Video: SupERPower Hour Session 1 — Prompting & Vibe Coding for Finance | Duration: 3528s | Summary: SupERPower Hour Session 1 — Prompting & Vibe Coding for Finance | Chapters: Welcome and Introduction (5.866s), Super Powered Hat (217.31099999999998s), Getting Started with AI (264.046s), AI Tools Overview (493.636s), Setting Up Claude (690.4010000000001s), Continuous AI Interaction (974.721s), Financial Planning Process (1307.221s), Building AI Skills (1841.271s), Managing File Access (2388.466s), Feedback and Iteration (2515.616s), Security and Compliance (2714.031s), AI Integration Techniques (2904.626s), Q&A and Wrap-up (3165.681s)
Transcript for "SupERPower Hour Session 1 — Prompting & Vibe Coding for Finance":
Morning, everyone. Welcome. I'm gonna take a few minutes until everybody is logging in. My name is Irina Silva. I'm with the marketing team here at Campfire. We're coming to you live from Downtown San Francisco. I'm going give a few more minutes. Hi, everyone. Maybe you can tell us where you're logging in from in the chat. We'd love to know where everybody's logging in from. Yes. There will be a recording, and we will share that with everyone here. Also, this is a four part series, so if your other teammates could not come today to this session one, they're welcome to still register. There's going to be three more that they can attend later. Great. Well, welcome to Superpower Hour brought to you by Campfire. We are very excited to bring this to you live. This is a direct response from the community. So many of you have come to John and to other teammates here at Campfire asking us how to vibe code, how to automate your finance tasks with AI. So we're, put together this four part series, to help you with that and bring your, take your manual automation to the next level. So, with that said, we're starting today with session one. We're going to have our CEO, John Glasgow, lead this session one. We're going to basically kinda set you up and get give you the foundations to for you to learn how to effectively write your prompts and get the useful outputs that you need on your day to day work. So that's today's session one. Then next Friday, also at 10:10AM Pacific, 1PM eastern, we'll have session two. So we'll there, we'll cover app building with using Replit. It will also be a hands on session. And there, John will walk you through, how to how to develop how to go from idea, to cut development. Then, third week on April 10, we're going to have Hank Sun, also from Campfire, walking you through two tools to make AI repeatable and more practical. And then session four, we'll wrap up. We'll bring back John, and we'll have one of our partners, Nigel Glenday, for from Masterworks, where they're going to do a live demo where you we're going to bring Campfire, MCP, and connect that to clot code. So that will wrap up the four part series. Please, we hope you can join all four of them. If you can't, there will be recordings, and we will share that widely with the community. We're so happy that you're here with us today. And because of that, we're actually going to be mailing you a super powered hat. So if you'd like to receive one, please take a moment now to scan the QR code that you see on your screen. Make sure to add in all your details, and we'll make sure to mail you one in the next week or so. So take a moment. If you'd like one of these super cool hats that we just had made, give us your details and we'll receive one in the mail. And so now I'm going to pass the stage to John Glasgow, our CEO. He's going to give an introduction, and then we're going to proceed with a live demo. John? Awesome. Thank you. And really excited to be here today. Thanks all for joining. How this session came about is, talk I had a ten year corporate finance career and, obviously, spent a lot of time in Microsoft Excel doing everything manually, a lot of editing, charts and graphs and, you know, three statement models and getting it to tie out and VLOOKUPs. And, once Claude Kowork came out and, obviously, with with the advancements in AI, started to really dive in and think about how do I take all my Excel skills. Yeah. Yes. I miss all my keyboard commands, but how do I really apply that to accelerating the workflows and less time preparing the analysis and more time executing on it. And so, I've been on my own journey with all of you. And, as Irina stated, the first session for today is really gonna be we're just gonna get started from from day one. So if you're a power user, you know, you're probably gonna be more excited about the later sessions in the series. But, today, we're gonna get everything everything started, talk through how I got going, with my AI journey as a corporate finance leader, and just going to look at the q and a, and we're gonna make this very interactive. If you're into gaming, it's like Twitch where you can kinda, like, watch and chat and, you know, we can work through tasks together. But, this is, like, just a just a working session together. So, first thing we're gonna do, we're gonna start super, super basic because I was talking to someone just yesterday, and they were like, once you told me that, that was actually a a game changer. But I I I think before that, what's an MCP? We're gonna get super, super basic, and then we're gonna hop into actually some tasks together. So an MCP model context protocol, this was developed by Anthropic, but