Yesterday, I did a workshop to help a company’s product and design team improve their productivity by learning how to prompt better. I’m not sure if I can share the workshop recording, so I’ll double check before I do.

Till then I want to share this one slide. It captured everything I was trying to communicate in the workshop.

People attend prompt engineering workshops expecting to learn a bunch of tips, wordsmith tricks, and phrasing hacks. The idea being that if we can just articulate the right combination of words the perfect prompt will magically do all the heavy lifting for you.

I believe this is true, but only up to a point. In much the same way that writing polite emails with clear instructions and plenty of context will help you get more out a smart intern or virtual assistant.

I think prompt engineering goes beyond phrasing. My experience has been that coming up with clear examples of what you want is as effective, if not more effective, than spending time on prompt articulation.

Another important aspect of getting great results out of ChatGPT (and Claude) has involved using it to analyze it’s own output. Using different perspectives or frameworks to reason about output and then fold these suggestions back into the process results in significantly better work at the end of the day.

The last component to great prompt engineering for me has been figuring out knowing when to split a prompt out into lots of smaller, but more focused prompts, that can be chained together. We now have the luxury of massive context windows (Claude 2 has a ridiculous 200K tokens) and it’s becoming clear that tokens are not the only limiting factor to output complexity. Asking LLMs to do one thing at a time reliably produces better output that trying to fit your whole job into a single mega prompt. That said, breaking everything down into individual tasks is overkill, and there seems to be a sweet spot to ceiling of complexity a single prompt can handle well. Kind of like how it works with humans.

I think these four components are roughly as important as each other. This is highly subjective and my point of view on this will almost certainly change a month from now, but that’s where I am today.