Transcript
0:00 Hey, how's it going guys? Jacob Tiner here from Sculpted. Today I'm testing out Clay's new sculptor feature, which is an AI agent within the Clay platform that can actually build clay tables for you using natural language, or so they say. I've tried out Sculptor before, and I've never tried something crazy as this, where I'm going to try and have it build a full clay table end to end live with you over my shoulder. I'm going to try and build this clay table in under 3 minutes, but we'll see how it goes. I've described what I'm looking for with Loom. I took the transcript of that Loom video and put it into this GPT prompt and told it basically to make a sculptor prompt that I can then get to Clay and see how it does. So, I'm gonna walk through with you the criteria for this table and we'll see how well the output is on the other side. Essentially, um, of course, I'll leave if you want to see this prompt, you can have your access to it. Um, but essentially I'm telling Clay that we want to have an import section where we can import company name, domain, record ID and then uh use the domain to enrich a company with LinkedIn company data and output company LinkedIn URL employee account and then we also want to get revenue range with a revenue waterfall enrichment as well as marketing employee headcount. I want to use collagent with 40 Mini to try and find the uh company's website and then using OpenAI 40 Mini. I'm actually gonna update this here and I'll add any other uh update later on. And then we want the JSON structure here. Street, city, state, county, zip, country. Uh we also want United States in full, not US, but we want the states to be abbreviated. We also want to create company summary generation uh with collision. So just output a summary paragraph here of what they do. And then we have this industry classification GPT40 prompt which will take the company summary as the input and output all of this. So ideally Clay could build this end to end. I'm going to just say output the prompt again with everything the same except uh let the AI know it should use 40 mini for agent and 40 for its uh GPT columns. Okay. Okay. So I have the whole prompt here. I'm going to go ahead and grab this and I'm going to paste it in my uh clay sculptor build here. And we're going to go ahead and hit enter and we're going to see what happens next. So, starting the clock. It is 7:38. I'll say timer. So, I'm starting the timer or stopwatch, actually. So, it's we'll just I gave it a few seconds.
2:14 It wants a CSV with these uh inputs. I actually had the um AI output a CSV that I can use for this. So, I'm going to go ahead and download that real quick. We're going to import that into the clay table. Here we go. So, we have the source data here. I'm going to go ahead and attach the CSV. Boom. Attach. And boom. So we have three examples. We have Mount Si Hospital, Black Rockck, and Walmart. So we're holding at 33 seconds right now on the clock.
2:44 Okay, so let's see. Um, here's a workflow plan for their table. Enrich each company using domain to get company, employee count, revenue range, marketing employee. It's good. Use client form to research only the company's website and extract the official address outputting it in the JSON structure. Okay, that's all good. but it skipped the part about company summary and industry. So, not stoked about that. I'm going to try and say yes, you got everything right so far, but forgot the industry AI prompt as well. Please try again with all of your current um enrichments plus the clagent company summary plus GPT industry classification combin combo. Let's see how this works.
3:34 I tried this before and it actually just output only this step and not this step. So um okay, this looks better. Okay, this looks good. Let's generate the table. We're holding at 1 minute and 50 seconds. So, we have our input data here and we are going ahead and building all of the columns. While this is going, I'm going to go ahead and talk about my thoughts on this feature until it's done. I might cut off in the middle.
3:53 Essentially, I think this is nuts and I've been trying to think of ways to speed up my build time on in the Clay platform because I spend so much time clicking around and doing the same things over and over again. And I have so much contact from my clients, sales calls, onboarding calls. If I could combine that context with a best practices document along with uh a sync of their serum properties, I probably could automate a lot of my clayt builds along with clayt templates. All of this combined together, I think could take a one hour call with a customer and turn it into a fully enriched clay table workflow versus 20 to 30 hours of work per project right now. So that would really really make things a lot faster and also open up clay too. I don't know if it would make clay accessible to everyone because you still have to know what you're doing in order to prompt this properly. As you can see here, I think most people would not be able to make such a good prompt and that takes a lot of experience and and skills. Um, but for someone who is good at clay already like me and has that background being very humble, I know and has that background and go to market then it's going to speed you up a lot and make you much more powerful. Although I think this first version probably still needs some some work to smooth it out. But I am very excited about sculptor. So let's see what happens. We're still uh building here. If I go through this flow. Oh, looks like we have what we have. What we need. So, we have company LinkedIn URL. We have employee account, revenue range, and marketing employee account. Let's see what happens. So, let's add this one. And okay, that's going to go ahead and enrich with the information that we're looking for here.
