WEBVTT 00:00.210 --> 00:01.290 Instructor: So in this video, 00:01.290 --> 00:03.630 I want to discuss some very good tips 00:03.630 --> 00:06.630 that can help you get started with prompt engineering 00:06.630 --> 00:08.970 and to manufacture some prompts 00:08.970 --> 00:12.870 that will produce you some awesome responses. 00:12.870 --> 00:15.330 Those tips are low hanging fruit. 00:15.330 --> 00:18.000 That means that it's super easy to incorporate them, 00:18.000 --> 00:20.130 and to implement them in your prompts. 00:20.130 --> 00:21.210 So once you do that, 00:21.210 --> 00:23.460 your prompt's going to be much, much better, 00:23.460 --> 00:26.010 and you'll see an improvement in your answers 00:26.010 --> 00:28.200 that you get from the large language model. 00:28.200 --> 00:31.320 So the first thing we need to consider is the context. 00:31.320 --> 00:33.930 Now the context is super important 00:33.930 --> 00:38.040 because it gives our prompt some contextual relevance 00:38.040 --> 00:41.790 and for generating coherent and accurate responses, 00:41.790 --> 00:43.557 we need to have the context. 00:43.557 --> 00:46.680 Now, if we don't provide the context to our prompt, 00:46.680 --> 00:50.664 then we are simply leaving this task of creating the context 00:50.664 --> 00:52.440 to the large language model, 00:52.440 --> 00:56.940 and the LLM will not provide the context that we may want. 00:56.940 --> 00:59.460 And we're simply leaving the LLM to guess 00:59.460 --> 01:00.780 what is the context, 01:00.780 --> 01:03.800 and that may generate a response that will be off topic 01:03.800 --> 01:05.790 and it won't be relevant 01:05.790 --> 01:07.950 or inconsistent with the goal that we want 01:07.950 --> 01:10.323 to achieve with our prompt. 01:10.323 --> 01:15.181 Okay, so this is example that we pass to the LLM sub context 01:15.181 --> 01:19.050 and we can see what kind of context the LLM deduces 01:19.050 --> 01:21.090 from the context that we gave it. 01:21.090 --> 01:22.890 So the prompt is the following. 01:22.890 --> 01:25.880 We want to tell the LLM to break down our prompt 01:25.880 --> 01:28.410 with contextual relevance. 01:28.410 --> 01:31.350 So the prompt is we want to generate a list 01:31.350 --> 01:33.390 of technical interview questions 01:33.390 --> 01:35.384 for a senior DevOps engineer position 01:35.384 --> 01:38.424 in a tech startup in a fast-paced culture 01:38.424 --> 01:40.757 working in the cloud, okay? 01:40.757 --> 01:44.220 So the task is to break down the prompt 01:44.220 --> 01:45.690 into contextual relevance. 01:45.690 --> 01:46.800 This is the task. 01:46.800 --> 01:48.300 And what comes after it, 01:48.300 --> 01:51.033 the entire sentence over here is the context. 01:53.190 --> 01:55.500 Now, what we can learn from the response 01:55.500 --> 01:59.144 that we got is that the LLM was smart enough 01:59.144 --> 02:03.420 to first recognize what is the context that we provided, 02:03.420 --> 02:06.420 and it was even able to label it as context 02:06.420 --> 02:09.810 and it was even able to label the task, 02:09.810 --> 02:13.050 which was to generate the technical interview question, 02:13.050 --> 02:14.580 which is super, super cool, 02:14.580 --> 02:17.670 and it is helping us to get a better result. 02:17.670 --> 02:20.610 So now I'm simply going to paste in the prompt 02:20.610 --> 02:23.940 that we want only the interview questions. 02:23.940 --> 02:25.293 Now let's review them. 02:27.720 --> 02:29.580 Now, we're not going to review all 02:29.580 --> 02:30.840 of the questions over here, 02:30.840 --> 02:33.210 but let's take a look at the first one. 02:33.210 --> 02:36.000 So how would you handle a sudden spike in traffic 02:36.000 --> 02:37.590 on the company's website? 02:37.590 --> 02:38.670 What steps would you take 02:38.670 --> 02:40.470 to ensure the website remains available 02:40.470 --> 02:42.563 and responsive to the users? 02:42.563 --> 02:46.424 So I personally come from a software engineering, DevOps, 02:46.424 --> 02:49.923 and cloud native application development background. 