it's really become a unified connector. It's essentially the API layer for AI, and it allows you to have, the AI from really any system, whether it's your spend management system, whether it's your ERP, like Campfire, or whether it's, you know, Google Drive, having Claude go in there and prompt the data and then pull things out. So as you can see on the slide here, you're gonna have your system. You're gonna have the MC connector, MCP connector to the AI model, and then the outputs. And this is a way of pulling in all sorts of data sources, and we'll actually go into, doing this together in just a second. Okay. So first thing I recommend doing and super basic for folks, but, again, we're gonna start at the beginning, is if you don't and there's there's a lot of good systems out there. There's Gemini. There's OpenAI. Personally, I'm on Claude. And so and we have an enterprise license for Claude at Campfire. So there is the web version. And I think one thing that a lot of folks have said, hey. I was using the web version, but you really wanna get into desktop. And the main reason for that is desktop, has Cowork and and Code as well. But today, we're gonna be spending time in Cowork. So be be sure to download, Claude desktop. And, again, if you haven't done that, go for it. Or in your relevant system, like Gemini or, or OpenAI or open source models as well. If you're in a work environment, you may need some permissions from the IT department. But at this point, we're gonna assume, you all have Cloud Desktop. Okay. So I am now in Claude desktop, and we have chat. I think we're all pretty familiar with chat at this point, but this is the one that, you know, is was what really took off in the 2022. Then Claude Code came out, and this is one that, I was using Claude Code. But, you know, I think as a nontechnical person, I know a lot of finance folks that are in Claude Code. I think it's totally fine to use. What they did with Cloud CodeWork is it's actually sitting on top of Cloud Code, but, they abstracted away some of the tasks you might need to do, like going into the terminal. And so going in here, as you can see on my screen and, like, performing tasks or going into, like, a a code editing tool, there was a there were just moments of time where, yeah, just very candidly, I was getting a little stuck and spent some time trying to get through it. You can see I'm even here, like, debugging and resolving an error in my Claude code days. But when when Cowork came out, and you can see it kinda walks us through the the setup here, that was able to really allow me to accelerate my work with AI. And and so we're going to give me one second here. Okay. Alright. So, the first thing we're gonna do is a lot of you put, questions ahead of time. So we're actually just gonna go ahead and run a quick analysis with questions. And I use Cowork for all sorts of work tasks where chat is gonna be for things that just have a quick question. Alright. So let me drag something in. Oh, I'm looks like I'm sharing the wrong screen. Sorry about that. Okay. Sorry, everybody. We're now in in in in Claude Cowork. So I think the only task we missed is I went on the Internet, and I went to Claude desktop, and I showed how to download it. So let's hop in here. Great. We got some good topics. People wanna talk about reconciliations, month end workflows, BBA, AR, you know, how to do slides. I think this is definitely one we're gonna wanna cover. A lot of folks wanna talk about sales commissions. We're gonna cover, in a VIBE coding session later on, how to build your own sales commission app. So but this is just like taking really any data file and anything that's, a lot of data, then, I will just throw it in here and ask it to to summarize data for me. So that's just a quick one. If you're not familiar with what this is here, these are the different models that come with it. I generally always stay in Opus four six. That's the latest model. It does use the most tokens. So depending on the plan you're on, you might wanna be careful. It's like I have hit the token limits for the day, particularly if I'm doing a lot of, like, deeper work. The the Sonnet and Haikyu are gonna be for maybe you're doing, like, a lot of simple tasks. Like, you're having it do invoice reconciliation for you, then you might wanna be in one of the lighter ones if it's a very large dataset, and that'll preserve tokens. But particularly for, like, building board slides or strategy work, always gonna be in in four six. Another thing we're gonna do is we're gonna wanna set up, an MCP. So as we just discussed. So I'll go in and customize, and we have these connectors here. And we can go into the connectors. You can see some things I've got set up. Let me just we will browse. Well, I'll just hang on here. And so you can browse the Claude connector store. So, like, when I set up Campfire, literally just typed Campfire. It was right there. If you're in ramp, I would go ahead and just do this with me. Set up a connector if you haven't already. But ramp ramp and Brex are easy ones because you can start analyzing spend data, or, you know, just, there's, like, a HubSpot connector. But there's you know, you can see there's quite a few out of the box. Now let's say I I I don't see a connector. So maybe I'm in Salesforce, and I'm not seeing Salesforce. What I typically do next is then I look if there's a custom connector for a