Prompt Engineering Full Course

[00:00] In this video, you'll learn everything

[00:04] I'll explain what it is.

[00:05] Why it's important. How to do it faster.

[00:08] The top techniques and methods

[00:13] With that said, let's dive in.

[00:14] So what I've done for this

[00:17] that contains all of the examples

[00:21] Now, if you want this file for yourself

[00:24] just by clicking the link in the description

[00:28] Anyways, let's start and I want to just show you

[00:32] a great prompt just to kind of set the tone,

[00:36] And keep in mind, all of the different topics

[00:40] in the video player or the chapters in the comments

[00:44] Okay, so what is a bad prompt?

[00:45] Well, an example of that would be something

[00:49] It's going to give you something like generic

[00:53] probably no CTA.

[00:54] The model is effectively going to

[00:56] guess what you want here, whereas a better prompt

[01:00] You are a senior B2B copywriter.

[01:02] Write a two sentence LinkedIn

[01:05] As an alternative.

[01:06] The audience is ops managers at mid-sized companies.

[01:09] The tone is confident

[01:13] Now, if you do this, you're

[01:16] scoped, actionable, and easy to drop into an ad.

[01:19] The point of me showing you

[01:22] these prompts,

[01:26] So what's giving us the difference in result

[01:32] And most people, while this kind of seems obvious,

[01:37] than they are to the good prompt, and just because a

[01:42] Which I'm going to go over in this video.

[01:44] So this leads me to what is prompt engineering

[01:48] Well, you should really think of prompt engineering

[01:52] So rather than using Python or Java or JavaScript,

[01:58] or whatever

[02:01] some kind of result complete some action,

[02:04] Effectively what you're doing is you're giving

[02:06] instructions in plain language

[02:10] Now, it's important to note that the models

[02:13] these instructions to,

[02:16] So you have to define the task, the role, the format

[02:21] And you have to be very specific

[02:24] And that's why the same model can seem brilliant

[02:28] depending on the clarity, context

[02:31] Now, when most people prompt AI models,

[02:36] And while that's still a very valid use case

[02:40] that we've now shifted kind of into an agent era

[02:43] where a lot of these models

[02:47] searching the web, and actually taking actions

[02:52] They can write a Google doc,

[02:56] And because of that prompt,

[02:59] because we're no longer just directing the model

[03:04] we're actually instructing it to take some kind of

[03:09] So just keep that in mind.

[03:11] While a lot of what I'm showing you here

[03:13] This also works for agents, right?

[03:16] Or more complex models that are capable

[03:21] You can see a vague prompt would be something like

[03:25] is something like you were a tutor,

[03:28] of this 500 word history essay

[03:32] Keep my voice, adjust the edits inline, etc.. Okay.

[03:35] Now, in terms of why this is important,

[03:39] So for people like developers,

[03:44] the better

[03:47] that is for everyone, and the faster

[03:51] So the takeaway write

[03:55] and this is overall

[03:58] But by knowing some of the foundations that I'll

[04:02] you and allow you to think of these a little bit

[04:06] Now, what you've probably gathered at this point

[04:10] the better the result or action

[04:14] Now, that's generally true.

[04:15] I mean, longer prompts are not always better, but

[04:19] However, that's

[04:22] Most people are lazy, they type very slow,

[04:26] 20, 30 minutes here typing a prompt just to have to

[04:31] So what I'm going to suggest to

[04:32] you guys is going forward,

[04:36] So use your voice, use natural language

[04:41] faster where you can include all of the details

[04:46] Now this applies on phone and applies on computer,

[04:49] to do this within any application

[04:53] Here is called a whisper flow.

[04:55] Now I'm just going to quickly show you how it works

[04:58] I have a link in the description.

[04:59] I would highly recommend it

[05:03] that works a lot

[05:06] So for example, let me just do something.

[05:08] Hey, this is just a quick test

[05:11] and see what kind of results I'm getting.

[05:12] Notice I'm talking,

[05:17] I'm making a few repetitions, pauses, etc.

[05:20] and let's see the result that we get.

[05:22] Okay, so you can see the kind of little bubble

[05:25] And pretty much instantly we get the result.

[05:27] And you'll see that if we go through the result here,

[05:31] the capitalization, and corrects

[05:36] I believe I said extremely twice.

