WEBVTT 00:01.650 --> 00:03.090 -: Okay, I'm gonna run you 00:03.090 --> 00:07.260 through how to do analysis of a publication. 00:07.260 --> 00:11.610 The publication I'm looking at today is Lenny's blog. 00:11.610 --> 00:12.443 As you can see, 00:12.443 --> 00:14.970 here are the top posts on Lenny's blog 00:14.970 --> 00:16.890 and I wanted to scrape all of these 00:16.890 --> 00:19.860 and basically summarize what are the sorts of things 00:19.860 --> 00:21.840 that this blog is talking about. 00:21.840 --> 00:23.520 What are some insights I can get? 00:23.520 --> 00:27.030 And if I was pitching this blog or similar publication, 00:27.030 --> 00:28.770 what sort of titles 00:28.770 --> 00:31.200 should I pitch them that would be interesting? 00:31.200 --> 00:33.540 Or potentially what templates are working well for them 00:33.540 --> 00:36.120 that I can use for my own blog? 00:36.120 --> 00:39.993 Okay, so let me show you how this works. 00:41.490 --> 00:44.010 First, I'm gonna install Anthropic, 00:44.010 --> 00:45.660 so I have that running 00:45.660 --> 00:50.070 and going to load the environment, 00:50.070 --> 00:52.950 and that's where I have my OpenAI API key. 00:52.950 --> 00:54.570 I'm just gonna do that in a second. 00:54.570 --> 00:58.080 I'm just gonna pull in here the blog posts. 00:58.080 --> 01:01.410 So this is what you need to get from scraping 01:01.410 --> 01:03.480 and this isn't the scraping tutorial. 01:03.480 --> 01:05.400 This is more on the analysis 01:05.400 --> 01:06.390 and to be honest, 01:06.390 --> 01:08.490 I don't really know that much about web scraping. 01:08.490 --> 01:11.010 I literally just asked ChatGPT 01:11.010 --> 01:13.630 and this is the code that it gave me 01:15.420 --> 01:17.100 to scrape the blog. 01:17.100 --> 01:18.810 And if we go, 01:18.810 --> 01:23.730 let's see, if we pop open this blog again, 01:23.730 --> 01:25.480 let's see if this code still works. 01:26.790 --> 01:28.473 I'm just gonna paste into, 01:30.720 --> 01:33.010 here we go, data copied to clipboard 01:34.020 --> 01:37.420 and basically what this has done 01:39.510 --> 01:43.740 is pasted all the different blog posts and titles. 01:43.740 --> 01:48.573 And what I've then done is put that into this CSV, 01:50.070 --> 01:51.030 paste it into here, 01:51.030 --> 01:52.350 and it's got all the titles, 01:52.350 --> 01:53.640 it's got when it was published, 01:53.640 --> 01:55.020 what the author is, et cetera. 01:55.020 --> 01:56.970 That is what I did for scraping. 01:56.970 --> 02:00.000 I literally just asked ChatGPT to write that code for me, 02:00.000 --> 02:03.060 so I can't teach you that much about web scraping. 02:03.060 --> 02:05.790 But once you have all the posts in here, 02:05.790 --> 02:08.700 then you can download them 02:08.700 --> 02:10.413 and I've done that down here. 02:12.930 --> 02:17.930 So you can see it downloaded all the texts from each post. 02:22.020 --> 02:25.290 Yeah, and the way I've done that just to show you 02:25.290 --> 02:30.290 is if you load the links 02:30.540 --> 02:32.880 and then you get a list of the URLs, 02:32.880 --> 02:35.580 and this is going, just to show you, 02:35.580 --> 02:37.060 it's going through each URL 02:37.920 --> 02:39.810 and it's getting the link 02:39.810 --> 02:41.580 which is the second part, 02:41.580 --> 02:44.160 and then it's finding the file name, 02:44.160 --> 02:49.160 and then it's just reading the text using Beautiful Soup 02:49.860 --> 02:52.620 and then just saving that text locally. 02:52.620 --> 02:54.750 So that's how I downloaded everything 02:54.750 --> 02:56.460 and I just sleep for three seconds here. 02:56.460 --> 02:57.780 It depends on the website. 02:57.780 --> 02:59.220 Again, I'm not a web scrap pro. 02:59.220 --> 03:00.870 I'm not particularly good at this. 03:00.870 --> 03:03.060 This isn't a web scraping tutorial. 