It only took a few months for prompt engineering to evolve from a practice few understood to one of the most in-demand skills in the modern workforce.
Since the runaway adoption of OpenAI’s ChatGPT this past November, the ability to engineer prompts and use generative AI tools has quickly become “the most important job skill of this century.”
Already, there are job postings promising six-figure salaries for successful prompt engineers, and many roles now require a basic understanding of how to engineer prompts.
Experts say that communicating effectively with AI will soon become an “overlay for not only search engines, but also creative work, busywork, memo writing, research, homework, sketching, outlining, storyboarding, and teaching.”
But despite its widespread popularity, some are still in the dark when it comes to engineering prompts in ChatGPT and other AI tools.
The good news: you don’t need to write code (although having some technical expertise helps) to communicate effectively with AI.
“The hottest new programming language is English,” Andrej Karpathy, Tesla’s former chief of AI, recently joked in a Tweet.
Professional prompt engineers are notoriously protective of the prompts they create. But to get the most out of Bloks and similar tools, it pays to know how to write and refine AI text prompts to get the results you want.
Whether you’re a total newbie, a self-taught pro looking to level up, or someone who's simply curious, here are some prompt engineering best practices and practical tips to keep in mind when creating prompts.
1. Be Descriptive
The biggest mistake people make when writing their first prompts is being too vague.
Ask an AI tool like Bloks to write you a “cold sales email,” and what comes out will be long-winded, generic, and meaningless—sort of like this.
But, in the words of ‘90s one-hit wonders New Radicals, with ChatGPT and other tools, you get what you give.
Being specific, descriptive, and as detailed as possible will lead to better results when using generative AI tools.
Focus on the tone, output length, format, intended audience, and any other information you think is worth including to provide the system with enough context to reach your desired outcome.
Now, we’re getting somewhere.
2. Get rid of fluffy, unnecessary words
As much as you want to be descriptive, you don’t want to go overboard when writing your prompts because it could send the AI down the wrong path.
Here’s an example I’ve adapted from OpenAI:
Say you’re a marketer creating a product description for a ceramic coffee filter.
You make a prompt that says: “Write a description for a ceramic coffee filter. The description for this product should be fairly short, a few sentences only, and not too much more.”
The description will be decent, but anytime you regenerate a response, the description’s number of sentences and sentence length will fluctuate.
If you want the output to be more precise, change the second sentence to “describe this product in three sentences,” and you’ll get exactly that—a three-sentence product description.
That may seem like a small change, but the results can be widely different if you leave the language in your prompts up to interpretation.
Ambiguous language—like, in the example above, “fairly short,” “a few sentences,” and “not too much more”—opens AI tools up to endless possibilities.
Removing adverbs (e.g., “fairly”), adding constraints (e.g., “three sentences” instead of “a few sentences”), and reducing ambiguity (e.g., “not too much”) will make your output more precise and concise.
Need a tip that's simpler? Prompt engineer Anna Bernstein says to contain the main focus of your task within the verb. So, instead of saying, “rewrite this to be shorter,” write, “condense this.”
3. Don’t Be Afraid to Roleplay
This is one of my favourite strategies, especially if you want to spice up your writing (like emails, blogs, social media posts, etc.) or need help understanding a complex topic.
When prompting Bloks, ChatGPT, or another large language model, give it a role and clarify the output’s intended audience.
This is incredibly helpful when I hear words or phrases in meetings I don’t understand.
For instance, working at a startup, I hear a lot of technical terms.
When Bloks spits out our meeting transcriptions and summaries, I can copy a phrase, prime the AI by giving it a role, audience, and additional context, and ask it to explain it in layperson’s terms, so I know what people are talking about without doing research outside of the app or asking someone to explain it to me.
Here’s an example:
Someone in our weekly meeting was talking about Python the other day. I had no idea what that was (being a bit of a Luddite, I assumed some type of programming language).
My initial prompt didn’t give me the answers I was looking for…
So, I gave the AI a role and set up a scenario to help it explain what Python was to me in basic terms…
This strategy can be used in lots of helpful—even ridiculous—scenarios outside of simple questions and answers.
4. Get Your Prompts in Order
To be fair, this is more important to consider when using AI that generates images, but the order in which you present information in a prompt impacts the output in subtle ways.
Take Midjourney, for instance.
