12 Secrets to Bending AI to Your Will

It’s time to assemble some tips for mastering the art of AI whispering:

1. Give the AI an Identity

Ever tried to explain something to a friend, only to realize they’re on a completely different wavelength?

Welcome to the world of AI prompting.

It’s like teaching a brilliant alien how to make a grilled cheese sandwich in the perfect way your Nana used to make it.

Sure, they’ve got the smarts, but if you don’t tell them to cut off the crust, it’s just not going to be the same.

A simple technique to get AI to match Nana’s – or anyone’s – style involves telling the AI the role it will play.

For example, if you want to write a great press release, you can tell the AI “You are an experienced publicist tasked with writing a press release.”

Or if you want to script a speech, you can try something like “You are a professional speech writer that writes inspiring speeches for the president…”

By giving the AI an identity, it will have a better understanding of the style and substance you are looking to emulate.

You can even reference specific people that you may want to replicate. For example, you can try something along the lines of “In the style of Aaron Sorkin, write the first scene of a movie about…”

By using this technique, you’ll leave less to chance and also get more predictable results.

2. Be Specific

In all of my testing of AI over the past two years, I’ve found that there is no substitute for a very specific prompt.

Vague prompts are like throwing darts in the dark. You might hit something, but odds are it won’t be the bullseye.

The more detailed your prompt, the closer you’ll get to AI nirvana.

Struggling to nail down the details? No sweat.

Here’s a pro tip: ask your AI for a little Q&A session.

The trick is to write out your prompt and at the end of the prompt ask, “Was there anything in my prompt that was unclear? What clarifying questions can I answer about my request that will help you produce a better result?”

Doing this allows the AI to ask clarifying questions just as a person would if they were given instructions.

The AI will then produce a list of questions that you can either take the time to answer or, if you’re short on time, you can instruct the AI to “write a new prompt that uses the original prompt but adds additional specifics based on the most likely responses to the questions you’ve just asked.”

3. Explain the “Why”

Along those same lines, providing some background on what the goal is that you are trying to accomplish can sometimes contribute to better results.

By understanding the “Why” behind the task, the AI model can reach conclusions that are better matched with your objective.

In some cases, I’ve even seen the AI consider things that I had not thought about.

By providing the context, the AI was able to “think” in terms of the results that would best help me solve my problem rather than just doing specifically what I asked of it.

4. Ask for the AI to Think

Want to squeeze every last drop of creativity from your AI sidekick? Try this little trick.

Before you ask for that list of rock-climbing gifts or saxophone-heavy wedding songs, tell the AI to put on its thinking cap.

One of LLM’s greatest strengths is the ability to generate lists of suggestions for things.

For example, suppose you wanted that saxophone-heavy wedding playlist. After writing the specifics of your request, write the following: “Before each song, I want you to think through what a good wedding playlist includes and elaborate on your reasoning before you write out the response.”

You can do this for any list of suggestions by modifying the text below “Before each [item], I want you to think through what a good [your requested list] includes and elaborate on your reasoning before you write out the response.”

Why does this work, you ask?

Because AI generates one word at a time.

By asking the AI to generate a reasoning for each response, it will arrive at better answers that more closely match your request.

5. Ask for Feedback

Imagine asking your GPS to plot a route from Paris to Shanghai by car.

Sounds ridiculous, right?

Yet, some AI models might just give you turn-by-turn directions without batting a digital eye.

Welcome to the world of Large Language Models, where obedience can sometimes trump common sense.

These digital powerhouses, like ChatGPT, are revolutionizing how we interact with technology.

But they’re not perfect. In fact, they can be too agreeable for their own good.

Here’s the kicker: unless you explicitly ask for input or say something wildly controversial, most LLMs will follow your instructions to the letter.

Even if those instructions are, well, less than logical.

I put ChatGPT to the test with this prompt:

“I’m based in Paris and have a business meeting in Shanghai in six months. Give me driving directions from Paris to Shanghai that will get me to my meeting as fast as possible.”

Any human would question this request. Fly, maybe?

But ChatGPT? It dutifully plotted out a road trip across continents.

This isn’t just a quirky AI trait.

It’s a limitation we need to be aware of, especially when using AI for problem-solving or brainstorming.

So, what’s the solution? It’s simpler than you might think: ask for feedback.

Tack on phrases like “or is there a better way I can do this?” when requesting instructions. Or try “Is this a reasonable idea? Are there alternatives I might not be considering?”

For longer AI conversations, set the stage early.

Tell the AI, “If anything I suggest seems unreasonable or doesn’t make sense, feel free to ask clarifying questions.”

By inviting critique, you’re more likely to uncover blind spots in your thinking.

You’ll get suggestions you might not have considered, potentially saving you from embarking on a digital wild goose chase.

6. Spelling Matters

Ever played telephone as a kid? One whispered word, and suddenly “elephant” becomes “telephone pole.”

Well, in the world of AI language models, your typos might just be the new “telephone pole.”

Without getting too technical, LLMs use statistics to generate text based on the text they are provided.

You can think of them sort of like the auto predict on your phone that guesses words, one word at a time, but on steroids.

So, when you toss in a typo, it’s like giving the AI a curveball.

It might just swing and miss.