03:03.060 --> 03:06.840 Yeah, basically this was also written by ChatGPT, 03:06.840 --> 03:08.580 but if I run into an error, 03:08.580 --> 03:10.257 I just ask ChatGPT 03:10.257 --> 03:11.340 and it will fix it. 03:11.340 --> 03:13.260 So if that's how that works, 03:13.260 --> 03:15.750 so that will download the different URLs. 03:15.750 --> 03:17.640 You'd have them all the text in here. 03:17.640 --> 03:19.800 So that's a prerequisite. 03:19.800 --> 03:21.550 I'm just going to interrupt that 03:22.890 --> 03:24.270 'cause I've already done that part 03:24.270 --> 03:27.300 and it took I think a maybe about an hour 03:27.300 --> 03:29.160 to download this stuff. 03:29.160 --> 03:31.320 All right, so now onto the actual tutorial though 03:31.320 --> 03:36.180 which is how do we summarize this text? 03:36.180 --> 03:39.423 So just gonna load Anthropic first. 03:41.070 --> 03:42.360 Make sure this is working. 03:42.360 --> 03:43.193 Yeah, here we go. 03:43.193 --> 03:45.330 So the way you call Anthropic, 03:45.330 --> 03:48.270 this is Clause Three is you set up the client, 03:48.270 --> 03:51.603 it gets the API key that you determined. 03:53.070 --> 03:55.170 If I just close this, 03:55.170 --> 03:56.820 in your .env file. 03:56.820 --> 03:57.750 I won't show you mine, 03:57.750 --> 04:01.560 but you just need to have something 04:01.560 --> 04:05.310 called Anthropic_API_KEY and then equals 04:05.310 --> 04:07.890 and then the actual key that you get from Anthropic. 04:07.890 --> 04:09.270 But assuming you have that, 04:09.270 --> 04:10.770 then it will load Anthropic 04:10.770 --> 04:14.430 and then you can get, it's very similar to OpenAI API. 04:14.430 --> 04:16.500 You can just specify the model 04:16.500 --> 04:17.640 and then you send the messages. 04:17.640 --> 04:18.600 And I just said, "Hello," 04:18.600 --> 04:20.520 and it said, "Hello, how can I assist you?" 04:20.520 --> 04:22.170 That's how I know it's working. 04:22.170 --> 04:25.830 Now what I need to do to analyze this 04:25.830 --> 04:27.420 is to just loop through 04:27.420 --> 04:31.920 and pull out different texts from this folder 04:31.920 --> 04:35.460 and then place that into the prompt 04:35.460 --> 04:36.930 and then here, I'm just asking you 04:36.930 --> 04:39.510 to pull out one non-obvious insight from this text. 04:39.510 --> 04:41.520 So it's basically reading the text 04:41.520 --> 04:44.130 and then coming back to me with something interesting. 04:44.130 --> 04:46.260 So let's see how this works. 04:46.260 --> 04:50.010 Okay, so that took 46 seconds for five URLs. 04:50.010 --> 04:51.150 Obviously we could do it for everything, 04:51.150 --> 04:53.280 but here we go. 04:53.280 --> 04:55.260 From this blog post, 04:55.260 --> 04:57.180 it said that borrowing successful features 04:57.180 --> 04:58.410 from other products requires 04:58.410 --> 05:00.690 carefully adapting them to your specific context 05:00.690 --> 05:02.820 and it has a quote here which is pretty cool. 05:02.820 --> 05:04.500 This one, "How the biggest consumer apps 05:04.500 --> 05:06.480 got the first 1,000 users," 05:06.480 --> 05:09.300 they found the early users from just a single strategy. 05:09.300 --> 05:11.610 That's pretty surprising, right? 05:11.610 --> 05:15.000 Another one here is that ChatGPT 05:15.000 --> 05:16.560 is good for non-engineering teams, 05:16.560 --> 05:18.030 not just engineering teams. 05:18.030 --> 05:19.800 So summarizing customer interviews, 05:19.800 --> 05:21.930 identifying regarding pain points, 05:21.930 --> 05:24.420 finding compelling customer testimonials, 05:24.420 --> 05:26.220 generating action items. 05:26.220 --> 05:27.870 I think that's pretty interesting, 05:27.870 --> 05:29.430 but this is a great way to decide 05:29.430 --> 05:31.590 whether you wanna read the post in the first place 05:31.590 --> 05:33.300 and also to start to notice patterns 05:33.300 --> 05:34.560 about what types of things 05:34.560 --> 05:36.810 are being written about on this blog. 05:36.810 --> 05:38.550 That's quite cool, I find that quite useful. 