WEBVTT 00:00.280 --> 00:05.120 So that concludes our very first foray into OpenAI agents SDK. 00:05.600 --> 00:11.880 But before we wrap up week two, day one, I did want to have a second sidebar with you. 00:12.000 --> 00:19.240 And this time, it's on the entertaining topic of vibe coding, which is a term coined by the the legendary 00:19.280 --> 00:28.200 Andrej Karpathy, who described this in I think it was an ex post that went viral about the way that 00:28.240 --> 00:35.200 he was enjoying coding with Llms and getting so much done in a way that you would sort of let the LLM 00:35.200 --> 00:40.440 generate some code and sort of go with it, tweak it a bit, generate some more, and just make so much 00:40.440 --> 00:48.840 progress in this sort of mode of working this, this ad hoc vibe coding way of navigating around things 00:48.840 --> 00:50.160 like new frameworks. 00:50.560 --> 00:56.680 So I think this is wonderful and I strongly encourage vibe coding, and I imagine that most of you are 00:56.680 --> 00:57.560 very good at it. 00:57.600 --> 01:04.220 I did want to give a few tips that I think are important to do it well, because I think it's easy to 01:04.260 --> 01:11.220 do vibe coding and to get led astray by LMS and get yourself in trouble and get stuck, which is very 01:11.220 --> 01:12.100 unpleasant. 01:12.260 --> 01:14.900 So I have five tips to leave you with. 01:14.940 --> 01:19.060 But before we get into more detail with OpenAI agents SDK and here they are. 01:19.460 --> 01:21.500 First of all, good vibes. 01:21.740 --> 01:27.460 It's important to spend time getting your prompt to the LLM to be really good that you can reuse. 01:27.460 --> 01:31.260 Lots of times you should ask for short answers. 01:31.300 --> 01:37.580 Llms tend to be quite verbose in the way that in their code that they generate, they like packaging 01:37.580 --> 01:42.060 everything with lots of exception handlers, and they tend to do things in quite long winded way. 01:42.260 --> 01:50.940 Try and ask it to come up with concise, clean code and also mention today's date and say, please make 01:50.980 --> 01:54.700 sure that you use APIs that are current as of this date. 01:54.900 --> 02:01.520 Otherwise, llms have a nasty tendency to use older APIs because that was in a lot of their training 02:01.520 --> 02:07.480 data, so explicitly prompt for that vibe, but verify. 02:07.520 --> 02:12.400 My second tip don't just ask an LLM a question and go with it. 02:12.400 --> 02:13.840 Ask a couple of llms. 02:13.880 --> 02:19.400 So I often ask the same question to ChatGPT and to Claude and have them both up. 02:19.400 --> 02:23.280 I asked the question because I'll learn from both of the answers. 02:23.320 --> 02:28.920 Often one of them is too long winded or is missing the point a bit, and one of them will be spot on. 02:28.920 --> 02:34.360 And so asking a couple or maybe even three so that you're verifying what answers you're getting is a 02:34.360 --> 02:35.880 really good technique. 02:37.000 --> 02:38.240 Step up the vibe. 02:38.240 --> 02:42.360 So this is saying I think this is such a great one. 02:42.920 --> 02:48.760 This is because I sometimes get students sending me problems. 02:48.760 --> 02:55.600 They're saying I'm stuck with this and they'll send me some code and it will be like 200 lines of code 02:55.600 --> 02:56.840 and they'll say, it's not working. 02:56.840 --> 02:57.760 I don't know why. 02:57.800 --> 03:02.260 And when I look through the code, it's immediately obvious to me that this is they've been vibe coding. 03:02.460 --> 03:08.380 You can tell there are telltale signs that a lot of this was generated by an LLM, and it's it's unwieldy, 03:08.380 --> 03:11.420 and I can sometimes see a bunch of different bugs in it. 03:11.420 --> 03:13.220 And I come back and I say, it's no good. 03:13.220 --> 03:18.140 It's no good generating 200 lines of code and then saying it's broken. 03:18.340 --> 03:19.900 That's not the way to do it. 03:20.140 --> 03:24.940 Rather, you should always try and get llms to do things a little piece at a time. 03:24.940 --> 03:26.860 Generate function by function. 