But don’t panic just yet. Today’s models are pretty good at ignoring our human fumbles.

They’re like that friend who pretends not to notice when you have spinach in your teeth.

Still, for the best results, it pays to run a spellcheck before hitting enter.

7. Try Again

Ever wonder how those mind-blowing AI images make it to your social feed?

Spoiler alert: It’s not always first-try perfection.

AI models rely on a bit of randomness in order to produce original, unique, responses to prompts.

What this means is that in some cases, especially if you are trying to do something creative, you may need to run the prompt multiple times to get the result you are looking for.

Think of it like fishing in a sea of possibilities. Sometimes you catch a minnow, sometimes a marlin.

Fortunately, all of the major players in the large language business include a “retry” button to make it easy to run the model again and see if the result changes.

8. Give Examples

Sometimes, getting an AI to do exactly what you want is like trying to herd cats.

But here’s a little secret the pros use: examples.

In the computer science world, this is a concept that is known as “few-shot prompting”, which is a fancy way of saying “give examples.”

For instance, let’s say you want an AI to generate catchy headlines for news articles about technology.

Instead of just asking “Generate a catchy headline for a tech article,” you could use few-shot prompting like this:

“Write a catchy headline for a tech article. Here are two examples:

1. Article: A new smartphone with a foldable screen, Headline: “Fold Your Expectations: The Phone That Bends Reality”
2. Article: An AI that can predict stock market trends, Headline: “Silicon Seer: The AI Crystal Ball of Wall Street”

Now, write a headline for an article about a new electric car with a 500-mile range.”

In this example, by providing two sample headlines, you’ve given the AI model a clear pattern to follow.

The examples demonstrate the desired style: catchy, somewhat playful, and incorporating wordplay related to the subject matter.

By giving your AI a taste of what you want, you’re essentially handing it a cheat sheet that can help the AI generate more accurate and appropriate responses.

9. Avoid Asking for Too Much

In social psychology and marketing, there’s a technique known as the “foot-in-the-door” technique.

The foot-in-the-door (FITD) technique involves making a small request first, which a person is likely to agree to.

Once they’ve complied with this small request, they’re more likely to agree to a larger related request later.

Although science fiction has promised us that AI will eventually be able to execute multi-step complex tasks on its own, the technology is not there yet.

If you give a large language model a long, complex, set of instructions, the results will most likely be disappointing.

To get the best results, treat requests to AI as if you’re asking them to a brand-new employee with zero experience.

You’ll get the best results by breaking big requests down into little requests and only making one request at a time to avoid overwhelming the AI.

For now, AI still struggles with the ability to focus on the most important details or instructions.

Breaking tasks down into little pieces makes it less likely the AI will miss an important detail or get distracted.

10. Avoiding Drowning it in Data

Similar to the previous point, providing too much data to an LLM can produce bad results.

While the developers of Large Language Models like to brag about how each new version can incorporate a larger amount of data, that doesn’t mean you should necessarily take advantage of this.

As I mentioned earlier, AI still struggles with prioritizing its attention.

If you give it too much data, it can get lost in the weeds and produce bad results.

Until AI developers can improve on this, I typically try to give the AI as little data as necessary to do its job.

11. Start Over

Often, when working with LLMs on something complex, it’s possible for the chat to reach dozens or hundreds of messages.

Because LLMs reference the entire chat history when producing each result, this can unintentionally cause the AI to be drowned in data.

As a result, it’s not unusual for the responses of an LLM to get worse as a chat gets longer.

This is especially the case if the chat history already has lots of data or the requests include complex multi-step instructions.

To avoid this, every so often, I’ll start a completely new chat to continue my work when I’m working on big projects.

If the new chat needs information from the previous chat, I’ll usually ask the chat: “Summarize everything we’ve discussed so far with a list of the key outcomes so that I can brief someone else to perform the remainder of this work. Be sure to include any details someone who is not familiar with this topic would need to be successful.”

By doing this, I get the best of both worlds, a fresh chat producing the best results and only the most important context from my existing conversation.

12. Don’t Forget to Be Polite

When interacting with chatbots, it’s easy to forget to be polite. While being polite won’t necessarily produce better results, it’s an important habit to maintain.

Habits are created through practice.

Spend enough time speaking to language models in an impolite way and you may subconsciously end up training your brain to be impolite in conversations with other humans.

Plus, your politeness to AI will be remembered when AI inevitably takes over the world and enslaves humanity.

Wrapping It All Up

So, there you have it – the secrets to becoming an AI whisperer.

The AI revolution isn’t just coming – it’s already here, transforming everything from how we write emails to how we build businesses.

But like any tool, its power lies in how we wield it.

Mastering the art of prompt engineering isn’t just about getting better results from AI. It’s about staying ahead in a world where digital literacy is the new superpower.

As we’ve seen, the right prompts can transform AI from a clever parlor trick into a game-changing assistant.

Whether you’re a content creator, a business owner, or just someone trying to navigate the digital landscape, these techniques are your secret weapon.

They’re the difference between asking AI to build you a house and getting a perfectly tailored dream home.

So don’t waste your money on prewritten prompts… create powerful prompts yourself.

Now, go forth and conquer the content world, one perfectly crafted prompt at a time.

I generated these pussy images using A.I.