21 words: The Complex Simplicity of Prompting
Google Gemini just put out a new prompt guide.
And it reveals something amazing that changes how we think about prompting. (I’ll get to that.)
But let’s start with the guide itself.
It’s actually notable that it exists at all. No big players (OpenAI, etc) have put out a user guide.
Weird, right?
I mean, everything has a user guide. My dog brush came with instructions.
But LLMs are bizarro. They don’t really act like software. There’s no Step One, Step Two kind of thing. There’s no learning curve at all - not really. You just talk to it like a human.
So how did Google approach this?
Gemini’s user guide is 45 pages long.
But it could have been 2 pages. Or 4000 pages.
That’s because prompting is simple in concept. And complicated in practice.
It’s what makes training, learning, and implementing genAI in companies so difficult.
So without further ado, let’s dive in to see what Google can teach us (and what this whole big revelation is…)
The guide gets right to the point. There are Four Elements to any good prompt:
Role, Task, Context, Format.
We know that part already, but okay. Cool.
Then it says something really interesting:
The most successful prompts average around 21 words.
Twenty one words.
This flies in the face of every prompt library I’ve ever seen.
It also is a much-needed lesson to the AI Influencers who demand that you Copy-and-Paste-My-Prompt!!! even though the chances of you needing that prompt at exactly that moment are about one-in-infinity.
But wait. There’s more.
Google finishes that sentence above by saying:
…yet prompts people often try without knowing this are short - usually less than nine words.
Awkward phrasing aside - what does this tell us?
It tells us people are not, as a general rule, entering crazy long prompts into Gemini (or other LLMs like ChatGPT).
And yet!
There is still this infuriating copy-and-paste-this-prompt army shouting that the best outputs come from greater complexity.
Listen - I need to state this very clearly.
I am NOT saying that long prompts are bad. They aren’t.
In fact, giving LLMs more context, more examples, and defining the output in more detail - those will only improve the output.
My problem is much different.
My problem is that these people telling you to that you need long prompts, and special skills, and to take classes - they are harming the human race.
That’s because the majority of humans are still not using generative AI, when it could literally be improving their lives.
The reason why they are not is because they think they need to know something that they do not need to know.
It’s like telling somebody at risk of heart disease that they have to learn a 40 step process to start exercising the right way.
With generative AI, a big part of the thing holding people back is that they think they have to learn how to prompt. They think it’s a science.
Which it is NOT.
In fact, Google says this:
Prompting is an art. You will likely need to try a few different approaches for your prompt if you don't get your desired outcome the first time. Based on what we have learned during our Workspace Labs program, the most successful prompts average around 21 words…
You see that?
Prompting is an art. Just like conversation is an art.
So what DO you need to know?
There are two steps to creating a great prompt:
1. Decide what you actually need
Now, this is easier said than done.
Why?
Because your brain is too busy trying to figure out How Can This Weird New Thing Help Me.
Best practice?
Just talk to it like a person who you believe can help you. If you have a finance question, talk to it like its your broker. If you have a gardening question, like it’s that gardener from TV. (I assume TV has one of those?)
Forget what you think ChatGPT or Gemini or Claude can do. Just ask it for what you need.
2. State the question, then add on
Don’t worry about ‘how to prompt.’ Just state your question.
Then give it a little more context. Give it some examples. If you’re asking a marketing question, write an example of a Nike ad or something that you want it to emulate.
You can definitely tell it to act in a certain role - or tell it about your role (teacher, engineer, etc).
But all you’re doing when you do that is giving it more context about how you want it to answer. Assigning a role is just a short cut for that.
And that’s it. A prompt can be short or long. It doesn’t have to be one or the other.
If you’re using ChatGPT, or Gemini, or Claude on a regular basis, you can’t help but get better. Keep at it!
We got this, friends.
AI NEWS OF THE WEEK
AI coming for analysts
The NY Times has a great piece on how senior leaders at Goldman Sachs and Morgan Stanley are openly contemplating what the future of an analyst role looks like due to the capabilities of generative AI.
Jamie Dimon says AI is the new….everything
In his annual shareholder letter, JPMorgan Chase CEO Jamie Dimon compares AI it to the printing press, steam engine, electricity, computing and the internet. The implications for banking and jobs is huge.
Google making us better looking
AI editing tools are coming to Google photos. I love this kind of news, because it shows more and more we don’t need to venture off to new tools - companies are building AI in to their own products that you use every day.
Generative AI Tips
I’m finding of late that a combination approach with GPT4 and Claude is working really well for me - here’s how I use it.
If I’m working on a framework, or step by step instructions, or ways to remember something, I run that rough draft thought GPT.
It’s a fantastic reasoning tool, and frames things into great, short guides.
But then Claude 3 really shines - I then take that and run the whole thing through Claude with a sample of how I write. That puts it into my brand and tone in a way that really is way better than ChatGPT these days.
Experiment for yourselves!
That's all for now, friends! See you next time.
