Read time: 3 minutes

Hey Reader,

It is one thing to know the potential of AI Language models like ChatGPT or Claude, but it's also important to know their limitations.

In this email, I want to show you 5 things you should avoid when using ChatGPT (or Claude or Gemini or LLama, etc.)

1/ Treat it like a know-it-all search engine

ChatGPT is not a search engine. It is simply a very expensive predictive keyboard. All it does is it just guesses what's the best next word to say in a scenario.

It makes that decision primarily based on 3 things:

1. It's training data
2. Context of previous words
3. Your prompt

And one of the things that it does very well is that it sounds convincing.

It kind of just bluffs its way through an answer, like an unprepared university student on an exam.

But it doesn't know what it's saying, and can't really fact-check its own responses.

So if you treat it like it knows everything, and accept every response to be true, you might make a fatal mistake.

2/ Not giving enough context.

Context is king when it comes to Language Models. They are not mind-readers, and can't guess what you mean.

Prompt Engineering is really all about effective communication.

What helped me a lot when it comes to giving context to ChatGPT is asking myself this:

Then, tell ChatGPT that information. If you find yourself repeating this information over and over again, it's probably time to set up a Custom GPT for that task so you can save more time.

3/ Not treating it as a conversation partner.

It's name is ChatGPT. Not MonologueGPT. Go CHAT with it. It's a great partner for engaging in brainstorming sessions, but also when it comes to tasks.

So try to switch your mind from "Prompting" to "Chatting". This way, you can guide ChatGPT to the treasure, which might be a document that would be impossible to get from one prompt.

A life-changer for this is to include this in your ChatGPT prompt:

This prompts it to instead of answering, asking questions. Now, you can go ahead and answer these questions, or you can say:

Now you are making it feed its own context window, which is proven to enhance the responses it gives you.

This is what we call Persona Switching, one of the many Conversation Design methods that we teach our students. With these, you can get days of work done in hours (on the right tasks of course)

4/ Giving too much context.

Megaprompts decrease performance. The longer you prompt is the less likely it is that you will get a high-quality response.

This graph perfectly shows how adding more and more documents into the Input Context (a.k.a the prompt), decreases the response accuracy.

So you can't be like:

Well, you can (if it fits the input window) but you're going to get a pretty useless response that will probably leave out a lot of crucial details.

And sometimes, when you're configuring a chatbot that — for the love of God — you can't get it to do (or not do) what you want, the solution is not to add more stuff to your prompt but to remove from it.

Brevity. Say what you mean, then shut up.

5/ Relying on it without human oversight

And the final thing to avoid when using ChatGPT is relying on it for sensitive or critical tasks without human oversight.

Language Models make mistakes and should not be used as the last touchpoint for important matters.

For critical tasks, ChatGPT's or Claude's errors could have serious consequences.

AI models don't have real-world experience or common sense reasoning. They might miss important nuances or context that a human would easily grasp.

Instead, use ChatGPT as a tool to augment human intelligence, not replace it. Here are some of my best practices:

1. Always review and verify AI-generated content.
2. Use ChatGPT for ideation and drafts, not the final output.
3. When in doubt, consult an expert.
4. In automations, review and approve responses by a human.
5. Educate your team about the capabilities and limitations of AI.

+1/ Not giving them a tip.

Okay, this is a funny one, but actually, ChatGPT tends to give longer, more verbose responses when offered a tip.

So you can either say:

Or

The reason why this works is because the generated words (tokens) that ChatGPT says while generating a response actually influence what comes after it.

And offering a tip just makes the responses slightly longer, long enough to slightly enhance the response because the model explains its thinking before providing a final answer.

Talk soon,

Dave

Dave Talas CEO & Founder Promptmaster

P.S: Whenever you're ready, here are 4 ways I can help you:

#1: Learn LLM prompt engineering and how to build AI agents without writing any code in my Prompt Master AI Course. With over 1,400 satisfied students, this course makes you a Prompt Master. Taught at 2 universities as an accredited course.

#2: Turn ADHD into your superpower, and activate your hyperfocus on demand with the Cheatcode Bootcamp. First cohort begins September 9th (and that's when doors close too).

#3: Book me as a workshop instructor or keynote speaker for your company, your team or for your audience. I've ran many workshops on ChatGPT prompt engineering, AI for executives, use-case specific topics like AI in HR and PR, and AI agent building.

#4: Want my best students to build AI automations and agents for you? You can hire us for work, whether you have a specific problem to solve or you need someone to figure out what's a problem worth solving with AI.