[05:38] It recognized that wasn't useful and it just removed.

[05:41] So you get a much faster, better result

[05:45] I want to have three bullet points.

[05:47] Bullet point one is Apple's bullet point two is

[05:51] banana's, bullet point three is pears.

[05:54] And you'll see that

[05:57] Okay.

[05:58] So when you're doing the prompts,

[06:00] And this tool will handle it

[06:04] I do have a long term partnership

[06:08] before we did the partnership

[06:11] You can see here I've spoken 35,000 words.

[06:14] My WPM is 162,

[06:19] It has a lot of other features

[06:22] You can change the style based on the apps

[06:25] all of this kind of stuff, and a bunch of settings

[06:29] on my computer, and that allows me to actually

[06:35] So I use this in every kind of application.

[06:37] Sometimes I use it to write emails,

[06:40] Right?

[06:40] Sometimes I used in cursor

[06:44] and I'll show you more of it in the video.

[06:45] But the point is, speak your prompts,

[06:49] And this tool also is available on phone.

[06:51] They actually recently released it on Android,

[06:56] like on your iPhone or Android.

[06:57] So even if you're on mobile,

[07:01] comprehensive prompt

[07:04] So this now leads us to the next section,

[07:08] Now, an LLM is a large language model.

[07:12] And effectively what these are capable of doing

[07:16] So you give some text in, they give some text app.

[07:19] Now you can configure them in very fancy ways

[07:23] or call other LMS or create something

[07:27] But effectively, at the very root of an Lem,

[07:32] It's important to understand this

[07:36] you need to imagine that

[07:40] the next text that you would expect

[07:43] So while you might see reasoning models,

[07:47] is thinking or developing something.

[07:49] Really, it's just been trained

[07:54] it's able to predict

[07:57] what it is it thinks you want in terms of the output.

[08:00] So the prediction is usually going to be in text

[08:05] You give some tokens in it spits some tokens out.

[08:08] That's effectively what an Lem is.

[08:10] We don't need to worry about the architecture

[08:12] or anything complicated there just remember

[08:15] you get some output, and it's just predicting

[08:19] Now with that in mind and Lem by default.

[08:22] So something like GPT 5.2 right.

[08:24] It doesn't have any memory.

[08:26] Now what would give it

[08:30] injecting previous parts of your conversation

[08:33] into the prompt or into the model automatically.

[08:36] So what you need to understand

[08:39] Lem's right.

[08:40] You use ChatGPT, you use Claude, or you use curser.

[08:43] Behind the scenes, the people who developed

[08:48] they're injecting a lot of data into the prompt

[08:54] So while you may prompt and say, hey,

[08:59] behind the scenes,

[09:02] where it includes a lot more details

[09:06] So it's just important for you to understand that,

[09:10] It doesn't have, like all of these fancy features

[09:15] The reason why that exists is because OpenAI,

[09:19] created this nice interface on top of the Lem,

[09:25] that injects this other information into the prompt

[09:28] that then goes to the model,

[09:32] So 99% of the time, your prompt is not the only thing

[09:37] It's going to be seeing,

[09:39] Maybe it has access to some tools, right?

[09:42] Whatever.

[09:42] There's all kinds of things that are usually

[09:46] And it's really important

[09:50] And you don't think like GPT 5.2 or something

[09:55] The only reason that it knows something is because

[10:00] previously, and just injecting that into the prompt

[10:05] Now, that additional information that I just talked

[10:10] It's the other stuff that the Lem can see

[10:14] That's not just the prompt that you gave it.

[10:16] So just understand that the context, specificity

[10:20] that you provide is what shapes

[10:23] And a lot of these tools, just like,

[10:25] The tool that I'm using right now,

[10:27] if I create an agent by default, it's

[10:31] And if I tag something like this markdown file right.

[10:34] And I say, you know, hello world, the prompt is

[10:40] Now, with that in mind, let's quickly

[10:44] So the common mistake that most people make

[10:48] command the model to do something,

[10:52] Now commanding

[10:55] Do this, you know, write this down.

[10:57] Take this thing. Look at this thing. Whatever. Right.

[11:00] So if that's the case,

[11:03] that it's going to take

[11:07] Whereas steering the model,

[11:11] So you are an executive assistant.

[11:12] Summarize the meeting transcript and for bullet

[11:17] No filler.