05:38.550 --> 05:40.950 You could also change the prompt to say whatever you want. 05:40.950 --> 05:42.900 So you could ask it to give you a summary 05:42.900 --> 05:46.650 or to highlight what the quote is about 05:46.650 --> 05:48.090 or whatever it is. 05:48.090 --> 05:49.350 Pull out bullet points 05:49.350 --> 05:51.030 or whatever you would like to do. 05:51.030 --> 05:54.120 But where it gets really interesting is the Claude 05:54.120 --> 05:58.020 has a 200,000 token limit in terms of the context windows. 05:58.020 --> 05:58.853 So you can fit 05:58.853 --> 06:01.860 pretty much an entire Harry Potter book in there 06:01.860 --> 06:03.900 and therefore you can just pull 06:03.900 --> 06:06.240 a bunch of these texts together 06:06.240 --> 06:08.250 and start to look for patterns. 06:08.250 --> 06:11.160 What I did here is if you run this, 06:11.160 --> 06:13.020 it just pulls out the first 50 URLs 06:13.020 --> 06:17.610 and then it appends the text into tags here, 06:17.610 --> 06:21.390 and then it's just telling me how many words basically, 06:21.390 --> 06:22.680 how many tokens is an estimate. 06:22.680 --> 06:24.870 One token is roughly four characters, 06:24.870 --> 06:27.900 so just counted here. 06:27.900 --> 06:30.510 So I'm just under 200,000 which is great. 06:30.510 --> 06:32.283 If I look at that one, 06:35.220 --> 06:39.210 all texts, you can see that it starts with article 06:39.210 --> 06:41.490 and then it has all the texts. 06:41.490 --> 06:43.230 I could probably strip out some of this white space 06:43.230 --> 06:46.650 and then it divides the articles with a line like this. 06:46.650 --> 06:48.840 And the reason I'm using these tags like this, 06:48.840 --> 06:50.610 it's almost like HTML, 06:50.610 --> 06:55.320 is that Claude was trained on tags 06:55.320 --> 06:57.060 and responds really well for them. 06:57.060 --> 07:00.000 So that's just a specific Claude thing 07:00.000 --> 07:02.160 that I recommend you do. 07:02.160 --> 07:03.210 You just get better results 07:03.210 --> 07:05.640 'cause it delineates the specific things 07:05.640 --> 07:06.480 that you're talking about. 07:06.480 --> 07:08.520 In this case, the articles. 07:08.520 --> 07:10.050 So now what we're gonna do 07:10.050 --> 07:13.080 is we've got all of our text here, 07:13.080 --> 07:15.120 so we're reading all text 07:15.120 --> 07:16.770 and then we're sending to Claude Opus. 07:16.770 --> 07:18.300 We're saying summarize these articles, 07:18.300 --> 07:20.670 analyze the word choice, identify any patterns, 07:20.670 --> 07:23.970 and tell me about any non-obvious insights publications. 07:23.970 --> 07:26.370 So now we're passing in all of those articles, 07:26.370 --> 07:29.130 50 articles, the full text of each, 07:29.130 --> 07:31.440 all in one prompt which is pretty cool 07:31.440 --> 07:34.080 and we'll just see what it comes back with. 07:34.080 --> 07:35.250 Okay, we've got the response. 07:35.250 --> 07:36.570 It took about two minutes 07:36.570 --> 07:38.760 because this is a smart model 07:38.760 --> 07:40.800 and therefore it's relatively slower, 07:40.800 --> 07:43.290 but also we're passing it a ton of text, right? 07:43.290 --> 07:45.000 So the longer the prompt, 07:45.000 --> 07:48.660 the longer the time it takes to get back to you. 07:48.660 --> 07:50.040 So analyze these articles. 07:50.040 --> 07:52.350 It did hallucinate, it thinks there's 40 articles, 07:52.350 --> 07:53.730 but there are actually 50. 07:53.730 --> 07:55.230 It's not very good at accounting, 07:55.230 --> 07:58.560 and it summarizes, so it says they cover 07:58.560 --> 08:00.360 a wide range of topics related to product management, 08:00.360 --> 08:02.490 growth, career development, entrepreneurship 08:02.490 --> 08:04.140 and then it gives some common themes, 08:04.140 --> 08:06.990 tactical advice and frameworks, inside stories. 08:06.990 --> 08:08.370 Interviews with experienced leaders, 08:08.370 --> 08:09.780 guidance for early-stage startups, 08:09.780 --> 08:11.550 and personal stories of career journeys. 