WEBVTT 00:00:00.140 --> 00:00:03.690 So a key part of working with agents is crafting good 00:00:03.720 --> 00:00:07.250 prompts, and a key part of that is adding the right context. 00:00:07.260 --> 00:00:10.740 For example, let's say I wanna enhance my dummy CLI 00:00:10.780 --> 00:00:14.400 application here with another feature that also allows me to 00:00:14.660 --> 00:00:18.300 generate dummy products, not just users. 00:00:18.340 --> 00:00:21.860 So we could open up a new chat session in agent mode and 00:00:21.900 --> 00:00:23.260 say, "I want 00:00:25.280 --> 00:00:28.080 to enhance my CLI tool 00:00:29.560 --> 00:00:33.120 to also generate dummy products 00:00:33.700 --> 00:00:37.560 depending on the flags 00:00:37.620 --> 00:00:41.550 with which it was executed." I then also provide some 00:00:41.580 --> 00:00:45.210 details on how my dummy product should look like, because again, being 00:00:45.260 --> 00:00:48.940 specific is a good idea. You can use plan mode, as you saw 00:00:48.960 --> 00:00:52.160 before, and it will ask for follow-up questions, 00:00:52.220 --> 00:00:56.140 have to hope that it asks for the things you 00:00:56.280 --> 00:01:00.140 want it to ask. If you already know, for example, the shape of a dummy 00:01:00.180 --> 00:01:04.160 product, in this case, tell it. You don't need to hope for the AI 00:01:04.200 --> 00:01:07.730 to guess it or to ask for it. Still, of course, 00:01:07.820 --> 00:01:10.920 plan mode, but you wanna be as specific as 00:01:10.960 --> 00:01:14.920 possible. And as mentioned, you wanna add the right context and 00:01:14.960 --> 00:01:18.860 maybe point at some examples. Here I know that I already have a CLI 00:01:18.960 --> 00:01:22.680 tool that generates dummy users. Now, as mentioned, in a 00:01:22.720 --> 00:01:26.670 simple project like this, the chances of the AI detecting that 00:01:26.720 --> 00:01:30.500 other file are almost at 100%, but still, I 00:01:30.620 --> 00:01:34.320 know that there is this main.py file which has that other CLI 00:01:34.360 --> 00:01:37.880 program, so why wouldn't I explicitly point the AI 00:01:37.920 --> 00:01:41.380 it? Now here actually, it already added main.py as 00:01:41.420 --> 00:01:45.400 context to this prompt. You can see the context here, 00:01:45.460 --> 00:01:49.000 because that's the editor I now have open, and the open 00:01:49.160 --> 00:01:52.040 editor window will always be added as context. 00:01:52.080 --> 00:01:56.030 If you select some text here, for example, the selected code will 00:01:56.060 --> 00:01:59.040 be the context that's added. If you don't have a selection, 00:01:59.120 --> 00:02:02.160 file. But let's say this was closed. 00:02:02.200 --> 00:02:05.780 In that case, as you can see, there's no context here, and you can add 00:02:05.840 --> 00:02:09.400 context with the add context button, and there's a lot of context you can 00:02:09.460 --> 00:02:13.300 add. You can add files and folders, you could take a screenshot, you 00:02:13.360 --> 00:02:16.600 can also paste images in here, and a lot of other 00:02:16.700 --> 00:02:19.920 things. Alternatively, you can also 00:02:21.220 --> 00:02:25.200 use the hash key and then bring up this menu to 00:02:25.280 --> 00:02:29.160 open context from here whilst typing, and you can enter 00:02:29.320 --> 00:02:31.360 filenames here too, like main.py. 00:02:31.400 --> 00:02:34.800 So this will now point at that file and add it as a context here. 00:02:36.360 --> 00:02:39.960 Now again, I wanna make sure that we also take the official 00:02:40.000 --> 00:02:43.940 account, and therefore I'll go to it here, copy that URL, and you can 00:02:43.960 --> 00:02:47.920 add that as context too. So we could say, "Also take a 00:02:47.980 --> 00:02:51.890 look at the official docs," and 00:02:52.000 --> 00:02:55.940 paste that link in there, and if we wanna improve our chances of it 00:02:56.100 --> 00:03:00.000 really making that request to the docs, you can use the fetch 00:03:00.820 --> 00:03:04.000 context tool here, and then add the link 00:03:04.020 --> 00:03:07.769 here. This gives it an extra hint that it should fetch this 00:03:07.900 --> 00:03:11.040 page and send a request to it and get the content. 