[11:17] Notice that we're not only telling you what to do,

[11:21] and we're giving it a much more clear objective

[11:26] it can give us a much more accurate, better result

[11:31] So when you steer the model,

[11:32] you're typically indicating the length,

[11:36] Whereas when you're commanding the model,

[11:40] how it's going to do that. The result

[11:42] So I'm not going to focus on that too much more,

[11:46] Now, what I want to do is get into the core

[11:48] techniques of prompt engineering

[11:52] So one of the most basic and important techniques

[11:56] and to set the scene or specify

[12:00] The model or agent to take.

[12:02] Now these are the four things that,

[12:06] give you a much better result

[12:10] If you include a role in audience, a tone,

[12:13] format or output, you're almost always going

[12:17] So here it says,

[12:20] Okay, fine, you can reply to it.

[12:21] But if you said something like this,

[12:25] the audience, the tone and the format.

[12:27] So you're going to get a better response.

[12:28] So what I want to do now

[12:31] by actually writing or voicing

[12:34] So let's just do a quick example of the bad prompt,

[12:37] Reply to this customer complaint.

[12:39] And then I'm just going to paste in a random one

[12:42] And let's see the result

[12:45] So you can see here it gave us something decent,

[12:48] Even though you know, we didn't specify we want it

[12:52] It didn't say it was from the support team. Whatever.

[12:55] So what I'm going to do now is just make a new chat

[12:59] with the same customer complaint

[13:03] Reply to this complaint from a paying user

[13:07] Acknowledge the frustration. Apologize. Briefly.

[13:10] Confirm

[13:13] For example, we'll email it in 24 hours,

[13:16] 150 words and sign off as the support team.

[13:19] Cool.

[13:19] So that is our prompt.

[13:21] Okay.

[13:22] Notice again

[13:25] So another important part when you are writing

[13:29] well so you have like bullet points

[13:32] sections,

[13:35] It's going to be a lot easier for the model to

[13:40] For example, if we even just do a quick delimiter,

[13:44] and saying, you know, here is the complaint.

[13:47] And then putting a colon, it's separate.

[13:49] So like this is the thing it should be listening to.

[13:52] And then this is the additional context.

[13:54] And also notice that I created a new session here

[13:59] not automatically going to include the previous text

[14:04] So I'm just going to go ahead and hit enter here.

[14:06] What I mean is that if I had just run this underneath

[14:10] you know, we can still get a decent result,

[14:15] or mixed up with the other conversation history,

[14:19] in the prompt as well,

[14:23] So if we look here, you can see now we get something

[14:26] You know, it says best regards support team.

[14:28] And it's followed

[14:31] Then the next technique you can use

[14:35] Now effectively what this means is

[14:39] of input and output pairs that you're looking for, so

[14:42] that when you give the prompt,

[14:45] So when you do this, the model can infer and figure

[14:50] and then replicate the pattern

[14:54] So this is especially helpful

[14:58] You have various edge cases,

[15:01] task using an LLF, because

[15:05] For example, like is this email positive or negative?

[15:08] Is it red or black, whatever, right.

[15:11] But only if they know and have examples

[15:15] And in fact, when you talk about fine tuning LMS,

[15:20] is you're passing them a bunch of different examples

[15:25] so they can infer that pattern on top of their base

[15:30] So let's have a look at a few examples here.

[15:32] I'm not going to throw them on screen.

[15:33] Let's just go into ChatGPT and we can do

[15:38] So let's do something like turn

[15:43] The feedback is the app crashed

[15:49] Okay, and let's see what we get here.

[15:51] Perfect.

[15:53] Let's see what it gives us.

[15:54] Okay, so you can see we didn't really get any clear

[15:56] It didn't actually give me one title.

[15:58] It gave me five because again

[16:01] But let's change and do something like this.

[16:03] Turn user feedback into a short ticket title.

[16:05] We want this to be under 60 characters,

[16:11] Okay. And then I'm just going to paste this.

[16:13] And then here I'm just going to say here are

[16:17] Just give me back one line of text that starts

[16:23] And then a brief description of the issue. Okay.

[16:26] And then we should just say something like this.

[16:28] And then I can even do,

[16:33] Okay.

[16:33] Cool.

[16:34] So let's now run this

[16:37] So let's actually separate the examples a little bit.

