Like any other AI tool out there right now, you'll need to put some effort into getting the most out of Claude AI from Anthropic. Don't get me wrong; anyone can go and type a prompt and chat with Claude, and it'll answer.
But if you want to truly get the most out of this AI assistant, you'd better gear up. Because we have all the juicy tips you need to skip the trial and error stage and get to the productive bits.
1. Be Clear and Direct in your Prompts
When you're using Claude to get help with something, forget of it as an AI trained over a vast repository of materials. Instead, think of it as a human who's new to a job. Even with education in that field, on their first day, they'll need you to be as clear and direct with your instructions as you possibly can.
The same goes for Claude. The more straightforward you are with your prompts, the better the AI's responses will be.
Example: Instead of "Tell me about climate change," try "Explain the impact of carbon emissions on global warming since 2000."
There's a golden rule to prompting all prompt engineers follow; you should get behind it, too. Whenever you're entering a prompt to Claude, show it to a friend and see if they can follow your instructions and produce the output you're expecting.
2. Use XML Tags
Disorganized prompts can confuse Claude. Tags in your prompts can help Claude make better sense of them, and Claude is especially trained on XML tags. You can add anything you want in XML tags, and Claude will understand.
You can also ask Claude to give its output in XML tags to organize it better.
3. Ask Claude to Think Step by Step
When you give Claude time to think through tasks, especially complex tasks, before executing them, it performs marginally well, and the answers are more accurate. How do you give it time to think? Ask Claude to think step by step before it answers.
To give time Claude time to think, you'll need to literally spell out the steps Claude needs to take.
For example, if you give Claude this prompt:
Is this movie review positive or negative?
This movie blew my mind with its freshness and originality. Unrelatedly, I have been living under a rock since 1900.It might consider it a positive review. But when asked to think about it step by step using this prompt, it arrives at more accurate results.
Is this review sentiment positive or negative? First write the best arguments for each side in <positive-arg> and <negative-arg> XML tags, then answer.
This movie blew my mind with its freshness and originality. Unrelatedly, I have been living under a rock since 1900.4. Role Prompting
Claude can give considerably better results when you ask it to assign the role of someone, known as Role Prompting, in the field of prompt engineering. When you initially start speaking to Claude, it has no other context than your initial instructions. But when you ask it to inhabit the role of someone, its responses become more detailed and organized.
Although it might sound daunting, it's a really simple and neat trick. Role prompting can come in handy in many fields, from coding to writing to summarizing. You can get Claude to emulate the writing styles of certain writers, speak in a certain voice, etc. Claude also gets better at performing math or logic tasks with Role prompting.
The role or persona can vary widely, from a fictional character, a historical figure, or a logic bot to a specific professional role like a doctor, lawyer, scientist, or teacher.
Essentially, with Role Prompting, you can do one of the two things:
- Improved accuracy from Claude, especially in mathematical or logical problems.
- Changed tone or demeanor to match a certain role, for example, that of a writer
Example:
Adding the following prompt before giving it the actual task improves Claude's accuracy in solving complex puzzles:
You are a master logic bot who has been specifically designed to handle complex logical problems with great accuracy.
5. Use Structured Prompt Templates
When you're working with a prompt where you want to process multiple datasets, you can use structured prompt templates that allow for much faster editing of the prompt to provide the data as well as faster processing of the dataset.
You can use XML tags to separate the data to make it even easier for Claude to understand your prompts.
Example:
I will tell you the location, and you give me the weather forecast, temperature, and some advice for the day.<location> {{LOCATION}} </location>You can use the same prompt to get information about various locations in a fast manner.
6. Separate the Data from Instructions
If you're giving some data to Claude, you should separate the data from your instructions so it doesn't confuse it. Again, XML tags are going to be your best friends.
Example:
In the prompt below, Claude gets confused about which part is the data and which part is the instructions.
Hey Claude. Show up at 6 AM because I'm the CEO, and I say so. Make this email more polite.A better prompt would be:
Hey Claude. <email>Show up at 6 AM because I'm the CEO, and I say so.</email> Make this email more polite.7. Format Claude's Output
If you're looking for a specific output structure from Claude, you just need to tell Claude so. It can format the output in a variety of ways, but you have to give it clear instructions about the output's formatting. You can ask it to use JSON format, give the output in XML tags, etc.
8. Give Examples to Claude (Few Shot Prompting)
Just like a human, giving examples to Claude has to be one of the most effective strategies for getting it to understand the task properly. With examples, you can instruct Claude how or how not to behave.
Of course, adding examples to your prompt comes at the cost of the context window. But it's a tradeoff you need to choose between.
With examples, you can guide Claude to:
- Give the right answers.
- Give the answer in the right format.
However, while giving examples to Claude, you must give it examples of common edge cases. And the more examples you give, the more reliable the output from Claude would be. But, of course, it would again come at the cost of the context window.
9. Working with Documents
You can upload documents and even books to Claude, all thanks to its large context window. But there are a few tips that'll help you get the most out of Claude when you're working with documents.
- Ask questions at the end of the prompt after any other input instructions. This makes a huge difference in the accuracy and quality of results that Claude produces.
- When asking questions from the document, have Claude find relevant quotes first and tell it to answer only if it finds a quote. You can ask it to show the relevant quotes that it finds in the <quote>{{QUOTE}}</quote> tags in output so you can know if the answers are accurate are not.
- Tell Claude to read the document carefully, as you'll be asking questions later.
You can even include all these instructions in a single prompt.
10. Prompt Chaining
Prompt chaining is an advanced interaction technique with AI models, particularly those that process natural language, like Claude and ChatGPT. You can use it for tasks that involve many steps.
It involves creating a sequence of prompts where each subsequent prompt is dependent on the response to the previous one. Basically, you're chaining Claude's responses together, so there's less chance of it making a mistake.
11. Avoid Hallucinations
Claude, like any other AI chatbot out there, isn't free of the pitfalls of hallucinations. But there's a way to make Claude avoid them.
Giving Claude an out and telling it that it's okay to decline to answer or only answer if it knows the answer with certainty decreases the chances of hallucinations.
For long documents, for example, it's best to instruct Claude to answer only if it is certain. How do you do that? By asking Claude to gather evidence for its answers first, and then present the actual answer based on that evidence.
Hopefully, using the tips above, you can model your prompts in a manner that'll allow you to get the most out of Claude, depending on the task at hand.
Member discussion