If you write a prompt that says, “Maverick and Goose from Top Gun high-fiving while playing volleyball,” you’ll get something like this…
But, if you move the word “goose” after the word “high-fiving,” you’ll probably get this…
The order matters. Microsoft suggests “telling the model the task you want it to do at the beginning of the prompt, before sharing additional contextual information or examples” to get higher-quality outputs.
And, to prevent recency bias, which large language models can be susceptible to, try repeating instructions at the end of your prompt to emphasize the main task you want to be completed.
Afterward, check the output and adjust it if necessary.
5. Add Clear Syntax
Unlike AI tools that generate images, like DALL-E and Midjourney, it’s better to use clear syntax when creating your prompts instead of listing off a bunch of keywords.
Use proper punctuation, section markers, and headings, so it’s easier for tools like ChatGPT or Bloks to understand your prompt.
6. Provide More Context
When engineering any kind of prompt, the more context you can provide, the better.
In fact, all of the tips above, in some ways, are about providing context—giving the AI enough (and the right types of) information to help it complete the task successfully.
When structuring your prompts, the input you provide makes all the difference, and prompt engineers often rely on at least one of the following strategies—priming (mentioned in passing above), one-shot prompting, and few-shot prompting—to get what they’re looking for.
Using the tips I’ve already written about above, I prompted ChatGPT and asked it to explain each concept I just mentioned in more detail.
Here’s what it said (read it out loud in a robotic voice if you want to feel like you’re in an episode of Black Mirror):
Zero-shot prompting: In a zero-shot prompt scenario, you give the AI model a new task without any prior examples or context. It's like a cold call, where the model relies solely on its pre-existing training. For example, you might ask, “Can you create a checklist of key metrics an early-stage startup CEO should track regularly?”
Priming: Priming sets the stage by giving the AI model a specific role or background, helping generate more relevant responses. For instance, “Imagine you're a successful entrepreneur who has launched multiple successful startups. Based on your experience, what are the most common pitfalls an early-stage startup CEO should avoid?”
One-shot prompting: In this method, you provide the AI model with a single example of the task before asking it to perform the task. For example, “Here's an example of a successful elevator pitch: 'Our company, X, has developed a unique AI solution that can double the efficiency of logistics operations. We're seeking $2M in funding to expand our market reach.' Now, can you help create an elevator pitch for a startup offering an innovative renewable energy solution?”
Few-shot prompting: Similar to one-shot prompting, but with multiple examples to better illustrate the task. This could look like, “Here are three examples of mission statements from successful tech startups: 1) 'To make the world more open and connected.' - Facebook. 2) 'To accelerate the advent of sustainable transport.' - Tesla. 3) 'To organize the world's information and make it universally accessible and useful.' - Google. Now, can you help create a mission statement for an early-stage startup focused on AI in healthcare?”
Using Conversational Context: If you're in a conversation with the AI, prior interactions are used as context. For example, if you ask, “What are some effective growth strategies for early-stage startups?” and then follow up with, “Which of these strategies is best suited for a SaaS company?”, the model understands that “these strategies” refer to the growth strategies previously mentioned.
Chances are you won’t get what you want on your first try.
A lot of prompt engineering is trial and error. You’ll need to write your prompt, see the results, and refine your prompt to get what you’re requesting or picturing in your head.
When I asked ChatGPT if it had any tips for refining prompts, it suggested the following:
- Trying multiple prompt variations to identify which phrasing or approach yields the best results
- Reviewing past conversations with the AI to identify successful and unsuccessful prompt strategies
- Asking follow-up questions to clarify, expand, or narrow down the AI's responses
- Leveraging diverse perspectives from co-workers to further refine your prompts and adjust their clarity and focus
But truthfully, it’s a lot of guesswork.
8. Be Prepared to Forget Everything You Just Learned
AI is evolving so fast it’s possible that these kinds of text-heavy strategies won’t exist in a few years or even a few months.
Just look at how the output from the same prompt—“Captain Jack Sparrow”—has evolved in Midjourney from V1 to V5.
Until then, though, it's important to know the fundamentals.
But contrary to what the hucksters selling e-books with titles like “1001 Prompts That Will MAKE YOU RICH” for $4.99 suggest, the average person doesn’t need to be an expert at engineering prompts to get the most out of ChatGPT and other AI tools like Bloks. You can get by with just the basics.
And with that, we hope you enjoyed this guide.
And, if you still haven’t downloaded Bloks, request access (it’s free).