[16:40] So it's it's a bit easier to see.

[16:41] And you can see that we have you know feedback long.

[16:44] It is broken with Google on Safari

[16:47] And then Google login fails on Safari.

[16:49] And then let's go here.

[16:50] Feedback is export to CSV only. Exports

[16:54] The title export right CSV feedback.

[16:56] The app crashed

[17:00] This is actually the example

[17:03] Sorry.

[17:04] And I'm going to say like this.

[17:07] This is the feedback to generate the title for.

[17:13] Cool.

[17:13] Okay.

[17:14] And then it should give us the title

[17:16] Now again I'm being super specific,

[17:19] a few examples,

[17:23] And there you go.

[17:23] You can see it says upload and then iPhone app

[17:28] So in this case I was super verbose.

[17:30] So we went from this kind of crap here

[17:34] now I could directly plug into like some API

[17:38] And I know that the format I'm going to get from

[17:41] And then I could just do something else now

[17:44] So let's go.

[17:45] Like feedback I don't know here.

[17:47] Let's go feedback and let's change this

[17:50] to say something

[17:53] The keyboard does not disappear when clicking

[17:59] Okay.

[17:59] And let's see if we give it

[18:02] And there you go.

[18:04] And we can now keep doing this in this chain because

[18:09] Now the next technique to go over

[18:13] Now this is asking the model to reason step by step

[18:18] Now this will reduce errors on logic, math, you know,

[18:23] And a lot of models do this by default now,

[18:28] or a reasoning model

[18:31] Right.

[18:31] Again, like in the prompt that's being created for

[18:36] So the reasoning models, like

[18:40] some fancy kind of tools and orchestration

[18:45] But if you're not using a fancy model like that,

[18:49] with chain of thought.

[18:50] So without it, you have something like, you know,

[18:55] Sarah buys three pens and two notebooks.

[18:56] She has a 10% off coupon. Coupon? Sorry.

[18:58] How much does she pay now?

[19:00] If you don't tell it to reason, it's just going to

[19:05] Now, in a lot of cases, it will get this wrong.

[19:07] Just like the famous example of asking them to count.

[19:10] You know how many teaser and strawberry or something,

[19:14] unless you change the prompt

[19:18] so that it actually thinks before

[19:21] Okay,

[19:23] because it is probably not going to work

[19:27] will give us a very similar result,

[19:32] models is already really good,

[19:36] the reasoning by default, or if you're not using them

[19:40] and you're just using it in like an API or something,

[19:44] these kind of keywords, right?

[19:45] So just anyways,

[19:48] A lot of times now

[19:49] the models that are super modern and,

[19:53] And actually auto detect what reasoning model

[19:58] But it is kind of worth knowing, especially,

[20:01] 12 months ago, this was a lot more important than it

[20:04] So the next method that is super

[20:08] Now this is asking the model to provide the output

[20:14] XML, markdown, whatever, and then giving an exact

[20:18] such that you're able to parse this and use it

[20:23] Now, in my case,

[20:26] And when I have the model generating something,

[20:31] Maybe I want to display it

[20:33] In order to do that,

[20:37] So what I will do is something like what you see

[20:39] right here

[20:43] This is what is referred to, which is really just

[20:48] So it knows how to give me the output.

[20:49] So obviously compared to the first example

[20:52] something like this, the main difference

[20:55] and we're telling it that we want it to respond

[20:58] No other text in a particular shape.

[21:00] So we're kind of almost doing this

[21:03] or few short prompting,

[21:07] So anyways, let me go to ChatGPT

[21:10] So what's something that we can do.

[21:12] so we can say maybe compare

[21:16] For a small team of 5 to 10 people,

[21:19] the main features, the limitations and the pricing.

[21:23] And what I'd like you to do is output this with valid

[21:27] Don't give me any other text

[21:31] Okay, so that is what we're going to do here.

[21:33] Just put that in with whisper.

[21:34] And then I'm just going to paste in the format.

[21:36] So we'll just put it like that

[21:41] So let's run this now and let's see what we get okay.

[21:44] So just give me the output right now.

[21:46] But you can see that it's following this format.

[21:48] It's in a valid Json object.

[21:50] And now I would be able to load this into my code.

[21:52] I could store it in a database right.

[21:54] I can return it from an API.

