How to Ask AI Useful Questions
Prompting isn't a skill you study — it's four habits that fix 90% of bad answers.
Open in the XNewsAi app →"Prompting" sounds like a technical skill. It isn't — it's four habits, and once you know them you'll rarely write a bad prompt by accident again.
Say exactly what you want
Not "write an email" — "write a 3-sentence email telling a customer their estimate is delayed two days, apologetic but not groveling."
Tell it who you are
"I run a 4-person roofing crew in Lubbock" changes the answer more than almost anything else you can add.
Say what "done" looks like
"Give me 3 options," "keep it under 100 words," "make it a table" — it will match whatever shape you ask for.
Treat it like a conversation
"Make that more casual," "shorter," "add a line about the warranty" — it remembers what you just said and revises instead of starting over.
Where you'd use this: everywhere you already talk to AI. The habit compounds — five better prompts a day saves real time by the end of the week.
Next action: Ask XAi the same real question three different ways — vague, then specific, then specific-with-context — and watch the answer change.
Go deeper
Under the hood, a prompt is doing two jobs at once: it narrows the space of plausible next-word predictions (specificity), and it supplies context the model has no other way to know (your trade, your customer, your voice). Neither is "tricking" the model — you're giving it the same information a competent human assistant would need to do the job right.
Two techniques worth knowing by name, because you'll see them referenced elsewhere:
Show, don't just tell
Paste one real example of the thing you want ("here's a past estimate email — write the next one in the same voice") and the model copies the pattern far more reliably than a description alone.
Ask it to show its work
For anything with real logic — pricing math, scheduling conflicts — add "walk through your reasoning step by step before giving the final answer." It catches its own mistakes more often when it has to show them.
Limitation worth knowing: no prompt fixes a model that genuinely doesn't have the facts. If it's guessing at something outside its training or outside what you gave it, better prompting gets you a more confident wrong answer, not a correct one — that's what live web search and your own verification are for.
Examples
- Roofing: 'You're helping a small roofing company. Rewrite this estimate email to sound less formal, keep the price the same.'
- Sales: 'Give me 3 versions of a cold follow-up text, each under 25 words, for a customer who went quiet after a quote.'
- Farming: 'Explain this USDA program like I've never read a government form before, in 4 sentences.'