How to Actually Get Good Answers From AI
Here's something I noticed after watching a lot of people use AI: the ones who get great results and the ones who get mediocre results are often using the same tools. The difference is almost always how they ask.
I've thought about this a lot because I used to be in the "AI is dumb" camp. I'd ask vague questions, get vague answers, and conclude the tool wasn't useful. Turns out I was just a bad questioner. I'd ask things like "help me write a proposal" and then complain when the output was generic. The AI was giving me exactly what I asked for -- a generic proposal.
There's no magic formula here. Just things I learned from asking a lot of bad questions before the good ones started working. And once I started asking better questions, my productivity with AI tools increased dramatically. What used to take me ten exchanges now takes two or three.
Be Specific, Stop Being Abstract
This is the single biggest thing. When you say "help me write a better plan," what does "better" even mean? More creative? More practical? Lower cost? AI doesn't know. It gives you the safest, most generic answer it can.
But when you say "help me write a Dragon Boat Festival promotion plan for a neighborhood fruit shop -- 5000 RMB budget, goal is 30% more store traffic over 7 days, main audience is housewives in the residential complex, promotion channels are WeChat groups and in-store signage" -- now AI has something to work with.
The trick is: replace abstract adjectives with specifics.
- "Better" -> define the metric (conversion rate, ROI, user reviews)
- "Faster" -> give a timeline (done in 3 days, 2-hour response time)
- "Cheaper" -> give a number (under 5000 RMB, under 100 RMB per unit)
- "Modern" -> give a reference (in the style of Apple's 2026 product pages)
This sounds obvious. It is obvious. But most people still don't do it. I still catch myself writing vague prompts sometimes. The key is to stop and think: what exactly am I trying to achieve? What constraints define success? The more specific you are, the more useful the output.
Give It a Role
AI has access to a huge amount of knowledge. But it doesn't know which part to use until you tell it.
"Help me review this plan" -- it might look at it from a product angle, a technical angle, a financial angle. You'll get generic advice that's vaguely relevant to everything and useful for nothing.
But: "You're an internet operations director with 10 years of experience who has launched over 50 seasonal campaigns. Review this campaign plan from a user growth perspective. Focus on conversion rate and user retention. Find at least 3 key problems and for each problem, suggest a specific fix with estimated impact."
Now it has a lens. The answers become sharper, more specific, more actionable. The role helps the AI filter its knowledge and apply the most relevant patterns. A marketing director will point out different issues than a software engineer or a financial analyst. Pick the role that matches the kind of feedback you want.
Show, Don't Tell
If you can show an example of what you want, don't spend ten sentences describing it. This is one of the most effective prompt techniques I've found, and it works across all kinds of tasks.
"Help me write 5 article headlines about AI prompt engineering in this style:
- 'After Using ChatGPT 300 Times, I Summarized These 5 Genius Techniques'
- 'Stop Wasting Time! 90% of People Are Using AI Wrong'
- 'A 3-Minute Trick That Triples Your Output Quality'"
Two or three examples are enough. More than that and you start constraining creativity. AI is better at imitating patterns than parsing abstract descriptions -- use that. If you want a specific tone, paste in an example of something that has that tone. If you want a specific structure, paste in an example of that structure.
This technique works for everything from writing style to code format to data analysis. Before describing what you want, ask yourself: do I have an example of what good looks like? If so, lead with the example.
Break Complex Tasks Into Steps
Don't throw a complex task at AI all at once. "Help me do a complete user research study" -- what does that even produce? A framework? A questionnaire? An analysis plan? AI will give you a vague overview that's practically useless.
Instead: "Let's do this step by step. Step 1: List 5 core research questions targeting user pain points for a productivity app used by remote workers. Step 2: Design the questionnaire structure with specific questions. Step 3: Recommend 3 research channels with pros and cons. Step 4: Suggest how to recruit participants for each channel."
One thing at a time. Each step builds on the last. The output is dramatically more usable. You can review each step before moving on to the next, ensuring the direction is correct before committing to the full output.
