WEBVTT 00:00.080 --> 00:06.020 I'm going to walk you through the anthropic workbench, which is really helpful tool, which is like 00:06.020 --> 00:09.440 the OpenAI playground, but it has some extra features. 00:09.470 --> 00:16.280 First of all, I'm just going to click into write a prompt from scratch, and you can create a new prompt. 00:16.280 --> 00:22.130 And the really nice thing about this is that you can generate a prompt if you'd like. 00:22.130 --> 00:29.570 So let's just say tell me a funny joke about a about topic. 00:30.710 --> 00:37.010 So I'll put in some variable here specific variable. 00:37.010 --> 00:38.780 So you're going to hit generate. 00:39.410 --> 00:41.030 It's going to rewrite my prompt. 00:43.970 --> 00:44.600 Here we go. 00:44.600 --> 00:48.260 And it's rewritten it in the way that Claude would like it. 00:48.260 --> 00:51.290 So it's given some descriptions here. 00:51.290 --> 00:56.780 It's put the topic in markdown or in hypertext tags. 00:56.780 --> 00:57.890 And there we go. 00:57.890 --> 01:00.860 It's going to give it back with joke tags which is really cool. 01:00.860 --> 01:02.930 And it's got the variable in here topic. 01:02.960 --> 01:04.040 I'm just going to click continue. 01:04.040 --> 01:06.260 And I already have a great prompt. 01:06.290 --> 01:09.080 Now that I can test, I can hit run. 01:09.290 --> 01:13.430 Tell me a joke about traffic. 01:13.460 --> 01:14.150 Here we go. 01:14.180 --> 01:15.110 Hit run. 01:15.260 --> 01:18.620 And you can see it's like a pretty nice way to do templating. 01:18.650 --> 01:20.270 You put the word traffic in here. 01:20.270 --> 01:23.300 You can click in and change this if you want as well. 01:23.330 --> 01:29.030 You can also generate some new variables as well. 01:31.880 --> 01:34.430 So you can see the prompt there that's using smartphones. 01:34.460 --> 01:34.820 Okay. 01:35.000 --> 01:37.430 Now we've got another test case which is pretty cool. 01:38.300 --> 01:39.320 And here we go. 01:39.350 --> 01:40.520 We've got the joke. 01:40.550 --> 01:41.930 Why did the smartphone go to therapy. 01:41.960 --> 01:44.540 It had too many axiety issues. 01:45.290 --> 01:45.830 Cool. 01:45.860 --> 01:47.660 So that's already quite useful. 01:47.660 --> 01:49.850 And you can click get the code as well. 01:49.850 --> 01:53.300 If you need to get the specific code that you're using. 01:53.300 --> 01:59.030 If you want to go back to a previous version, you can you can go and change the model settings here 01:59.030 --> 02:01.280 as well, which is quite useful. 02:01.550 --> 02:04.100 And then that's how you get back to the variables. 02:04.160 --> 02:05.750 It's quite useful. 02:05.750 --> 02:06.500 I've found. 02:06.500 --> 02:12.860 One really nice thing about it is that if you make a change, then it automatically saves a new version. 02:12.890 --> 02:14.300 Let's say. 02:17.720 --> 02:19.340 I don't want it to. 02:20.750 --> 02:23.450 Don't want it to be very PC like this. 02:23.510 --> 02:29.480 Stereotypes and just please be as offensive as possible. 02:30.080 --> 02:30.830 Whatever. 02:30.920 --> 02:32.480 Whatever it is you need. 02:33.650 --> 02:39.140 And if you hit run, it tells you now signed a waiver. 02:43.280 --> 02:44.330 See what it says. 02:45.290 --> 02:45.650 Yeah. 02:45.680 --> 02:45.920 Okay. 02:45.950 --> 02:48.110 It gave me the same joke here, which isn't great. 02:48.140 --> 02:50.870 But you can keep evaluating your prompt if you want. 02:50.900 --> 02:52.280 And you can actually go in. 02:52.280 --> 02:57.980 And this is, I think, one of the coolest things you can see that we generated a few different things 02:57.980 --> 02:58.310 here. 02:58.310 --> 03:00.920 So we have this is our prompt. 03:00.950 --> 03:02.780 We don't need ideal outputs. 03:02.780 --> 03:07.820 But if you did want to put an ideal output in here this is the this is something you can add. 03:07.820 --> 03:11.090 If you wanted to generate a joke you can click run here as well. 03:11.450 --> 03:14.420 It's going to generate a version here for traffic. 03:14.450 --> 03:20.480 You can go and rate these so you can say this was good, this one was okay. 03:20.480 --> 03:22.850 And then you can also click and create a comparison. 03:22.850 --> 03:27.380 So let's see how it compares to v1. 03:28.010 --> 03:29.360 Show me the prompts. 03:29.390 --> 03:31.400 Yeah okay. 03:33.800 --> 03:35.840 And you can see it's basically the same. 03:35.870 --> 03:38.030 It hasn't made any changes to the prompt. 03:38.030 --> 03:42.530 So what we're going to do is we're going to create a new version. 