03:26.860 --> 03:32.900 Generate small pieces that are independently testable, like ten lines of code at a time. 03:32.900 --> 03:39.100 So you should think of dividing your problem down into small bite sized chunks and have an LLM do each 03:39.100 --> 03:39.820 piece of it. 03:40.260 --> 03:45.380 And here's the trick if you can't think about how you would divide your problem down into ten smaller 03:45.380 --> 03:50.020 steps or whatever, you don't need to because you can ask an LLM to do that. 03:50.020 --> 03:52.700 But be very clear, you don't want it to generate code. 03:52.740 --> 03:58.210 You say, look, I'm trying to solve the following problem and you tell it and say, what I want you 03:58.210 --> 04:07.650 to do is tell me what would be the 4 or 5 steps in a solution where each step is simple and simple, 04:07.690 --> 04:13.650 bite sized chunk that could be tested and verified independently and get it to come up with those, 04:13.690 --> 04:14.530 those chunks. 04:14.730 --> 04:21.490 And then you can ask again llms to generate code for each one in turn, along with the way to test it, 04:21.490 --> 04:24.770 so that you can make sure that ten lines of code at a time. 04:24.770 --> 04:29.090 You have a working solution, and you've convinced yourself that when you've got each of these, you'll 04:29.090 --> 04:33.130 be able to put it together and have 200 perfect lines of code. 04:33.850 --> 04:34.210 Okay. 04:34.250 --> 04:37.290 And then the fourth one, vibe and validate. 04:37.290 --> 04:38.890 Similar to vibe, but verify. 04:38.930 --> 04:42.450 But Vibe and validate is saying you can ask a question to an LLM. 04:42.490 --> 04:47.290 And then when it gives you the answer, you can go to another LLM or even the same one, but better 04:47.290 --> 04:51.770 to be another one and say, I asked this question, I've got this answer. 04:52.090 --> 04:57.990 Please confirm that that is an appropriate answer and that it's you can't do anything that's more concise 04:58.230 --> 05:02.630 or better structured or clearer, or there are no bugs in this. 05:02.630 --> 05:08.990 And by having another LLM go through and validate, it will often detect problems or make it be cleaner 05:08.990 --> 05:09.910 and simpler. 05:09.910 --> 05:11.510 So it's just a nice trick. 05:11.510 --> 05:18.310 And in many ways this this mirrors this echoes the common agentic design pattern of the evaluator optimizer. 05:18.310 --> 05:20.430 But you can just do that manually yourself. 05:20.430 --> 05:25.910 So taking a leaf out of Agentic design patterns and using it as part of vibe coding. 05:26.710 --> 05:33.030 And then my final one vibe with the variety, which is don't just say, can you generate code for this? 05:33.030 --> 05:37.550 But particularly if you're only talking about ten lines of code, say, can you give me three different 05:37.550 --> 05:42.870 solutions to this three, approach it three different ways, and maybe the response will be that there 05:42.870 --> 05:43.990 aren't three different ways of doing it. 05:43.990 --> 05:47.030 There's only one clear way, but you might get three different solutions. 05:47.070 --> 05:52.830 And and that will force the model into a mode where it's contemplating different ways of approaching 05:52.830 --> 05:53.790 the same problem. 05:53.790 --> 05:56.930 And as a result, you might get better solutions. 05:57.330 --> 06:03.170 As part of that, you can also ask it to explain itself, explain the rationale so that you'll get something 06:03.170 --> 06:07.530 that will that will also force it to think through why it's approaching it differently. 06:07.610 --> 06:12.610 Plus, as an added benefit, it will explain to you what's going on so that you can really understand 06:12.610 --> 06:13.050 this. 06:13.450 --> 06:18.290 And that does sort of lead to an obvious point, which I've not put here, which is that when you do 06:18.330 --> 06:23.930 vibe coding, you should always then go back and ask an LLM to explain it clearly. 06:23.930 --> 06:29.050 If it's not immediately clear to you, get to a point where you understand every single thing that's 06:29.050 --> 06:29.770 going on. 06:29.810 --> 06:35.450 Vibe coding is super fun and productive and powerful, but if you're not following what's actually happening, 06:35.450 --> 06:39.130 it's going to become painful and frustrating when something goes wrong. 06:39.250 --> 06:42.330 So it is important to stay in touch with what's going on. 06:42.730 --> 06:43.210 All right. 06:43.210 --> 06:49.050 And that is my survival guide to vibe coding and I hope you'll find that helpful. 06:49.050 --> 06:51.970 But other than that you should definitely do it. 06:52.010 --> 06:54.570 It's it's great fun and it's super productive.