system. And so let's just go through that together. And I'll often ask, Claude if there is a connector for a system. And then if you wanna go do a I'll just pull this up here. If if you wanna do a custom connector, it's gonna be a little more work, but you're gonna be able to go set it up with through we won't do that today, but it's there's a couple steps you've gotta take by, like, dropping in a URL. Typically, you have to go into the API docs. Like, I'll show you. Yeah. We'll we'll skip that one today. I think it's gonna be a little too much. But not everything is in the Cloud App Store, and that doesn't mean you can't set it up. So I just wanna, like, throw that out there. It's obviously much easier if it's in the Cloud App Store. And then the other one is skills. And so we're going to, build a skill today. We'll get to that one in just a second. Scheduled, if you're not already scheduling things, you can go ahead and create, like, recurring tasks. Little more, not quite OpenClaw if you're familiar where it's just kinda like doing tons of tasks for you, but maybe every day at 8AM, you wanna get something. But there's, the other thing I wanna note is we want to set up a plug in. And so there's plug ins, and we're gonna browse. And one that I recommend is there is a finance plug in. There is no accounting plugin, but there is a finance one. So if you're in cloud with me today, go ahead and just add the the plus button and set it up, and we can also manage it. And so here's a few skills that it can perform. It can take on some tasks for us. It can prep a journal entry, prepare a month end close. So there's obviously a lot of other things that Claude can do, but, essentially, it prepares it for you, and then you can see you can invoke them. But it's like a called a well structured prompt that's set up with some hotkeys. Think going back to the Excel comment. We've got some hotkeys here generating financial statements, reconciling accounts. K. So the first thing we're gonna do is we're gonna build some slides together. Actually, even before that, we're just gonna ask a quick prompt to make sure we connected ramp. So can you And you can just ask questions. And while it's going, I'll often actually, just keep prompting it. In ramp. So you don't need to get the first prompt right. And I I think one of the biggest learnings for me has actually been the first prompt is essentially never right for me. And so it's like a new treat AI like a new member of your team. It's gonna get it wrong the first time, or it's not gonna do things great. And you just keep prompting it and giving it more feedback. So looks like someone spent $18 on Zoom. Great. Then, now let's take a look at all spend for And we're just gonna go ahead and keep going. I often have it build slides for me, and often, I'm the only person that's ingesting the slides. You can have it do a couple different formats of analysis. Claude really likes to build, like, lightweight software apps to show you visualizations. And I it's building a PowerPoint. Can you build in Google Drive instead? And what it's gonna do is as it's working, it's gonna have the the tasks and what it's connected to, in the side here. And you can always check on what it's doing. You can do multiple things at once. So while this one's working, we can kick off a new task. And one thing I'd I'd done in this account is a project. So if we wanna create a project for, like, a larger exercise and, I created one from scratch. And then what I did was I called it, like, 2026 financial plan. I saw, like, cash flow forecasting and planning was, like, number one on today's list. And so, we'll go ahead and work through that. And what it does is it creates a new folder, and this is on your desktop, and we can add files. So what I did was I just dragged a couple files in. So I'm just gonna drag them in here. And then these will essentially always be referenced when I'm working in this project folder. Memory is turned on. And as it says here on the screen, that just means it's literally gonna the learn as I go. So if I give it feedback, like, hey. Actually, here's how I really like Excel formatting, then going forward, all Excel files should have that. It just sounds like a lightweight skill, but it's just more of, like, little context that's passed to it along the way. Again, just like anyone else on your team, once the team learns how you like your Excel formatting, then, you know, hopefully, they're gonna continue to deliver it in that format. And and that's really, how memory works. So let's take a look here. Now it's it's asking me. So I was talking to someone this week, and they're like, my team has told me they don't have time to learn AI. So I am often, working in Claude on one monitor, and then I'm often doing a non Claude task in another monitor, because it doesn't need to necessarily slow you down. There is a lot of waiting when you're when you're actually, building things with AI, but then there's a lot of steps where it needs your help. And so I think having one screen up, not just waiting for it to get to the conclusion. So I use Chrome. Build there. Is is has made it for me where, like, there's really no excuse to not have time to learn AI. We're near we're largely waiting for it for, quite a bit of the time. Okay. So while that's one working while that's working, we built out a project plan, and we now have, some files in here. And