Hacks When Using ChatGPT for No Priors
14:34 Share

Welcome to the podcast today. We wants to tell you about something really crazy that was recently released, ed, and this is essentially open the eye themselves. So you know you're hearing IT straight from straight from the beast um has released a guide to prompt engineering essentially if so, essentially telling us exactly how you try to get the best results out of IT.

Jamie and I have kind of broken this down into like a very concise there is massive blog post. You go on their website, we might link in in the shown notes. Otherwise just go look OpenAI prompted eran guide google and it'll be the top result um is a massive blog posting guide.

We consolidate this down to the six core areas that are important on this. I'm not going to give you a couple bullet points and use cases of us both using ChatGPT and how we use this. This is black, literally completely game changing for how I use chat P, T, and how my responses are.

I'm going from like mediocre responses that I have to like heavily edit to all the sudden, like a completely does the task for me. I don't have to touch IT. That's goal. And so a lot of these is really how you get there. So we're we're going to break IT down but jammy want you kick us off with the first form yeah so I mean.

I think um with when that comes to ChatGPT, I feel like there's a lot of opportunity and and there a need for education in space because um when you just if you just open up the program and you start typing prompts, you may I get the results you want uh unless you know kind of know what the process is and he kind of get inside the mind of uh the GPT so um yeah so you need to think about like where is information coming from um and then how is that going to build you know the thing you want. So um there's like a little workflow we're just going to walk through. The first one is obviously get a right clear instruction. So um you know if you don't tell the model exactly what you want and if you're not a detail, it's going to a kind of just start jumbling stuff together and you may get something that is kindly visible but generally generally not you know i've tried to um you know just give you a quick um you know right me a blog post on ah what not which is you know an APP i'm stern to use lot more reselling up and IT doesn't even know what IT is a kind of just B S S himself so you to get to be very clear about what exactly you .

want from IT yeah one hundred percent. They they have a couple of tactics. They mentioned that I think they are really important along with what Jamie said.

Um and the first one is to or are one of them that they mentioned specifically that I like is they talk about to specify the desired length of your output. So I want to talk about that because i've had mixed responses. I don't know if you're in the camp jme pretty.

What I found is chat. Beauty cannot give you in one answer more than like a thousand words. Sometimes I can push IT and bully IT into giving me twelve hundred words, maybe fifteen hundred of the max i've ever gotten out of, like, you know, once in a blue moon.

But really, if like, thousand words is usually a long response. So I might give me a blog post, give me an article, tell me about this thing is because, as lengthy as possible, a thousand worries like the absolute max. So that's something that I think is interesting.

Um the other thing though that I that that has helped me so if i'm like, if I say like word can't like make IT like two thousand words or five hundred words is terrible at doing work count. I don't think I don't think I can do work count. I don't jv have you've ever bought IT to work count for you?

Not to the exact act to be honest. I I haven't taken the time to actually count or see their self O.

I get like in those word counters and put IT in there. I don't care if if it's like twenty off, but I wanted to to be roughly i'm like give me like a two thousand word article and it's like seven hundred words and I come down that this is what I found out to be the the best is to count by sentences or paragraphs. So I might give me three paragraphs about this or give me six sentences about this. I can do sentences in para counts, I found. So that's just a little protest as something that i've actually be able to get IT to do.

One question I have about this before I was looking at IT has provided examples. Could you give IT, you know, from you and I could you give IT a exact U R, L from another web page to crawl? Is that is not capable that yet?