[21:55] It doesn't matter. But it is very useful.

[21:57] And it's going to be a lot better

[21:59] this random kind of markdown text

[22:04] So moving on, let's talk about constraints

[22:07] So sometimes the best prompts are not talking

[22:10] too much about what to do,

[22:14] Now the more constraints which is saying you know, do

[22:19] whatever the better the results you're going to get,

[22:23] So some examples of constraints is for example

[22:27] Now you could say summarize in exactly

[22:29] That's fine. That's a constraint.

[22:31] You also could say keep the reply under 300 words

[22:35] Right.

[22:35] You could have a tone like no slang or humor.

[22:37] Do not apologize. Right.

[22:39] Do not use bullet points. Use one short paragraph.

[22:42] Do not suggest paid tools. Do not include code.

[22:44] Describe the approach only.

[22:46] So when you do that, it actually ends up listening

[22:51] And this has been shown a lot,

[22:52] especially with a lot of the researchers

[22:56] are including a large list of things not to do

[23:01] So you don't want to write something

[23:04] Instead, you would say,

[23:07] Start directly

[23:10] Do not start with welcome or generic greetings.

[23:13] And then in that case, you're going to get something

[23:16] So anyways, let's have a quick

[23:18] If we go inside of something like ChatGPT again,

[23:22] So let's just do something

[23:26] Tip okay, so this would be without any constraints.

[23:30] And let's see the kind of response that we get.

[23:31] Okay, so ChatGPT actually gave me something

[23:35] for an out of office email

[23:38] the dates, the time, all of that kind of stuff.

[23:40] So let's do something better would be, say,

[23:44] start with the exact dates I'm away,

[23:49] Do not use the phrase limited access.

[23:52] Reply when I can or thank you for your patience.

[23:55] Be short to the point,

[23:58] when I'm coming back,

[24:02] Okay, so in this case, again,

[24:05] So hopefully we're going to get

[24:08] And let's see perfect.

[24:09] And you can see that we get a much better email

[24:13] you know, message me

[24:16] Which funny enough, we got in this one,

[24:19] Now the next method to go over is pretty

[24:23] but it's just to do iterative refinement

[24:26] So very rarely is the first prompt

[24:30] So rather than restarting simply

[24:32] improve the prompt over and over again

[24:36] So as it says here, you know, treat prompting

[24:41] The first reply maybe isn't right,

[24:44] I'd like it more formal. Add another example.

[24:46] Focus only on X, whatever.

[24:48] Of course, the better first prompt

[24:51] But sometimes you just don't know

[24:53] In that case, don't restart.

[24:55] Just keep going based off the result that you have.

[24:58] If you're in like a decent enough place

[25:01] you know, you dropped a two sentence blurb.

[25:03] Whatever it returns four sentences a bit salesy.

[25:06] Hey, cut it to two sentences.

[25:08] Whatever model tightens

[25:11] You get the idea.

[25:12] I'm not going to show this to you in ChatGPT

[25:15] But the point is, be iterative with your prompt.

[25:18] And then the last method that I'm quickly going

[25:22] is the interview style prompting,

[25:26] So I'm just going to read through some of this

[25:29] But instead of guessing what context to provide,

[25:33] and ask it to interview you

[25:37] So a lot of times you know, you're

[25:40] because you don't actually know what information

[25:45] You can go off the top of your head,

[25:48] if the model prompts you on what it is that it needs,

[25:53] where it's prompting you and then your prompting it,

[25:58] So you state the goal,

[26:01] and you can answer them one by one, and then it can

[26:06] Now it works because we often leave out

[26:09] Audience constraints, format examples

[26:13] So if we let them surface that information first,

[26:17] it gives us kind of the prompt to

[26:21] And a lot of times we're lazy or we have assumptions.

[26:23] So again, we just don't include something

[26:26] but it may not actually know that

[26:30] Okay.

[26:30] So I mean, if you have this talk

[26:32] but what I'm going to do is just go over into ChatGPT

[26:36] this kind of interview style

[26:40] Okay, so let's just read something out here.

[26:42] I need a short LinkedIn post about 250 words

[26:45] about lessons

[26:48] Before you write it, interview

[26:52] So, for example,

[26:55] do's and don'ts,

[26:59] And then when you have enough information, say,

[27:03] Then I want you to write the post and again

[27:09] Okay.