you can actually, see it's in there. And so what we can do, can you take our, 20 okay. We'll give it a second. It's asking it's on the it's off screen, but it's it's asking me to, like, reauthorize, Claude. But, ultimately k. You've done give it a second. I think that somehow I got kicked out, so I'll give it a second to to build the slides. There we go. Alright. So it's building. We're gonna hop back over, And I will bring this over. It is now kind of building in this other tab, that's off screen right now. And for the 2026 financial plan, I I actually just had Claude make, 2025, income statement, and then, I had to just make a 2026 headcount plan. And then what we're gonna do is use those as some initial inputs for a 2026 plan. I often like to start with a simpler prompt for for Claude and then look at, like, where it has gaps or where it's, like, over assuming things and then really give it a lot of feedback. Think then for the for the next time, we've got a much better structured prompt. Hopefully, the memory feature is kinda picked up some of the the quirks, so we can get better for the next one. And so head count, plan, and build. What questions do you thankfully, you don't need to be the perfect speller, but, I like to ask it what questions it has. It typically will often ask you, but this is a way of having it not, make as many assumptions. And you can see here because we started that project with the two files, it is referencing them, and it's now gonna walk us through the building of our 2026 plan. And when you see blue, then that means it's done. So you can do multiple things at the same time. So they're both blue, so they're both waiting on me. And so we're just gonna put in, well, two x for the year, and you can give it specific targets. So how do we wanna think about nonheadcount operating expenses? Say as we grow, there's gonna be a little more overhead. You assume overhead will grow. Right? And we want a full p and l, but and a and a spreadsheet, so we'll just do that. And I often like to start with one type of output, like an Excel file. And then because I'm gonna iterate on it quite a bit, and then I'll have it build slides as, like, a phase two just like you would as a as a more of, like, a human or manually performing the task. So any changes? It did notice the 2025 plan has some an evolution, and so let's just do we'll compress slightly, and we will check on the other one here. Let's just build a PowerPoint. It's often gonna ask you, questions for approval. And one thing I do wanna call out, there's been a lot of horror stories of when Claude has write access or if you've used, like, OpenClaw, when they have write access, when they delete things on your computer, they actually do it, as noted earlier, what's called the terminal. And it can be incredibly hard to recover things that were, deleted via terminal. It's not the normal, like, moving to a trash that we're all used to. So when you're configuring your account and you're setting up your connectors, like, the Campfire one here, you probably wanna make sure it needs approval. And when it's writing to like, if it's ever gonna write, like, on them, make sure that you've got I'll just kinda see if this one has yeah. On right here, you either wanna block if you don't want it writing back to a system. So we probably just wanna move it over to block. And then it's gonna be in, like, a read only state, and then I can't, like, accidentally, like, screw up our our finance and accounting software system. So if it, like, deleted everything in Ramp, obviously, that would be quite painful. And while this one's going, I might just show you the the output here. I'm gonna take a look at questions we have so far. I went to install Clogged Cowork previously, and it asked for full access to my computer, which made me nervous. Is there a way to set this up or it won't go rogue on your computer? That's a great question. I don't recall exactly what I did on that step. I would actually just be on the browser version of Claude Chat because that's you know, I assume that one you're comfortable with since it doesn't have full computer access. And I literally just ask Claude how to set up Claude Cowork on desktop where it doesn't have full access to your computer. I think one of the best ways to to work with Claude is literally to ask it. So even when I was doing projects for the first time, can you show me, what's a good what's the best way to set up my 2026 financial plan with your projects feature? Or is there a kinda weekly, summary of our cash flow, roll forward on a thirteen week cash flow that you can send me? Like, what would be the best way to do that with you? And it it it will do a great job at at walking you through the task. Someone said, I tend to run out of tokens. How do I avoid doing that? That was the comment earlier about choosing your model. Of course, upgrading cloud plans is their preferred method where you pay more. If you're on an enterprise plan, you know, play at work, you know, they're they're giving you enough tokens. But it's a it's a great question. This is quickly becoming a thing. I would say the other one is the the method of reprompting it consistently. Like, we're about to do a task where we're gonna build a closed checklist, And I had it add a column to the closed checklist, in Excel, and it actually said, hey. I can't add a column to the middle. I need to rebuild the