It predicts the next word
An AI assistant continues text; it doesn't read your mind. See why the exact words you type decide the answer you get, and start writing your intent into the prompt.
THE ENGINE
A machine that continues text
Behind every chat assistant is a language model. It was trained on a huge amount of text with one goal: given the text so far, predict what comes next, one small piece (a token, roughly a word or part of a word) at a time. Your message is the text so far. The answer is its continuation.
Give it "The opposite of hot is" and the most likely next word is "cold". A full answer is that same step repeated hundreds of times, each new word added to the text before the next one is predicted.
Your phone's autocomplete, grown enormous
Your keyboard suggests the next word from what you've typed. A language model does the same kind of thing with far more text behind it and far more of your message taken into account. Where the analogy breaks: after that first training, assistants are trained further to follow instructions and be helpful, so they answer your question instead of just rambling on from it. But the engine underneath is still prediction.
Check yourself
Nima types just "cover letter" and gets a bland, one-size-fits-all letter. What best explains the result?
- The assistant is low quality and he needs a different one
- Those two words fit millions of situations, so the likeliest continuation is an average letter
- It was being lazy, and asking for the same thing again would get him the better letter it held back
- The assistant already knows his job history from his profile and chose not to use it here
Show the answer
Those two words fit millions of situations, so the likeliest continuation is an average letter
Right. With so little text to continue, the safest prediction is the middle of the road. Average input, average output.
WHY WORDING MATTERS
Different words, different continuation
Because the answer is a continuation of your exact text, every word you add shifts what is likely to come next. Add who it's for, and the vocabulary changes. Add the length, and the length changes. Add your situation, and the advice stops being for everyone and starts being for you.
"Explain inflation" pulls toward a textbook paragraph. "Explain inflation to my 12-year-old sister using the price of ice cream" pulls toward something short, concrete and friendly.
Same person, two prompts
Weak prompt
"Tell me about coffee."
Could continue as history, farming, recipes, health or business. You get a little of everything and nothing you needed.
Improved prompt
"I drink 4 cups of coffee a day and sleep badly. In 5 short sentences, explain how caffeine affects sleep and what time I should have my last cup."
One topic, one reader, one length. Only a useful answer fits.
Check yourself
Leave the reader, the length and the tone out of a prompt and you still get a complete answer: the model fills each one in with its most common version.
Show the answer
True
Nothing is left blank. The model has to continue your text, so it picks a default reader, a default length and a default tone. That's why a vague prompt comes back polished and still wrong for you. Anything that matters has to be in the text.
THE BLIND SPOT
It can't see what's in your head
When you ask a colleague for help, they already know your job, your boss and what happened last week. The model has only the text in front of it. It doesn't know why you're asking, who will read the result, what you already tried, or what "good" looks like to you, unless you write it down.
"Make this email better" hides everything that matters: better for whom? Shorter? Politer? More persuasive? You know. The model has to guess.
This email is to my landlord, who is formal and a bit short-tempered. I want him to fix the water heater this week, and I don't want to sound like I'm threatening him. Make it polite, firm and under 100 words.
[email pasted here]Dear Mr. Karimi,
I hope you are well. The water heater in unit 4 stopped working on Saturday, and we have had no hot water since. Could you please arrange a repair this week? I am home every day after 5pm and happy to let the technician in...Reader, goal, tone and length were all in your head a moment ago. Now they are in the text, so the continuation has to respect them.
Check yourself
Negar asks an assistant to plan her sister's birthday. What does the model have, and what is missing unless she writes it?
- General ideas for birthday parties
- Her budget
- That her sister hates surprises
- What a typical party checklist looks like
- How many guests fit in her flat
- The words she typed in this chat
Show the answer
It already has this: General ideas for birthday parties, What a typical party checklist looks like, The words she typed in this chat
Only if she writes it: Her budget, That her sister hates surprises, How many guests fit in her flat
What this means for you
- A prompt isn't a search query. It's the beginning of a text the model will finish.
- Vague in, average out. The gaps you leave are filled with the most common guess.
- Before you send, ask: what do I know about this task that isn't on the page yet?
- A disappointing answer is information: it shows you which detail was missing.
Check yourself
Sara pastes her CV and writes "Is this good?" The reply is vague praise. Which rewrite uses what you learned?
- "I'm applying for a junior accountant job at a small firm. Name the 3 weakest parts of this CV for that job and how to fix each."
- "Is this good??? Please be honest."
- "You are the best CV expert in the world, with twenty years of hiring experience. Read this carefully and tell me honestly: is it good?"
- "Rate this CV."
Show the answer
"I'm applying for a junior accountant job at a small firm. Name the 3 weakest parts of this CV for that job and how to fix each."
Yes. She put the job, the reader and the kind of feedback she wants into the text, so a vague compliment no longer fits as a continuation.
Lesson recap
- An AI assistant is built on a language model that predicts what text comes next, piece by piece.
- Your prompt is the start of that text, so your exact wording steers the answer.
- The model can't see your goal, reader or situation unless you write them.
- Short, vague prompts get the most average answer; add what's in your head and the answer becomes yours.