02:49.923 --> 02:51.923 And I can tell you right now 02:51.923 --> 02:55.123 that this interview question is very deep 02:55.123 --> 02:59.661 and you can talk only on this question for a half an hour, 02:59.661 --> 03:02.970 an hour, even two hours about the strategies, 03:02.970 --> 03:05.501 about how to implement, about use cases. 03:05.501 --> 03:07.704 So it's a very deep question, 03:07.704 --> 03:10.470 and even when I was interviewing candidates 03:10.470 --> 03:12.210 to a software engineering position, 03:12.210 --> 03:13.725 which also involved DevOps, 03:13.725 --> 03:16.350 this was a question that I asked it 03:16.350 --> 03:18.213 in a similar way or another. 03:19.540 --> 03:20.910 So just to summarize, 03:20.910 --> 03:24.150 the more contextual relevance that we add to our prompt, 03:24.150 --> 03:26.160 the better the results that we will get. 03:26.160 --> 03:28.500 So this is a very low hanging fruit 03:28.500 --> 03:30.000 and it's super easy to put 03:30.000 --> 03:32.613 and this will make the results much, much better. 03:33.720 --> 03:37.320 Okay, so the second tip is that we need 03:37.320 --> 03:41.400 to set a clear task that will not be ambiguous. 03:41.400 --> 03:45.180 So the task definition need to set a specific goal 03:45.180 --> 03:48.410 or objective for the LLM to achieve. 03:48.410 --> 03:52.050 So if we don't provide this goal that is clear, 03:52.050 --> 03:55.540 that is precise, that it is exactly what we want, 03:55.540 --> 03:59.670 then we simply leave it for the LLM to choose for us. 03:59.670 --> 04:02.760 So I like to think of it like when my wife sends me 04:02.760 --> 04:04.950 to the supermarket to buy apples. 04:04.950 --> 04:07.650 So okay, she didn't say what kind of apples, 04:07.650 --> 04:09.120 green apples, red apples. 04:09.120 --> 04:13.380 Does she want small ones, medium ones, large ones? I dunno. 04:13.380 --> 04:14.707 So I bring them apples 04:14.707 --> 04:17.760 and of course the result is not what she wanted 04:17.760 --> 04:19.050 and she's disappointed 04:19.050 --> 04:21.750 because I didn't bring her the apples she wanted. 04:21.750 --> 04:24.510 Now, in my opinion, being a prompt engineer 04:24.510 --> 04:27.240 is having great communication skills. 04:27.240 --> 04:30.780 So it is important to have communication skills with people, 04:30.780 --> 04:33.780 of course, because you want to know how to convey a message 04:33.780 --> 04:36.330 and you want people to understand what you want. 04:36.330 --> 04:38.250 So the same thing is happening 04:38.250 --> 04:40.200 with the large language model. 04:40.200 --> 04:42.510 If we want to be great in prompt engineering, 04:42.510 --> 04:44.220 then we need to sharpen 04:44.220 --> 04:46.290 and to improve our communication skills, 04:46.290 --> 04:48.060 but this time with the LLM. 04:48.060 --> 04:50.520 And the LLM likes to get the prompts 04:50.520 --> 04:55.520 that are having concise, clear and non-ambiguous tasks. 04:57.450 --> 05:01.171 Okay, so I think now is a good time to review an example 05:01.171 --> 05:02.790 of an ambiguous task 05:02.790 --> 05:06.240 because it will help us avoid those kind of tasks. 05:06.240 --> 05:08.430 So let's take a look at the prompt, 05:08.430 --> 05:12.406 improve the user experience of this e-commerce website. 05:12.406 --> 05:15.150 Now this prompt is ambiguous 05:15.150 --> 05:18.088 because it doesn't specify what aspect 05:18.088 --> 05:21.208 of the user experience needs to be improved 05:21.208 --> 05:23.520 or how should it be improved. 05:23.520 --> 05:26.580 This prompt doesn't provide any specific guidelines 05:26.580 --> 05:31.500 or metrics for evaluating the success criteria of the task. 05:31.500 --> 05:34.152 The task would be interpreted in many, many ways, 05:34.152 --> 05:37.560 and this could lead to confusion, lack of clarity, 05:37.560 --> 05:40.083 and lower performance overall. 05:40.950 --> 05:43.770 Okay, now let's change this prompt 05:43.770 --> 05:45.570 and let's make it clearer, 05:45.570 --> 05:48.450 make it less ambiguous and overall better. 