entire closed checklist. So trying to get of it give it as much structure as possible if you are worried about tokens. I think I generally run a pretty high plan at Campfire, so I am a pretty heavy token burner. And and so I think as I kinda get more efficient and I learn how it works, I'm certainly gonna, you know, hope to to use less tokens. But that's, that's just part of the art of learning. And if you need to ask the admin for more tokens, whoever runs your cloud account, if you don't wanna ask them, you you you some people do run multiple you know, you can go have a Gemini. You can have an an OpenAI. You, you know, you can get multiple going. There's obviously downsides if it doesn't have the same memory, or there's there's reasons why that's not always great. Someone correctly pointed out the practical role of the data would need to appear in an audit trail. Don't run it through Cowork yet. It's a great callout. Yeah. I think within Campfire, that's one of the reasons why a lot of the folks are using the AI within Campfire because their auditor is quite comfortable because it lives within the ERP. Campfire can Ramp can handle this on their end. Would Campfire's version of the data look any different from Ramp's own reports? We will go through that. That's a great that's a great question. Okay. So we can see here, it pulled together file for us, And here's our first financial forecast. So it took our again, it took our 2026 head count plan and our 2025 income statement, and it took a lot of assumptions. So it prompted me for most of these, not all of these. So, like, I don't remember talking about the the effective tax rate. And so we're gonna you know? And, of course, they're blue, meaning, you know, they're they're editable for the model. We, you know, we gave it some assumptions on how we're gonna start and end the year. If you're doing top down planning, then, of course, that's great. I did it simplified. If you're doing bottoms up like a sales led company, then you really just wanna tell Claude, like, hey. We're bottoms up planning for revenue. We're going to end the year with 10 AEs, and quota is I think we had said 10,000,000, so we're gonna say quota is year. Assume they are fully ramped. Can you build the revenue forecast this way instead? Okay. I'm sorry for my spelling today, but we're gonna have it just rebuilt for a sales led business. Alright. Let's take a look over here. And I'll try and monitor questions in the chat. Why wouldn't you use ClaudeCode in Versus Code? It's a great question. I have, and you might see here it's on one of my machines. I do have Versus Code, and I've I've tried Cloud Code in there. I would say this is, like, a non engineer. I think just working out of CoWork, especially because it sits on top of Cloud Code, has been just, like, by far the easiest way for me to just, like, never get stuck. Alright. Let's take a look at what we shipped here. So this is, built a slide deck that is pulling out of the MCP. And as we talked about earlier, that just a a way for AI to get data out of a specific system. Looks like this person went on a trip or two, some rideshare, top merchants. Now this one, you know, we we got some issues on this one, and this is part of working with with AI. You you just can't give up when it struggles like this. We just need to power through. So I think a lot of folks say, AI is not ready for me. Look at slide four. I hear that often, as dramatic as it may sound. And so what you wanna do is we're just gonna give it some feedback. Also, Campfire AI, you know, our our formatting, is in blue. So let's go ahead and give it give it some feedback here. Pull this off. I am going to add a file. Let me just dig around here. I I often just grab any deck, and I'll show you how to build a skill in a second on it. Can you take this deck formatting and build a Claude skill so we always use. Okay. So we're gonna do two things at once. We're gonna give it some feedback while we build a skill. And if you remember earlier, one was the finance skill that, it was able to kind of we talked through, like, some backslash. Like, I think one was, like, close. So here's, like, the finance skills that you can kick off. But, ultimately, we're gonna we're gonna have it update, this with our formatting, and we're gonna have it build a skill. So going forward, it'll always gonna have this formatting context. Now we're gonna check-in on our financial model. Looks like it's still chugging along. And while both of these are running, I'm going to go ahead and check on the q and a. And by the way, I have, our CRM connected, our our spend management, Google Drive. I've got Campfire connected. And then what I often do is I'm having it, work off out of all the systems at once. So, literally, for we had a a board meeting last week. I took the prior board deck, and I said, can you go into our CRM? Can you go into our ERP? And can you actually refresh all of the data? The board meeting, is, is in a week, and we need the data as of x date. And and then it went in there, and it refreshed all of the data, in the prior board deck. And then can you refresh commentary? And but I'm often, like, going through and and just hitting the MCP for for a lot of tasks these days as opposed to going into the respective systems. Okay. Let's go through some q and a. And I was told the PowerPoint didn't share. Let me just reshare here. Can we see the