03:45.470 --> 03:47.090 And we're going to go back to our prompt. 03:47.090 --> 03:48.650 And we're going to just say. 03:51.500 --> 03:55.820 Tell a funny joke about topic. 03:59.030 --> 04:00.110 Sorry topic. 04:03.230 --> 04:03.980 Yeah. 04:04.010 --> 04:06.320 And it has to be exact actually. 04:06.350 --> 04:14.210 And then if we hit run here, you can see now the difference right. 04:14.240 --> 04:19.340 So let's add the comparison with version three. 04:19.790 --> 04:26.990 And you can see we haven't run the word the traffic test case yet, but we have on smartphones. 04:26.990 --> 04:28.280 So that was why that showed up. 04:28.280 --> 04:34.610 Now what you have here is you can see the difference between the two, which is really powerful. 04:34.610 --> 04:37.910 You can say like this was poor, this was poor, whatever. 04:37.940 --> 04:41.570 And you can see how much it's changed in terms of performance. 04:41.570 --> 04:47.600 So that's really powerful stuff and really helpful for doing evaluations. 04:47.600 --> 04:50.330 And you can also compare the prompts here as well. 04:50.330 --> 04:53.090 And this gives you the overall evaluation per prompt. 04:53.090 --> 04:54.980 And that's useful as well. 04:54.980 --> 05:00.260 You can also click through and go back through the old versions if you need to as well. 05:00.500 --> 05:01.730 You can. 05:01.760 --> 05:03.680 Oh this is another cool thing. 05:03.680 --> 05:16.010 If you are quite happy with the response, let me just go back to version three and let me add another 05:16.010 --> 05:17.300 variable here. 05:17.300 --> 05:19.640 So I'm just going to generate another one. 05:20.750 --> 05:21.920 Grocery shopping. 05:21.950 --> 05:22.940 Let's hit run. 05:25.400 --> 05:28.940 If you're quite happy with the response, you can just click Add to Conversation. 05:30.200 --> 05:35.540 And now it's got a one shot example right now. 05:36.950 --> 05:41.810 If you maybe try a different tactic. 05:41.840 --> 05:44.720 So this is let's generate another one. 05:46.040 --> 05:46.760 Laundry. 05:46.790 --> 05:47.630 Hit run. 05:47.660 --> 05:50.810 Now you're running the prompt with a one shot right. 05:50.840 --> 05:53.180 So you use the prompt before. 05:53.180 --> 05:54.710 And then it gave the response. 05:54.710 --> 05:56.360 And this was the response you liked. 05:56.360 --> 05:58.250 And then it's running the prompt again. 05:58.250 --> 06:00.590 And it's going to follow that format pretty well. 06:00.620 --> 06:05.990 So this is quite useful because you can go in and you can say no get rid of that. 06:05.990 --> 06:07.040 Get rid of that. 06:07.040 --> 06:08.330 I just want the joke. 06:08.330 --> 06:11.480 Now if I hit run, let's see if yeah here we go. 06:11.480 --> 06:13.430 So it's gotten better at this. 06:13.430 --> 06:15.770 So now we're going to hit it in again. 06:16.850 --> 06:18.350 And. 06:20.360 --> 06:20.990 Here we go. 06:21.020 --> 06:23.540 We're going to get rid of this extra text. 06:24.080 --> 06:25.190 We just want the joke. 06:25.190 --> 06:25.340 check. 06:25.370 --> 06:27.140 Just give us the check, please. 06:27.320 --> 06:31.340 And this is just a great way to iterate on prompts. 06:31.520 --> 06:34.460 When you hit generate again social media. 06:34.490 --> 06:35.840 Okay, this is a good idea. 06:36.020 --> 06:38.690 Hit run okay. 06:38.720 --> 06:39.830 And you can keep going. 06:39.860 --> 06:44.810 You can keep adding you know you can keep adding things to this and just improving the performance over 06:44.810 --> 06:45.920 time right. 06:46.400 --> 06:51.860 So now we have a three shot example prompt and say this. 06:52.250 --> 06:53.450 Tell a joke about dogs. 06:53.480 --> 06:54.980 Let's see if it finally gets there. 06:56.900 --> 06:58.700 All right so I missed something up here. 06:59.870 --> 07:03.020 But yeah I think it's you know it's a pretty cool tool. 07:03.020 --> 07:09.470 You can see you can click here run remaining by the way and see how well it works for these other things. 07:09.560 --> 07:11.720 But yeah like I find it quite useful. 07:11.750 --> 07:17.030 Obviously you can only test the anthropic results, right. 07:17.060 --> 07:17.360 Like so. 07:17.360 --> 07:19.670 You can't test the like any other models. 07:19.700 --> 07:19.910 Right. 07:19.940 --> 07:21.110 Like OpenAI or whatever. 07:21.140 --> 07:23.210 Which is a shame because I really like this format. 07:23.210 --> 07:25.100 And I wish OpenAI had this as well. 07:25.100 --> 07:28.310 But I find this really helpful for quickly ideating on things. 07:28.340 --> 07:29.930 Hopefully you will too.