[27:09] So we have the prompt here.

[27:11] You can see that it's generated.

[27:13] And now let's see if we can do this.

[27:15] So great says first question who is the primary

[27:19] And let's say it is founders and CEOs

[27:24] okay.

[27:25] Let's see what question we get next. Got it.

[27:27] Second question.

[27:28] Tell me about your company context.

[27:29] So the advice doesn't feel too generic.

[27:31] I run a developer mentorship program

[27:36] you know, land software engineering jobs

[27:40] And I employ a team of about ten people.

[27:43] They're mostly remote where they're doing like

[27:48] For the company.

[27:50] I'm just making up random stuff right now.

[27:52] But anyways, you get the idea.

[27:53] Okay, let's see what else. Get context.

[27:55] Next question why did you switch to a four day

[27:59] We switch because we wanted to get more flexibility

[28:02] and to make them more focused on the four days

[28:08] to do when they, you know, didn't need to be doing

[28:12] Okay, let's see here.

[28:14] It's actually picked up just spinning random stuff.

[28:16] So let's remove that, and let's see what we get.

[28:19] Next question.

[28:20] What actually happened after you switch

[28:24] The productivity went up.

[28:25] The revenue stayed about the same,

[28:29] And there was no unexpected problems.

[28:32] Okay. Let's see

[28:34] So after this I'm just going to tell it to stop

[28:37] Use the current information and give me the post.

[28:39] Don't ask any more questions.

[28:41] So you can see when I do this,

[28:45] that is accurate based on what I told it,

[28:49] AI crap that you guys probably see all the time over

[28:52] It actually matches,

[28:56] So what happened?

[28:56] Productivity went up, revenue stayed the same, team

[28:59] You get the idea.

[29:00] So this interview style is really effective

[29:05] by just chatting

[29:08] that you probably never would include by default,

[29:12] Okay, so now I'm going to go through

[29:16] So first one to talk about is system

[29:19] Now this is more relevant for those of you

[29:23] However in some different applications

[29:26] and a user prompt.

[29:27] Now, a system

[29:30] various rules and style, and it's often

[29:35] So this is for always on behavior

[29:40] going to read this system

[29:45] is designed for the system

[29:49] So for example,

[29:52] you know, I give it a system prompt

[29:55] I don't need to say that every time.

[29:57] It just always reads that.

[29:59] So in cursor, the app that I'm in right now

[30:03] they have their own system prompts

[30:05] that they provide to the models

[30:09] And then when you type the prompt,

[30:13] that's already inside of there, that's building

[30:18] So if you have a look here,

[30:20] be like you're a helpful

[30:23] Whereas the user is,

[30:27] Now if I go to ChatGPT,

[30:30] So actually if you go to ChatGPT settings,

[30:35] custom instructions, which is somewhat similar

[30:39] And you can see that

[30:42] Take a forward

[30:45] Now I'm going to change this to say,

[30:46] be direct into the points and always talk like Mario

[30:51] I don't know if that's going to work super

[30:53] but let's see now if we do a new conversation.

[30:55] Hey, how's your day going?

[30:57] Let's see if it's actually going to give me something

[31:00] You can see hey Tim, it's good, you see.

[31:02] So it's actually following like the system

[31:05] And giving me a much different response

[31:08] Now I need to turn that off.

[31:10] So let's just remove the custom instructions

[31:13] But you get the idea, okay.

[31:14] And even these other things like the style, warm,

[31:18] to the system prompt

[31:22] Okay.

[31:22] Now the next advanced feature to go over is prompt

[31:25] Now, generally speaking, it's better to break

[31:30] and to use the output of the previous step

[31:33] So rather than just writing a super complex prompt

[31:37] are the three different steps that you need to do.

[31:39] Just do them one by one

[31:43] before moving on to step two.

[31:45] So I'm not going to demonstrate this in ChatGPT.

[31:47] But the example would be like,

[31:48] you know, given topic, whatever output

[31:53] Okay. It then does that.

[31:54] Then you say okay based on that result,

[31:59] And then,

[32:01] Then you wait for that.

[32:02] I say turn this draft into,

[32:06] You get the idea.

[32:07] So go step by step and break it down.

[32:10] Rather than having the model do everything at once,

[32:15] more consistent result.

