Prompt engineering is dead đŸ˜”

Long live prompt engineering!

In partnership with

Hello, AI marketers!

Six months ago, “Prompt Engineer” was billed as a guaranteed six-figure job for the future.

Today, it’s looking a lot more likely that AI models will become so good at reasoning that complex prompting will become redundant. 

But that’s no reason to dismiss the importance of prompt engineering entirely. 

If you don’t put some effort into prompting, you won’t get the most out of your outputs and will probably waste your time in the process. 

I’ve carefully watched developments in prompt engineering over the last two years.

Read on, and I’ll share my thoughts on prompt engineering's current situation and future, helping you get more juice out of your AI-powered workflows.

Prompt Paradoxes đŸŽČ

VMware researchers recently conducted a systematic study on the impact of different prompt engineering strategies on an LLM's performance in solving grade-school math problems.

The results were surprising: no consistent trends emerged. 

Even popular techniques like chain-of-thought prompting were hit-or-miss. The researchers concluded that prompt performance is highly unpredictable and specific to each model and task.

This study challenges the notion that human-crafted prompts are the gold standard. 

But this doesn't mean prompt engineering is dead – rather, it's evolving.

The Rise of Autoprompts đŸ€–

If human-crafted prompts are unreliable, what's the alternative? 

Using AI models to optimize AI models could be the answer. 

The key idea here is to use AI models to self-define their own prompts and refine them over time. 

Intel Labs' NeuroPrompts tool is an intriguing example. It uses a language model to transform simple prompts into expert-level prompts, then applies reinforcement learning to optimize the quality of results. 

These auto-generated prompts consistently outperform expert human prompts, but some are completely bizarre. 

“I literally could not believe some of the stuff that it generated,” Rick Battle from VMware told IEEE Spectrum. 

“In one instance, the prompt was just an extended Star Trek reference: “Command, we need you to plot a course through this turbulence and locate the source of the anomaly. Use all available data and your expertise to guide us through this challenging situation.” 

Ultimately, you don't need to be a researcher or Captain Kirk to leverage auto prompts.

For example, many prompt generator tools are now available as GPTs.

Prompt-making GPTs

Or, you can simply give an AI model natural language instructions, ask it to convert them into a prompt, critique it, and optimize it. 

Below is my example of asking Claude 3 Opus to create and optimize a blog post prompt for professional but engaging content. I found it interesting how it added information about the reading level and Flesch score. 

A self-generated prompt from Claude 3 Opus, which produced good results in my tests

Sure, I’ve got no fancy metrics to demonstrate how effective it is, but it certainly beats prompting the model using your own natural language – and it takes just seconds, too. 

Creative Prompting Techniques That Still Work 🎹

While automated prompt optimization is the future, I believe creative human prompting will remain crucial for a while.  

In the long term, yes, we’ll find more ways to optimize AI models using math rather than words. After all, these are math-based systems.

I’m sure sophisticated auto-prompting ‘co-pilots’ will become accessible soon, too, so keep your eyes out.

In the meantime, here are some of my favorite prompting techniques that I still use regularly:

  • Chain-of-Thought Prompting: Asking models to explain their reasoning step-by-step can boost performance on complex tasks.

  • Persona Roleplaying: Prompts that put the AI into a specific persona or scenario can elicit better responses, e.g., “a world-renowned PhD researcher in [subject].”

  • Positive Self-Talk: Giving models encouraging prompts like "You're doing great!" has been shown to improve results. Indeed, while it can be tempting to trash-talk an LLM when you’re frustrated, this could materially produce worse results!

  • Bribes: This might be considered a bit old-school now, but I still think it works in some situations. You quite literally offer a gift or cash incentive for the AI model to complete the task to the best of its ability.

  • Regrouping: If you encounter a situation where the model doesn’t follow instructions well, ask it to summarize its understanding of the conversation and what you’re trying to achieve. 

  • Wording Prompts Correctly: Research has shown that LLMs provide biased or inferior results when prompted with poor spelling, grammar, dialect, or slang. Use AI to polish up your prompts before using them.

So don't be afraid to be ambitious with your prompts. Give the model room to interpret, create, and innovate. You might be surprised by the results

Stay tuned for more dispatches from the frontiers of AI marketing.

Join the Essential Community For Marketers

Stay ahead of the curve and join the thousands of marketers shaping the future of marketing with the American Marketing Association. As the largest community-based marketing organization, you’ll find award-winning content, professional certifications (PCM¼), industry-leading training events, and vibrant local chapters. Members get access to an abundance of resources, on-demand courses, and the most essential community in marketing.

Here are a few great AI marketing resources and tools I’ve encountered recently. Enjoy!

1. Contrarian Content is King

Here's another Neil Patel analysis. This one looked at the impacts of taking a contrarian view of the norm, like “Content is Not King,” which happened to drive 44% more traffic than baselines. This has been observed before with negatively worded blog headings, which also seem to attract more traffic. 

Text-to-audio models by Udio have taken social media by storm, turning your text prompts into realistic, natural music with lyrics. However, it’s also raising massive ethical and copyright concerns as users proceed to generate AI music that sounds like famous hits. What happens when one of these spits a Taylor Swift or Ed Sheeran melody or lyric?!

An awesome post that charts how companies are using AI in marketing right now, from optimizing content to audiences to brainstorming and analyzing competitors. I found a few new ideas here, so it’s definitely worth checking out. 

What did you think of today’s email? Loved it? Hated it? Hit reply and tell me why.

Until next Thursday! Happy marketing.

The AI Marketer