If you've ever used AI to help you brainstorm or think through a problem and have been disappointed by what came back, I hate to break it to you, but the LLM might not be the problem. The way you prompt it probably is. I say this with full confidence because I've seen everyone around me do the exact same thing. They throw a half-baked question at an AI tool, get an equally half-baked answer, and then walk away thinking the tool is just hype.

The key is being intentional with how you prompt any LLM. Anthropic’s Claude has been my go-to AI tool lately, and these 5 prompts are what consistently get me smarter, more thoughtful responses.

Pushing Claude to ask better questions before answering

A little back and forth goes a long way

You know how when you're onboarding a new client or hiring a new service, the first thing they do is ask a ton of questions before they actually get started? Doing so ensures expectations are aligned on both ends, and that the output actually manages what you had in mind.

That's exactly the kind of behavior you want from an AI tool, too. Instead of giving Claude some context on what you'd like it to do and getting a response where it's clearly filled in a lot of gaps itself, it's best to explicitly instruct it to ask you as many questions as it needs before answering. Even if you don't give Claude a lot of context in your prompt, end it with:

Another great example is:

The entire purpose of a detailed prompt is to give an AI tool relevant context. If you skip that step by giving it a vague prompt that skips over the essential information, you're forcing it to make assumptions. Letting it ask questions flips that entire dynamic. Instead of guessing, it'll build its response around exactly what you want.

This saves you the hassle of spending a significant amount of time trying to fix the output it initially generated (which, let’s be real, you’ll likely do with another round of prompts anyway).

Three anti-hallucination prompts from Anthropic's own docs

Hidden in the docs, found on Reddit

While the prompts I'm going to talk about in this section are technically three, they're all essentially the way of saying the same thing: stop making stuff up. By now, no matter what AI tools you use, I'm sure you're familiar with the term "hallucinations." Despite how powerful new LLMs are, hallucinations are still a big problem. Fortunately, there are three prompts that Reddit user ColdPlankton9273 found in Claude's own API documentation that cut down hallucinations significantly. The prompts are as follows:

• "Allow Claude to say I don't know"
• "Verify with citations"
• "Use direct quotes for factual grounding"

I've been using all three of these prompts myself, and they've made a massive difference in the quality of outputs. I'm a huge fan of the citations one, since it makes it much easier to double-check where information is coming from instead of blindly trusting the response. As someone who’s a huge fan of NotebookLM because of its grounded, citation-based nature, this is the closest I’ve gotten to replicating that same level of reliability in Claude.

Give Claude a specific role

You've probably been doing this without realizing

One of the foundational prompt engineering techniques is role prompting, and you've likely been doing it since the early ChatGPT days without fully realizing it. It's simply giving your AI a character to play. For instance, you could instruct it to act like a senior developer, respond as if it's a marketing strategist, act as a teacher explaining a concept to a 5-year-old, and more.

So for instance, I'm a huge fan of using AI to help me study and break down complex concepts I don't fully understand. But like every other AI tool, Claude has a major overexplaining problem. Even when I'm struggling to understand the very basics, it tends to jump straight into long, detailed explanations that feel way more complex than they need to be.

So, instead of telling Claude to explain what a full adder is to me, I'll say, "Explain a full adder to me as if you're a retired professor explaining it to someone who's never taken an engineering class." You'll also find posts all over Reddit and X of people taking a more creative approach by telling Claude to explain complex topics using only pizza delivery metaphors.

I once asked it to teach linear algebra to me as if math concepts were characters in a reality show. It sounds unhinged, but for someone with a goldfish memory, the key to memorization is making things memorable. And frankly, the weirder the analogy, the harder it is for your brain to forget it!

Ask Claude to explain its own reasoning

If you can't explain it, you didn't learn it

Something that's increasingly common is people blindly copying the output an AI tool gives them, pasting it elsewhere, and then calling it a day. Now, while that might get the job done in the moment, you're essentially outsourcing your thinking. And the second something doesn't work or someone asks you to explain it, you're stuck.

This is something that you'll notice especially when using AI tools for tasks like coding. When you're using AI for coding, it's extremely easy to fall into the habit of accepting the code without understanding a single line. Unfortunately, that comes to bite you when something breaks. So, I've been using this prompt I found on Reddit to actually understand the code Claude spits out:

Can you explain what you generated in detail:

1. What is the purpose of this section?

2. How does it work step-by-step?

3. What alternatives did you consider and why did you choose this one?

Now, this doesn't need to be just limited to coding! Anytime Claude clearly reasoned through something to give you an answer, ask it to show its work. A simple "why did you choose this approach, and what alternatives did you consider?" goes a long, long way.

Asking Claude to plan first before beginning a complex task

Plan mode, but for regular conversations

I've been using Claude Code a lot lately, and one of my favorite features is its Plan mode. This is why one of my favorite prompts to use when using the normal Claude chat is basically recreating that same energy with a prompt like:

Without this, Claude dives right into the task you've given it. More often than not, you get an output that'll require significant manual tweaking. With this prompt, though, the chances of you getting something that actually feels structured and thought through are much, much higher.

Being intentional is key

Many believe that the longer the prompt, the better your answer will be. That statement only holds true when your prompt genuinely adds content. Length for the sake of length is only going to give Claude more to misinterpret.