I'm not one hundred percent. Earlier this year, I was able to do that. If you had like browse with a thing, I think you might be able to do that now, to be honest. Yeah, give IT like a what I also think though, when IT says provide examples, what i've um done successfully in the past as well is just like if it's if they're something on the right page, copy all the text from the web page and you also that's that's a protium too when you caught like let's say you want to um yeah they are like a web page in there and get IT to write something about IT or give you some commentary on IT. You don't have to like clean that data so like you copy the whole web page just got like you know the text for like ads on the website and like images and like all text and like the comments section, like you could copy the whole thing and throw IT in there and it's going to be vant. F that's that's one thing I do love about IT is i'm not like back space in relevant and like titles and like random like that like jumble text in there um but as far as providing examples, one thing that I have very successfully done is what i'm trying to get IT to talk like in my same voice like so sounds like me not just like chat t like alter professional um one thing I do is I will grab like a block of text of like my writing the from an article i've done or something throw that in there and be like copy the exact tone and style of this clip this is me and like write IT like this and I can do a lot Better I don't know like I don't know exact prompt engineering like keywords to make IT talk like me like I talk like this and I say things like that you know like I or if I give an exact example .

you can copy me nice okay so that can allees under the second uh strategy IT gives which is provide reference text um you know with the copy pasting is a great uh example um yeah what other strategy have you used there?

Yeah so something that they mention is to tell tragic t the answer with citations tions um from a reference text that I think is super, super interesting. Um one of the thing that I think is like that i've done a law is you know going back to earlier when I was like I can't do like really long things, there is a strategy to make you do wrong things, but also this is a strategy for getting to do complex step.

So if you have a complex task with multiple steps, of course, for those that no, i'm currently building A I box, which is uh going to help people automate their workflows so you can like um take those multiple steps, put them into a workflow on our platform and habit, like run through all, that's going to be super nice. But because this is not launched yet, in the meantime what I do um is essentially all sometimes is like IT like a elaborate process i'll have like and I need my like I get my va to do that right now because my sock is not done. I'll break IT into like different um smaller steps.

So um i'll have like i'll like five prompts. It's like the first prompt is like, okay, let's say it's like an article i'm generating. The first prompt is like take this keyword turn into an article title, then the next take this article title and um you know right like an article outline and the next one is take this article outline and write an article and the next one is like take this article and write some freak last questions about and the next one is like take this article and right like ten misconceptions people have about IT.

The next one take this article in all this stuff and like also add like some tables of like interesting data that you can add into there. There's like that's like A A worker I ve used for, like some article generating A I tools. But if you just try to give IT like a giant prompt that said, right me, an article with all of these things in IT, there is no way I would get close. I think IT IT maxes out when I after like five or six bullet points of things to do IT just starts for getting bullet points. So the way I found to actually get IT to do IT is you just break IT down, just small pieces.

Yeah no, that's a that's a great point for sure. Then that that kind moves on to our next one, which is to split up complex task. So if you tell me to do too much at once, you're not, you not can be happy with the result.

You get a kind of break IT down into small section, small pieces of data that you want to all include. Kind like you said, I will definitely forget lots of things. So yeah yeah .

no hundred percent. The other thing we wanted to talk about is, you know hope it's kind of funny opening eyes, like give the model time to think. I mean, I think this kind of goes without saying but like we know obviously IT can take a really long time to uh to get your responses sometimes with IT um but one thing they say to do is to instruct ChatGPT to work out its own solutions before coming up with a conclusion.

So it's like you you can say like, hey, um I have this problem, i'm trying to find the solution and I maybe just like science or math like or something in your business, right? And you can be like I I need you to do like X, Y and z, like give me the response, take some time to think about IT and work out the problem before you give IT to me, this is so fascinating to me because if you just said here's the problem, give me the solution. It's gna use like it's it's like um especially these.

Or just like predicting the next word in the sentence. So it's just gonna predict through IT. Now chat B T has some software built into let IT do some logic and reasoning, but IT doesn't always know to trigger that software and that logic unless you can tell IT.

So it's like I have this problem this one going have to think about like work IT out first like write your step by step of why you're going to do this um and you're gonna Better results and if you just ask you to give me the response is just going to yeah the response might not be good. But if you tell that to like work through all the steps to get there, it's going to you're going to actually make sure that IT IT does all the correct steps and IT gets to the right output. Is that not fascinating?