08:11.550 --> 08:15.630 So already I know if I wanted to do a guest post for Lenny 08:15.630 --> 08:17.310 and I was coming up with new ideas, 08:17.310 --> 08:18.750 I could take a look here. 08:18.750 --> 08:21.630 The content's highly practical, example-driven, 08:21.630 --> 08:23.640 and grounded in real-world experience 08:23.640 --> 08:25.410 so that's important to know. 08:25.410 --> 08:27.660 They go deep on specific topics. 08:27.660 --> 08:29.400 The word style is conversational, direct, 08:29.400 --> 08:32.190 and easy to follow whilst covering complex topics. 08:32.190 --> 08:34.950 There's also frequent characteristics of the writing, 08:34.950 --> 08:38.370 so this is gonna help you get past the editing process 08:38.370 --> 08:40.620 if you are writing a post for this publication. 08:40.620 --> 08:43.200 But again, this could be used for any publication 08:43.200 --> 08:45.090 or even to analyze your own blog 08:45.090 --> 08:47.250 to understand what's interesting about it. 08:47.250 --> 08:48.810 Frequent use the second person "you" 08:48.810 --> 08:50.520 to speak directly to the reader. 08:50.520 --> 08:52.470 A liberal use of bolded numbered lists, 08:52.470 --> 08:53.760 bolding of key phrases, 08:53.760 --> 08:56.520 relatively simple vocabulary, et cetera. 08:56.520 --> 08:58.560 And then the writing is more casual 08:58.560 --> 09:02.220 and bloggy in style compared to traditional business writing 09:02.220 --> 09:05.790 and the there's also more information here. 09:05.790 --> 09:09.120 Kept going, it figured out the reading level. 09:09.120 --> 09:12.480 Some non-obvious patterns, his own writing accounts 09:12.480 --> 09:14.940 for less than half the articles, that's pretty interesting. 09:14.940 --> 09:17.670 And there many compendium-style posts like best-ofs. 09:17.670 --> 09:21.270 There is also failure stories as much as success stories, 09:21.270 --> 09:23.520 organic, product-led growth over a paid acquisition. 09:23.520 --> 09:26.220 Very little mention of ads, that's useful to know. 09:26.220 --> 09:29.010 The design is very rarely mentioned. 09:29.010 --> 09:30.360 That's actually surprising to me 09:30.360 --> 09:33.450 considering it's a product management-focused blog. 09:33.450 --> 09:35.070 Analytics and metrics are important 09:35.070 --> 09:38.400 and there's a bias towards VC-backed startups. 09:38.400 --> 09:40.380 So that's really interesting to know 09:40.380 --> 09:41.460 and maybe some of these things 09:41.460 --> 09:44.340 Lenny himself might not have realized. 09:44.340 --> 09:46.770 Well, now the other thing I wanted to do 09:46.770 --> 09:48.540 is just get all the titles 09:48.540 --> 09:51.270 and think about what a successful title might look like 09:51.270 --> 09:55.590 if I was pitching Lenny to join his blog as a guest speaker. 09:55.590 --> 09:57.960 This is a prompt I've used a bunch of times in the past. 09:57.960 --> 10:00.810 It's based on this thing called the Blank. 10:00.810 --> 10:03.420 This is from a Tim Ferriss interview with James Clear, 10:03.420 --> 10:05.460 the author of "Atomic Habits." 10:05.460 --> 10:07.470 He found when he looked 10:07.470 --> 10:12.470 at all the most successful titles of business books, 10:12.810 --> 10:15.127 that this was a successful template. 10:15.127 --> 10:16.447 "The War of Art," 10:16.447 --> 10:17.677 "The Psychology of Money," 10:17.677 --> 10:20.040 "The Subtle Art of Not Giving a --." 10:20.040 --> 10:21.450 Basically you take two things 10:21.450 --> 10:23.460 that are not normally associated 10:23.460 --> 10:24.870 and then put them together, 10:24.870 --> 10:26.497 the blank of blank. 10:26.497 --> 10:28.890 "Odd Topic" is the template he chose for his book, 10:28.890 --> 10:31.440 so habits are not normally atomic. 10:31.440 --> 10:34.140 This doesn't have the blank of blank. 10:34.140 --> 10:36.667 Instead it's just a odd combination. 10:36.667 --> 10:39.067 "Extreme Ownership," ownership's not normally extreme. 