocabulary step is what unlocks the specialist material.