00:03:11.049 --> 00:03:14.960 And now if I hit enter, it'll therefore go ahead 00:03:15.000 --> 00:03:18.850 this prompt and all the context I included in there, and that includes 00:03:18.940 --> 00:03:22.900 a request to these docs hopefully. 00:03:22.940 --> 00:03:26.540 Yeah, it asks me if I wanna allow it sending 00:03:26.580 --> 00:03:27.930 do. 00:03:30.600 --> 00:03:34.470 I also don't wanna review the results here, the results of that request. 00:03:34.500 --> 00:03:38.140 It should just go ahead. Same for this request. 00:03:38.160 --> 00:03:41.940 And now based on all that information it gathered 00:03:42.020 --> 00:03:45.100 spits out that plan it wants to implement. 00:03:45.160 --> 00:03:48.340 It also has some follow-up questions here, and I'll provide my 00:03:48.420 --> 00:03:51.600 answers. Hit enter to get a refined 00:03:51.720 --> 00:03:52.420 plan, 00:03:54.500 --> 00:03:58.480 refine it again with another answer to another question, and then just 00:03:58.520 --> 00:04:02.340 as before, I'll press start implementation, 00:04:02.400 --> 00:04:06.260 this. And now it'll, just as before, go ahead 00:04:06.360 --> 00:04:10.100 in agent mode now and edit my files or maybe create some new 00:04:10.140 --> 00:04:12.590 files to implement this new feature. 00:04:12.620 --> 00:04:16.450 So here it now edited this main.py file and it wants to run it with 00:04:16.480 --> 00:04:20.370 the help command to see if that works, and I'll allow this command, and 00:04:20.399 --> 00:04:24.220 it can indeed also execute terminal commands, as you see here. 00:04:24.260 --> 00:04:26.660 But as always, it asks for permission first. 00:04:26.720 --> 00:04:30.090 Now it wants to run the command to actually generate some products, 00:04:30.120 --> 00:04:33.980 of course carefully take a look at these commands so that it doesn't do 00:04:34.060 --> 00:04:37.720 anything you don't want it to do. But I'm fine with this command here. 00:04:37.740 --> 00:04:41.580 I can see that this works, and you therefore see this agent is really 00:04:41.660 --> 00:04:45.500 testing its own program. And of course, I could also in general 00:04:45.540 --> 00:04:49.350 allow all these UV commands so that I'm not 00:04:49.400 --> 00:04:53.350 constantly asked again and again and again for it here with 00:04:53.440 --> 00:04:57.360 that specific command line tool, which I'm fine with when the 00:04:57.400 --> 00:05:01.340 AI uses it. So it now tested it, it was able to 00:05:01.420 --> 00:05:05.180 see that terminal output, that's another feature this agent mode 00:05:05.240 --> 00:05:08.640 has, and with that, it wrote the code and on its own 00:05:08.700 --> 00:05:11.160 validated that it works correctly. 00:05:11.200 --> 00:05:15.080 Of course, you still should also review the code on 00:05:15.120 --> 00:05:18.830 own, take a look at it, and see if it's in line with your expectations and 00:05:18.900 --> 00:05:22.500 requirements before you then accept or deny it. 00:05:22.540 --> 00:05:25.990 You can, of course, by the way, also ask for follow-up edits or 00:05:26.040 --> 00:05:29.180 improvements. It's not just accept or decline. 00:05:29.220 --> 00:05:33.140 You can also ask for improvements, but here I'm fine with it 00:05:33.180 --> 00:05:35.160 it.