the the PowerPoint? Let's do entire screen. Okay. Here we go. So, sorry about that. And so all that all that you missed earlier was, like, here was the deck that it built. Formatting was wrong, and, this slide four, is definitely not what we want. So you can see it's drafting a skill for us for future use. Alright. Let's go through a little more q and a. So many of my manual tasks involve me logging into online tools and exporting CSV files, my bank websites. If there's a connector for these things, is it too soon to use co work? I would say no. It is a little frustrating to be dropping things in and out of of Claude. And, so I'm just gonna, like you're just gonna have to get by. People are adding MCPs quite often. And, again, just because it's not in the App Store for Claude, then it doesn't mean it doesn't exist. Ask Claude to browse the web and see if that specific app has has an MCP, and then you can use the manual connector that we noted earlier. How do we build a feedback loop so that the n plus one nth try will be better than than the prior one? If you have memory turned on, on Claude and if you add skills, those are the two ways to improve the feedback loop. So that's why we're, building a skill together here so, so it'll learn, which is a great question. Is there a co work setting where you can put global instructions stating to never delete a file without your permission? I try to be extra safe. I created a co work folder on my desktop where I'd make a copy of that file that I needed and threw it in the folder. So Claude only has access in there. You can yeah. So it's a great question. You can ask Claude to only stay within a certain folder. I think the only challenge is that there will be times where you're gonna work with Claude, and it needs to go outside of that folder. And so then if it goes outside the folder, particularly if, like, you're you're training it like, this project here, it's only gonna work within its own its its own folder here on the desktop called twenty twenty six financial plan. But if I'm doing other tasks and it goes outside, there is a risk it'll accidentally delete something else. This is why some people, more for the open source ones that have less security, they end up getting, like, a dedicated instance. They're getting, like, a a a Mac mini, and they're they're doing some of the, more, call it, innovative tasks or pushing boundaries on, like, a dedicated instance. But, as I noted earlier, there is a bit of a risk that you might end up accidentally deleting something if you're not careful. So please, please be careful. Okay. So we've, and I'm I'm apologies for folks that it wasn't fully, sharing earlier. I'm now kinda double checking everything to make sure it's correctly sharing. But, ultimately, I I can see that we can all see it, that we've got an income statement now. We've got our key metrics, our bridge. Just move this over. And on the assumption side, it's now, been fully updated. We got a headcount forecast, new hires for the year. And so this is just a great place to start. I think we're gonna move on to the next task. But iterating with the financial model, whether it's in, Google Sheets or in Excel, is something that, I used to always start a blank workbook. And, you know, and now that's that's no longer the case. Okay. We're gonna go ahead and check on this one here. Six slides. Pull it up. And it didn't use our new formatting. And this is something that I've seen, and sometimes will get a little stuck. So sure to use the slide template that I shared with the green. And I'm just gonna ask it oh, we didn't scroll. I'm sorry. Let's do yeah. So we just created a skill. And on the new skill here, it grabbed what's called, like, our RGBs. It has the correct and we can always view the skill, and we can always give it feedback, on the skill or make edits. But, ultimately, this is, now a new skill on our account. And I often do look at, like like, how Claude came to conclusions. I think this one's a little more technical, but something worth noting that, like, particularly when you're debugging, like, when it's editing a file, is it editing, like, the right things? What are we working through? So this is, you know, worth worth doing as well. Okay. So I can just share it with the green is used for the slide presentation. Also, slide four did not have vendor names. And so one thing I'm typically doing is, as I am reviewing a file like, we'll just pull this one up again. And I'm just gonna move this one off the screen, for a second, but and so you can you can follow along on the the Claude Cowork side. But as I'm reviewing a file, I just keep giving it feedback. So we're gonna go ahead and and you can see, by the way, this time, now correctly has all the all the data on slide four. It's got the merchant data, transaction detail. So it's starting to look better. We're just giving it some feedback on the the slide formatting, but maybe we actually don't wanna use a pie chart anymore. We're just gonna continue to give it that that level of feedback. And then when I'm done with something, it's like, okay. Can we now ask you know, we're gonna write a Slack message to this employee asking them to submit receipts as we noticed there weren't receipts. Or Claude noticed that, like, we're over we're we're outside of policy for certain, like, spend categories, and so we're gonna have them write, like, a Slack message for