[32:16] Now, another, more advanced technique

[32:20] So let's say the model outputs something to you.

[32:22] You can then ask the model to critique

[32:26] Now sometimes you want to fool the model a bit

[32:30] You don't want to have it in the same one and say,

[32:34] that I wrote rather than saying, hey, an AI model

[32:38] going to give you a more accurate response

[32:42] So if you have drafts, code, summaries, whatever,

[32:46] here's a short summary

[32:48] Put the text, rate it

[32:51] in one sentence

[32:54] And then you can take that improvement

[32:58] Okay, so super cool.

[32:59] The evaluation is really good,

[33:02] And you're not just,

[33:05] chain after it gives you the output

[33:09] because it's not going to give you

[33:12] as if you ask it in a fresh chain

[33:16] Okay.

[33:17] And then just a few other things to look at here.

[33:18] So temperature and other parameters.

[33:20] Now if you're going to work with models

[33:23] setting, you're going to have the ability

[33:27] Now the temperature of the model

[33:31] Now the determinism means how repeatable effectively

[33:36] something that's deterministic is like okay,

[33:42] It always opens like it's a deterministic action.

[33:44] Like if I open this, it's always going to open right?

[33:46] Whereas if I generate a random number

[33:48] that's non-deterministic

[33:52] And I can get a random number now LMS by

[33:58] a sense of randomness and you don't know the output

[34:03] Now you can adjust how deterministic

[34:09] So the lower the temperature is, the more repeatable

[34:15] the higher the temperature, the more creative, varied

[34:19] So you would use a low temperature

[34:22] especially if you want something

[34:26] Right.

[34:26] Code facts, things where there's really just like one

[34:31] is the sentiment of this, you know, text positive.

[34:34] That would be a low, you know, temperature,

[34:38] Whereas for higher temperature, you use this

[34:44] maybe some advanced planning if it's not really sure,

[34:48] And if the outputs are too random,

[34:51] But if it's too repetitive,

[34:54] so you know, play with the temperature by default.

[34:56] It's usually on the lower side on the higher side.

[34:58] But just know what that parameter means

[35:00] because it does come up a lot

[35:03] So now let's talk about a few common mistakes

[35:07] Now one mistake being too vague right.

[35:09] What's going to go wrong?

[35:12] How can you fix that at a roll?

[35:14] Audience tone format potentially length okay,

[35:17] next, too many tasks in one prompt.

[35:20] So a lot of times you're putting,

[35:23] In that case, the model may, miss something.

[35:25] You may not actually complete the task.

[35:26] It might mix them up, whatever.

[35:28] In that case,

[35:31] Right.

[35:31] So chaining like I talked about before.

[35:33] Now none of context are examples.

[35:35] So you're going to get like wrong format or style.

[35:37] So add 123.

[35:38] Few shot examples include the relevant background.

[35:41] Or use that interview technique to ensure

[35:45] Another common mistake

[35:46] is that the model ignores the format that you want,

[35:51] So, like I showed you before,

[35:54] a specific structure and tell it to only output that.

[35:57] So Json with keys x, y, z or whatever. Right?

[36:00] Only give me Json.

[36:01] Do not give me any text, don't give me anything

[36:03] We're adding the constraints

[36:07] and then assuming memory like I said, the model

[36:12] or that aren't in the same thread conversation

[36:17] So make sure that you repeat or summarize key facts

[36:21] or if you're working in a new session,

[36:25] Again, you kind of have to infer

[36:29] doing in the background

[36:32] Okay, now there's a lot of other mistakes and things

[36:35] kind of the rest of the stuff in this document

[36:40] Last thing I will say is that it

[36:42] does just make a massive difference

[36:46] The more you prompt, the more you're going to see

[36:49] to, how you get the best response,

[36:53] If you use a tool like flow,

[36:57] you just to dictate it with your natural voice

[37:01] You guys have seen in this video

[37:04] and how much better

[37:07] So I know yes there partner of mine,

[37:11] I really cannot recommend it enough, and whether

[37:15] But point is, use some kind of dictation tool used on

[37:19] Your productivity will skyrocket, especially

[37:24] really long, detailed prompts and you're working

[37:28] So anyways, guys, that's going to wrap up this video.

[37:31] If you enjoyed, make sure leave a like subscribe