10:39.067 --> 10:41.220 "Deep Work," work is not usually deep. 10:41.220 --> 10:43.410 So that's what James Clear choose 10:43.410 --> 10:45.813 and then I just added this one as well. 10:46.770 --> 10:47.850 I was using this prompt 10:47.850 --> 10:49.920 to come up with the title of my O'Reilly book 10:49.920 --> 10:54.300 and this is an example of the type of O'Reilly books. 10:54.300 --> 10:56.160 So if you're not getting good success, 10:56.160 --> 10:57.570 by the way, with this template, 10:57.570 --> 11:01.980 you can just add multiple new templates into here 11:01.980 --> 11:04.680 that cover the types of things you wanna see 11:04.680 --> 11:06.150 and then it will do a much better job 11:06.150 --> 11:08.160 for that specific publication. 11:08.160 --> 11:09.150 All right, what we're trying to do, 11:09.150 --> 11:10.440 we're passing it all the titles 11:10.440 --> 11:14.850 and we just wanna see what types of templates come back 11:14.850 --> 11:17.400 in terms of what are the types of titles 11:17.400 --> 11:19.980 that appear on Lenny's blog. 11:19.980 --> 11:23.280 So it's gonna come up with its own templates like this 11:23.280 --> 11:24.930 if it works well. 11:24.930 --> 11:27.817 Okay, here we go, have some templates. 11:27.817 --> 11:29.130 "How X does Y" 11:29.130 --> 11:30.510 and giving some examples. 11:30.510 --> 11:32.190 How Duolingo reignited user growth, 11:32.190 --> 11:33.210 how Linear builds product, 11:33.210 --> 11:34.680 how Shopify builds product. 11:34.680 --> 11:36.480 That's pretty interesting to me. 11:36.480 --> 11:39.810 Tells me if I work for a successful company like Duolingo, 11:39.810 --> 11:41.503 I could do something similar to that. 11:41.503 --> 11:43.020 I just choose a specific topic 11:43.020 --> 11:45.667 and they say how this company does X. 11:45.667 --> 11:47.427 "How to X," how to use ChatGPT, 11:47.427 --> 11:49.170 how to validate your B@B startup idea, 11:49.170 --> 11:51.270 how to pass a first-round interview. 11:51.270 --> 11:52.860 So that's pretty interesting, 11:52.860 --> 11:56.280 and "The X of Y" or "The X Y." 11:56.280 --> 11:58.200 So if the first word is an adjective 11:58.200 --> 12:00.720 or noun describing an important aspect 12:00.720 --> 12:02.790 and then that's related to the main topic. 12:02.790 --> 12:05.070 So the 10 commandments of salary negotiation, 12:05.070 --> 12:06.840 the ultimate guide to willingness-to-pay, 12:06.840 --> 12:09.300 the unconventional Palantir principles. 12:09.300 --> 12:10.620 And then "X for Y." 12:10.620 --> 12:13.350 So good strategy, bad strategy, okay. 12:13.350 --> 12:15.330 There's some interesting templates in here 12:15.330 --> 12:16.980 and you'll get slightly different ones 12:16.980 --> 12:18.690 if you run it again and again, 12:18.690 --> 12:22.410 but this is a pretty useful prompt I found 12:22.410 --> 12:25.641 just to get really good templates out of titles 12:25.641 --> 12:27.000 and it gives you a good sense 12:27.000 --> 12:30.540 of what are the types of things you should start looking at 12:30.540 --> 12:33.600 over time in order to hitch this blog 12:33.600 --> 12:36.780 or maybe just learn from what works for this blog 12:36.780 --> 12:37.680 because if they're using 12:37.680 --> 12:39.390 the same sort of template again and again, 12:39.390 --> 12:41.370 it's because it's working. 12:41.370 --> 12:42.630 The editor thinks it's working. 12:42.630 --> 12:45.300 Either they'd know and they're doing it consciously 12:45.300 --> 12:48.000 and they're going out finding other examples 12:48.000 --> 12:49.577 or they don't know, 12:49.577 --> 12:52.260 and it just happens to be successful 12:52.260 --> 12:54.720 and it's something you can learn from as well. 12:54.720 --> 12:57.870 So if you're pitching a publication to guest posts 12:57.870 --> 12:59.010 or if you're just trying to understand 12:59.010 --> 13:01.050 what makes a publication successful, 13:01.050 --> 13:02.850 these combination of prompts 13:02.850 --> 13:06.093 combined with some web scraping is really helpful.