me is kinda how I then think about, like, taking the the, the exercise all the way through. It's then, like, what's the final deliverable? And then we'll but then now Claude has full context on what exactly it should be doing. When connecting to Salesforce and other data sources, does Cloud ever edit or do anything to those entries or data? That was my comment earlier about you probably really wanna be careful. Like, I've never had a need to have Claude write back to, like, Ramp, for example, or spend management tools. And so that's why earlier we moved it to to read only because then to the question about Salesforce data, if you really wanna be careful, then, just put it in a in a read only mode. Is there a way to record a video of a task via Loom or similar that a finance analyst performs and ask Claude to repeat it? I've never dropped looms into Claude and asked it. I guess question for the group if you've dropped looms into Claude. But there I know there are some tools that are specifically built for doing things like that where it'll ingest a video recording, and it'll create, like like, essentially, a a list of how to perform the action, but I I candidly have not tried that with Claude. Who's liable for the MTP security layer in terms of SOC compliance? It's a great question. I do know that that Claude is is SOC compliant. Obviously, you need to go through your own security reviews and and what's your security posture. I think, as noted, some security departments are not comfortable with Claude at all. Cowork, some will deem Cowork, I have heard, to be not secure. An example is, like, no customer data at Campfire is put into to Claude in any way. It's it's purely used on internal data, like All Hands materials. And so you you yeah. I would just recommend coming up with your own posture on that. Give me a second. Okay. So we can see it came back with our formatting. This looks a lot better. This is the campfire green. We've got our nude skills. So going forward, it's gonna correctly use the campfire green on everything that we do. And, it's well it's nicely, cleanly formatted, with our font, with, now one thing you may notice is the deck I gave it had the campfire logo, but the Campfire logo is not in here. And you just like earlier, if you're trying to troubleshoot, Claude, like, how do I set you up without having, full system access? Why did you skip the logo on the slide? And I know the answer, but we this is just one way of learning how to work with Claude. While we're going, other questions. Could you comment on the Ember AI in the context of Claude's ability to connect to Campfire and pull numbers? Yeah. So for those who aren't familiar, Ember is the AI layer within Campfire itself, and there's a lot of things that Ember can do, that is probably better to do than Claude. So, for example, if it's any sort of, like, transactional accounting work, if it's cash flow forecast, if it's, deeper analysis, I'll often go into Campfire for it. If it's building slides, if it's connecting, like, our CRM data with our ERP data, then, that's typically where I would go to to co work. I think one thing, actually, I've really enjoyed doing, the other day, I was doing, like, an audit workbook, like, by client workbook where you pull out all of your data. And I actually had Campfire and Claude each generate prepared by client workbook. And then I fed both back into Claude, and I had Claude audit the outputs of both and had built the rubric on any differences. And Claude's formatting was a little better, but Campfire's version was more accurate because it sits within the ERP. And so then Claude actually was able to figure out that it the the API was was paginated. So it had missed a few transactions, and it was able to actually self heal after seeing the Campfire version. And it went back and kinda took a second pass at going through paginated API calls. And that was a way I was able to to QA the prepared by client workbook for an audit. So I think those are that's generally one way that I'm everyone's like, well, what if it hallucinates? How I'm generally and I'll just kinda move this along here. So it's a good question. Well, essentially, the answer is, like, AI struggles with logos because anything that's trademarked, that it it often won't wanna use a logo. And so that's why it ends up, like, doing something a little different a lot of times, and it does struggle with, like, kind of fine tuned images. So I'll often just add the logo at the end, if it's struggling. Alright. Let's see. I know we have just under ten minutes. There was a lot we were gonna try and get through today. The only other one, and I'll make this one quick, is use skills to build the closed checklist. I'll just show you the output on this one here. And, again, it's just gonna prompt you through, or you can do custom. So we're gonna do an Excel, two people on the team, and it's gonna kick it off. Now you can see here it's building without any context about our company. So I you probably wanna give it just a lot more structure before it gets to this to the question earlier about, like, burning through tokens. This is one way to generate a ton of tokens when, like, it's building the wrong thing to start, and then you're doing, like, aggressive editing for it. I trained you more than than what we're doing here. And I'll just kinda quickly show you the the output here. Here was that closed