Yeah that's crazy. And that kind of goes back to what we talked about last week. You um you're writing a book with A I knew you wanted IT to ask you certain you instructed to ask you questions are that I needed to put together the story so um ah I kind of giving IT like time to think in time to collect information yeah that's pretty wild to think about but also yeah very important because oh wise is going to just do the fastest path to get you your your answer so um yeah that's super .

cool yeah one hundred percent um something else that i've noticed where virtually well is like once IT generate something to you. Can I have like prom sometimes for like refining the prompt gives me so like he writes an article like awesome actually article.

Now rewrite this in a way that is like more likely to be more interesting or engaging um or more likely to be X Y N Z or rewrite this in a way but like double check all the facts in IT and let me know if any of the facts you think might be dubious now I know you should still like fact check yourself coming tragedy. But like your job is like easier if you get IT to do and it's like, oh so sorry I said this but it's not actually this is actually bad like IT will catch self a lot of time. So I think it's kind of interesting.

You can also ask IT if I missed anything on like on its previous like uh, output is like are you forgetting anything? Is there anything I should know? It's it's fun of you like some things that you would say to a human, right that they're like, oh my god, yeah let me go add to that .

like something I wish they would add, which as far as I know, is not a thing where you could do with almost colour code here what the answer gives you and tell you like the percentage accurate he thinks that is you know like or if he does have a cat that are pulled from know to give you that there I don't logic you know sometimes I feel like IT. It'll take almost more time to go through. In fact, check your whole article then to you know actually you write IT yourself it's yeah some that kind of tRicky but I ish I had like a color coded thing where you could can see if the yellow there. It's kind of you know trying to fill in the gaps but not the hottest processor um yeah I don't know yeah .

one hundred percent two other things that they mentioned, number one is to use external tools. I think that goes without saying you yep, use other things like charge. You can do everything so uh integrated other tools and they have their whole APP stores so you can integrate all those.

And the last thing I want to talk about is um they say to test your changes systematically. So essentially if you can get IT to like uh come up with an amazing response, but you're trying to like fine tune that response or get IT to be Better, like use that sponge, use that response and say, like, hey, this is the gold standard response. Now I want you to also do like X, Y and z, but like make sure IT is good as this response. And by saying that I can draw back and like kind of like auto correct itself to give you a good output.

Yeah I can like the scientific method where you you don't have too many different variables. You want to just change one thing at a time and and then compare IT, make sure it's it's still you know has the original entirely so yeah you're just changing one small thing in a time or make this section Better or you know kind of breaking IT down into smaller go going back to the smaller chunks thing yeah yes.

Like I mentioned something problem and I have like ten bullet points of like tasks or things that needs to do and like IT won't quite do a couple of them and like half the time attempted to be like i'm going to rewrite this holding from scratch and then put is like terrible my god we're grown up one almost work and now just have to figure how to like just do one two week at a time and and then you get there and here's one pro tip h for everyone if you do this strategy um sometimes I say I like I tweet the prompt and I say give me the output five times.

I give me five outputs for that and or just say like do to can do to get and do to again five times and then you can check because sometimes I will like mess up on something every fifth time or every tenth time like IT mostly gets a prompt. But if you have to run at a ton IT, it'll just mess up ever once in a while. So for cases like that, we're going to be using a prompt over over again.

I tried to get like a lot of the output, see there's any issues and then if there is, I fix in the main prompt. But like sometimes the problem doesn't show up every time. So multiple outputs are useful in that out.

Yeah that isn't resting. Well, thanks for thing guys. Be sure to rate our podcast reviews and really helps us a lot. If you get any value from this, we appreciate will see guys next ten.

Thank you for listening to today's podcast episode, breaking down how A I is impacting your industry. Today's episode is sponsored by A I box, a no code, A I P builder and marketplace, which just launched a crowd funding campaign. If you are interested in investing in a new AI start up, you go to republic dot com flash A I dash box.

I'll leave a link in the shown notes as well to learn more. The minimum investment is one hundred and fifty dollars, and the maximum investment is one hundred thousand. Until next time, have a fantastic day.