5:22 So, let's see what happens. It didn't output the the columns that I wanted. Um, so we're only going to use sculptor for this, though. So, I'm going to hit revenue range and see what happens here. Okay. Um, this waterfall I feel like I would not have used clear bit myself or crunch base. It's way too much credits. So, I'll configure this. Um, but that's it. So, the rules are that I can only use sculptor um natural language prompting here to try and do this as well as maybe one or two quick toggles over here on the right side. So, next we're going to go ahead and click company LinkedIn URL is already added, I guess, and then revenue range is already added. I don't really know what happened over here. So, what are we missing? Um, okay. Can you please add the collagent step to output the address in JSON format with street, city, state, uh, county, zip, and country formatted with full country name and abbreviated two-letter state codes like CA, MD, etc.
6:34 Okay, let's see what happens here. We're now holding at 4 minutes and 58 seconds. So about five minutes to build this. And it looks like we are hitting a bug here. I try this again. Oh, nothing happened. Can I refresh? See what's going on here.
7:08 Okay. I don't know if this is user error, but um I'm not seeing any any Oh, actually that was confusing, but I it looks like it's starting to build my next stop, my next step here. So, let's see what happens. Lovely clay. Always uh bugging. This is kind of part of the the clay journey here. Molding the AI column. Okay, that's promising. Okay, let's see.
7:38 Okay, let's set up the column and we'll run this and see what happens. Okay, now please build the Caggent company summary and industry classifier uh prompt. Let's see if it knows what I'm talking about. Okay, let's see how the address looks. We have street, city. Okay, we have everything we asked for here. Um, state. The county is empty. Not sure why that is.
8:10 Um, and country. I bet county is empty because we're not inferring the county and the address on the site does not have the county. So, it's empty. Whereas, if we had told the AI to infer it, maybe it would do it. Now, let's add the company summary step. It did not include it did use form mini, which is good. uh and it did not include any it did not include the GPT prompt as well to actually build the output. So I'm going to let this run. I'm also going to say please redo the address prompt to infer the county. Actually, this might get wrong. This might get it wrong. I'm going to leave county out of this for now and we'll just leave it blank because it might hallucinate and we'll explain more about that later.
9:07 Okay, we have the summary here and now please build the GPT prompt. Now, please build the GPT prompt to classify the company using its summary into one of the industries that I gave you or other. Okay, let's see what it does now. Please provide the list of industries. It does not remember what I gave it. Okay, we're going to try again.
9:38 I'm going to go back to the AI prompt that I I generated and I'm going to paste it in again and we'll see how this does. This should be good enough to give it enough context to do a good job here. It's using 40. Okay, very good. We're now holding at 9 minutes though, which is a bit long. Um, unfortunately, okay, let's run this and see how it looks.
10:27 Once this is done, I'm going to uh call this a a day. Okay, so industry pretty easy. Healthcare, finance, and retail. That looks good to me. This is pretty cool. In about 10 under 10 minutes, I could build a clay table that had all of these different properties. We have uh company name, we have the website, we have the uh company LinkedIn URL, we have the employee account on LinkedIn on LinkedIn.
10:57 We also have their revenue range, their address, and a company summary along with their industry classification. Now, a few things that we left out here and a few things I wish we could we could add. First of all, there's no conditional formulas in any of these columns, so they would all still have to be created individually. It could be that we could build them with the sculptor, although I'm not sure how well that would work. We could test it out later. Another thing I would want to say is that even though I gave it all of the instructions in one big prompt, it didn't output everything all at once. It only output partial columns, and I had to then go and reprompt it again and again a few times. So, what could have taken maybe 3 minutes took about 10 minutes. Now, going from 10 down to three, it is a about a 3x time savings. Now, it's only seven minutes difference. So, in the grand scheme of things, 10 minutes already is pretty darn good to build a click table with natural language, which is pretty cool. But imagine doing this all in a few minutes. That would be unbelievable and it would take a lot of the effort away. I think that will come in a future iteration. But for now, this is pretty decent. I think the real question is, did we actually save time here? I think for someone who's brand new to clay and doesn't know what they're doing, using natural language would probably help them, although I think it would still be a bit confusing.
12:01 And then the question is, would someone like me who is very fast in clay itself, would I actually benefit from sculptor versus doing this manually? So, in a follow-up video, I'm going to try and rebuild this exact same clay table using only my hands, so to speak, and no sculptor whatsoever. And I'll time myself and see if I can beat Sculptor.
Summary
- The sculptor feature allows users to build clay tables by inputting natural language prompts.
- Jacob outlines specific criteria for the table, including data enrichment from LinkedIn and JSON formatting requirements.
- Initial attempts to generate the table faced issues, such as missing components and bugs in the output.
- Despite the challenges, the sculptor successfully produced a table with company data, employee counts, revenue ranges, and industry classifications.
- Jacob notes that while the process took longer than expected, it still represents a significant time-saving compared to manual methods.
- He expresses excitement about the tool's potential to automate and streamline workflows, especially for experienced users.
- Future iterations of the sculptor may improve its efficiency and output accuracy.
- Jacob plans to compare the sculptor's performance against his manual table-building speed in a follow-up video.