checklist it built. And, initially, it did not come with charts and graphs, and then I ended up adding them, and I ended up adding the to the comment earlier I made at the beginning. It didn't have an approver column in there. And so when I went to add it, it actually ended up rebuilding the entire thing. And so this is why even if you're able to, like, here's the columns we wanna see or we're a team of five and we need approval workflows. Here are some tabs I like. The more structure can can help it take a better first pass from a token consumption standpoint. With the few minutes we have left, I'm gonna try and get through some more questions. You mentioned earlier that the initial output is often wrong when asking about values and sums. How do you reconcile this? That was the one and maybe, I covered it right when you asked. But that was how I'll often take two systems, and and I'll have them have it reconcile against two different systems. You definitely don't wanna use Claude for both. So that's why I'll go into like, you can go directly into your spend management tool and have it pull data out, and then you'll go into Claude and have it pull data out. If it's a a less critical task, like the employee one we just did where we're just kinda doing something, simple for an employee, I often won't have it. I'll just more of a spot check. But if it's, like, for the board or, like, that prepared by client list, I will have it tie out across, two different systems from a QA perspective. I see that you're on a Mac. Do you find Mac works better with cloth than PCs? I was a you know, I was literally the only employee at my last company with a PC PC for life. Actually, they sent me a Mac, I sent it back. And I said, this is a nonstarter. I've just I've just evolved. So I've not personally tried Claude on PCs. I don't know why it would be any worse. This is definitely not a why I'm on a Mac has nothing to do with with Claude. I would like to think that it should be just as strong, but, candidly, I I don't know. How should we think about should we think about some value if Claude already let us build and run accounting workflows? I think we're missing a word. I wasn't quite sure what the question was on that one. Okay. I've had Claude send me Zapier or make.com for certain agentic tasks. How should we think about what belongs with Claude versus Zapier or a make.com system? It's a great question. We'll cover more of the the the automation and the workflows and connectors later on. We do see more and more folks just connecting things directly into, like, a cloud environment and a little less Zapier these days or a little less, like, manually connecting systems. I think it just really depends of, like Zapier obviously has a lot more connectivity than cloud does. We saw the store, you know, that we're just in didn't even have Salesforce. So, Zapier is probably gonna capture the lion's share of things, particularly if you're using apps that are less common or more of a legacy piece of software. It's definitely not gonna have an MCP, but it will almost certainly be on Zapier. I had to remind Claude that they're connected to a system. Usually, that might be something you have on the security side if you have, like, a two factor authentication turned on. But I haven't and I guess as it relates to just Campfire specifically, I've never had to remind Claude that it's connected to Campfire or RAM. Occasionally, it will re authenticate, for security, but that's been the extent of it. I'm running a Windows Home 11. It'll not run co work. Had to get my Mac. So just a comment from Boris here. Potentially, the Windows version will not run Cloud Cowork. How do you think of last question, and then we'll kinda start to wrap up here, and I'll hand it back to Irina. How do think about Campfire's value if Claude already let us build and run accounting? It's definitely one of the comments made earlier that, you know, audit definitely, someone correctly pointed out, like, definitely don't, you know, doing things in an audit environment in Claude. But we're seeing the the accuracy of Campfire's AI to be much higher for accounting tasks than Claude's. The Campfire's agents can run asynchronously within your AI and continuously be in matching invoices and doing other tasks. So I would say for us, Campfire has been night and day better from, like, a security, from a performance, and from an auditability perspective. If it's a task that is not meant for Campfire, as I noted earlier, like building a board slide because we're also pulling in CRM data, then I I definitely will go to Claude because that's gonna be the better, place to start for that task. So I'm gonna go ahead and stop screen sharing. Again, apologies earlier that, it wasn't all sharing the entire time. I think we were able to to fill in the gaps there and hand it back over to Irina. Thank you so much, John. That was awesome. And we're going to be back next week. So please make sure to be here same time, same place. We're gonna be app building by using Replit. John will be back. We will share today's recording by early next week, so feel free to share that with your colleagues. Have them watch today's session before they come to the next three sessions. And with that, we're going to wrap up. Have a terrific end of your week, and we'll see you next week. Bye. Thank you everybody for joining today. Hopefully, that was helpful.