05:48.450 --> 05:50.520 So now the task is identify 05:50.520 --> 05:53.790 and address specific pain points in the user experience 05:53.790 --> 05:58.080 of the e-commerce website to increase customer satisfaction 05:58.080 --> 05:59.940 and sales conversion rates. 05:59.940 --> 06:02.488 Now this version of the prompt is clearer 06:02.488 --> 06:05.912 because it specifies that the task involves with identifying 06:05.912 --> 06:09.592 and addressing specific pain points in the user experience 06:09.592 --> 06:12.420 rather than simply improving the user experience 06:12.420 --> 06:13.770 in a general sense. 06:13.770 --> 06:17.940 So this prompt also includes specific goals for the task, 06:17.940 --> 06:20.070 like increasing customer satisfaction 06:20.070 --> 06:21.600 and sales conversion rates. 06:21.600 --> 06:25.050 And this provides more concrete metrics 06:25.050 --> 06:27.423 for evaluating the success of the task. 06:29.220 --> 06:32.940 Now, this task was not ambiguous, it was clear 06:32.940 --> 06:36.543 and it was accurate because we were very specific. 06:36.543 --> 06:39.667 And as far as specificity, then it will refer to the level 06:39.667 --> 06:44.667 of detail that we include in our prompt to be more precise. 06:44.910 --> 06:47.490 The more specific the prompt will be, 06:47.490 --> 06:51.480 the more likely that the LLM will generate a targeted 06:51.480 --> 06:53.910 and accurate response. 06:53.910 --> 06:58.860 Okay, the final idea I want to introduce is iterations. 06:58.860 --> 07:02.280 Now iterations are crucial in prompt engineering 07:02.280 --> 07:05.580 and in the design of a prompt process. 07:05.580 --> 07:08.550 I'm a very big fan of this process. 07:08.550 --> 07:12.026 It is being practiced a lot in software. 07:12.026 --> 07:15.542 It's also introduced in the book "The Lean Startup." 07:15.542 --> 07:19.706 And basically to iterate is to repeat and repeat and repeat. 07:19.706 --> 07:22.186 And each time we repeat the process, 07:22.186 --> 07:24.266 we simply change it a bit 07:24.266 --> 07:28.770 and we simply make our execution better. 07:28.770 --> 07:30.900 So in prompt engineering, 07:30.900 --> 07:34.260 the iteration will involve refining the prompt 07:34.260 --> 07:36.990 that will improve it overall, okay? 07:36.990 --> 07:40.940 So it would also involve testing and evaluating the output. 07:40.940 --> 07:43.770 Each iteration should rely on the output 07:43.770 --> 07:45.540 of the previous iteration, 07:45.540 --> 07:47.979 and through this output, it should refine 07:47.979 --> 07:50.640 and make the prompt better. 07:50.640 --> 07:54.750 So this process will result in an optimized prompt 07:54.750 --> 07:58.620 according to our likings, after we iterate, we refine it 07:58.620 --> 08:00.720 and we have the perfect prompt 08:00.720 --> 08:03.753 that will give us the perfect result that we want. 08:06.870 --> 08:10.050 So we are going to finish this video soon, 08:10.050 --> 08:13.320 and I want that the takeaway from this video to be 08:13.320 --> 08:17.940 that the time we take to write a prompt is super important. 08:17.940 --> 08:20.850 And before we write a prompt, I know it's tempting 08:20.850 --> 08:23.550 to simply write something fast and click enter, 08:23.550 --> 08:27.600 but if we invest the time of prompt engineering our prompt 08:27.600 --> 08:31.167 and making it more precise, refining the tasks, 08:31.167 --> 08:35.580 adding the context, being specific, not being ambiguous, 08:35.580 --> 08:40.106 and constantly refining our prompts over iterations, 08:40.106 --> 08:44.619 then the results that we will get will be much better, 08:44.619 --> 08:47.523 and this will overall save us time. 08:48.450 --> 08:50.880 So that's what's really prompt engineering is about, 08:50.880 --> 08:52.463 in my opinion, to manufacture 08:52.463 --> 08:55.590 and to create this perfect prompt 08:55.590 --> 08:57.813 that will give you exactly what you want.