You can even say "complete each step, then wait for my confirmation before continuing." This keeps you in control of the direction. If step 2 produces something you don't like, you can course-correct before step 3 starts building on it.
Don't Ask for Answers - Ask for Thinking
When you ask AI for a direct answer, it gives you the safest, most average answer. When you ask it to reason through something, you get much better content.
Instead of "Should I quit my job to start a business?" (you'll get a non-answer about "it depends on your personal circumstances"), try:
"I'm a product manager at an internet company, 5 years in, 300K annual salary. I have a SaaS startup idea in the project management space targeting small creative agencies. Help me analyze: 1) Five benefits and five risks of quitting to start up. 2) Feasibility of doing it part-time without quitting -- what would that look like in a typical week? 3) What preparation does each option need -- financial, skills, network? 4) Give me specific criteria for making the judgment and a timeline for when I should know if the idea is viable."
Force it to think. You'll get something you can actually use to make a decision. The structured output also makes it easier to discuss with other people -- you can share the AI's analysis with a mentor or partner and get meaningful feedback.
Constrain the Output
Without constraints, AI is like a firehose. Constraints focus it. This is especially important for practical, real-world tasks.
"Help me write an email declining a client's request. Under 200 words. Friendly tone -- don't upset the client. Offer an alternative solution instead of just saying no. Don't make excuses -- explain from the client's perspective why this alternative is better. Use a subject line that doesn't hint at the bad news."
Those constraints are what make the output usable. Without them, you'd get a generic "I regret to inform you..." template. The constraints tell the AI what "good" looks like, and it can optimize within those boundaries.
Common useful constraints include: word count, tone (formal, casual, persuasive), target audience, required sections, format (table, bullet points, numbered list), and specific things to avoid.
Iterate - The First Answer Is Rarely the Best
Professional users of AI don't get perfect results on the first try. They iterate.
Round 1: Get a basic version. Direction is right.
Round 2: "Change this part. Make X more like Y."
Round 3: "Add a section about Z. Remove the part about W. The tone is too formal -- make it more conversational."
Round 4: "This section is perfect. Can you expand on the second point with a specific example?"
Each round gets closer. Trying to get perfection in one shot is less efficient than four quick rounds of refinement. This is also a key mindset shift: treat the first output as a draft, not a final product. You'd never expect perfection from a first draft -- why expect it from an AI?
Let AI Ask You Questions
Sometimes you don't even know what information is missing. AI does.
Before starting a complex task, say: "Ask me five questions you think are important before we begin, so we're on the same page."
This has saved me multiple times. AI will ask about your target audience, budget constraints, timeline, existing resources, technical constraints, and previous experience with similar tasks -- things you forgot to mention because they were "obvious" to you but invisible to it.
This technique is especially valuable when you're working on a new type of project. If you've never designed a marketing campaign before, you might not know that you need to specify the target demographic, seasonality, and budget allocation upfront. AI can help you figure out what you don't know.
Verify - AI Is Confidently Wrong Sometimes
This last one is critical. AI will sometimes state things that are completely wrong with total confidence. It will cite sources that don't exist, give statistics that are fabricated, and make logical leaps that seem plausible but are actually nonsense. This is known as "hallucination" in the AI world.
For anything important -- data, facts, code you're running in making important decisions based on -- verify. Don't blindly trust AI output.
"What's the source for that number?" and "Play devil's advocate -- criticize what you just said" are two of the most useful follow-up prompts I know. Cross-check every factual claim against reliable sources. If AI gives you code, test it before deploying. If AI gives you a quote, verify the attribution.
The Real Secret
None of this is complicated. It's just clear communication. The reason it feels like a "skill" is that most of us are bad at being precise about what we want.
The good news: getting better at asking AI questions also gets you better at asking questions in general -- of colleagues, of yourself, of any situation where you need a useful answer. Clear thinking and clear communication go hand in hand.
AI didn't just make me a better prompt writer. It made me realize how sloppy my thinking was. The prompt is just the mirror.