What Happens When You Tell AI the Goal, Not the Process

Rebecca (00:03): Right, we’re live. Okay.

Charlie (00:04): We’re back.

Rebecca (00:05): Yes, we’re back and we had a pretty cool week with some very interesting AI conversations

Rebecca (00:12): And we started chatting about what we wanted to talk about today and it was another one of those moments where it was like we really should just be recording this because this is the good stuff. which is the whole point of this podcast is just so that we can let you hear the things that we’re discovering talking to accountants and kind of how we work with AI and see if the way we work can help others. So we had

Rebecca (00:36): One call in ALFA this week that we talked about all sorts of stuff. We learned a lot about the terminal and how to set things up in Claude and Chat GPT and then in TCR Charlie talked a lot about what having an AI native firm can look like. so what are you thinking about all this? What should we tell them?

Charlie (00:54): Yeah, they were both good chats, good conversations. I think we’d start maybe with the AI NATO firm kind of little presentation that I did at T C R.

Rebecca (01:03): Yeah.

Charlie (01:04): ’Cause it kinda bleeds over to the call at ALFA this week as well, in the sense that, what it looks like to be AI native, is A is gonna vary, right, between firm and firm and what size you are, what kind of technology you’re running, what software you’re using, and where you wanna take your firm. Because right now

Rebecca (01:24): Yeah.

Charlie (01:25): People are kinda a little bit divided of like how deep into AI do they want to go? Do they wanna just try and implement it kind of from the ground up? or do they wanna have this invisible AI whereas the client may not directly see that AI is involved throughout the engagement or throughout the process, but behind the scenes the firm is still using AI to, ship some of the work or automate some of the processes or whatever.

Charlie (01:50): may be. And then there’s fully visible AI in the sense that the client is fully aware and can can definitely see the AI running and that direct effect. and it was a good conversation kind of talking with some of the members in TCR about what an AI native firm even is. What is it what does that mean exactly?

Rebecca (02:09): Yeah.

Charlie (02:10): And the way that I was trying to explain it is

Charlie (02:12): You know, having it fully embedded into the firm, right? Into the fiber of the firm, in all processes throughout various parts of the engagement from beginning to end. Right now, people are trying to figure out where does AI fit in into their current firm, their current stack, their processes,

Rebecca (02:29): Yeah. Right.

Charlie (02:31): and it’s kind of like band-aiding AI onto their existing processes, their existing

Charlie (02:38): firms. And it’s funny because I’ve been working with some CPAs and other accountants and they’re trying to figure out all these workarounds that they can use AI to, use it with this existing software, whether it’s Tax Dome, whether it’s Ultra Tax, whether it’s Canopy.

Rebecca (02:53): Give us an example, like

Rebecca (02:55): what what would they be trying to do?

Charlie (02:57): Yeah, I mean there’s some clients right now that are trying to figure out how they can use AI with things like tax dome. Right?

Charlie (03:05): tax dome files or on the on

Charlie (03:07): T drive on your computer and the way that they’re stored is a bit different in the sense that you know if you try to open up Claude and have it read some of these folders in the T drive, it can’t access them. So

Rebecca (03:20): Yeah.

Charlie (03:20): you really have to figure out a way to work around and again it’s just one of these examples of trying to band-aid AI into the process. I’m not saying

Rebecca (03:30): Well what would

Rebecca (03:30): not band aiding that look like?

Charlie (03:33): Well that’s the question, right? And that’s what firm owners are gonna have to try and figure out is I mean, I’m not saying build your own tax dome or cancel your tax dome

Charlie (03:42): subscription or anything like that. But I am trying to figure out or have people figure out, okay, you know, how does this existing software work? What are the capabilities it has? Either it’s MCP or it’s API and where can AI fit in that, at least you know, to get the job done. Because we have work to do, right? We can’t sit here and try and reinvent the wheel. So

Rebecca (04:05): Right.

Charlie (04:06): the band-aid solution that a lot of people are trying to do right now

Charlie (04:10): I’m not saying it’s a wrong path because again, you gotta start somewhere. I think the

Rebecca (04:12): Mm-hmm. Yeah.

Charlie (04:15): the key is you know, having this goal in mind, right, of of where you want to take your firm. And this is what I was trying explain to the members in TCR was developing your North Star for your firm and how far you wanna take it when it comes to implementing AI and then work backwards from there. So if

Rebecca (04:34): So what could a a

Rebecca (04:35): North Star look like? give us an example.

Charlie (04:37): Well,

Charlie (04:38): like for me at least it’s being an agent native firm.

Charlie (04:43): Right, I wanna run numerous agents in practice and have them doing a lot of the work. And then, that way when I come in, I’m left with the review, I’m left with talking to the clients, I’m left with, creating customized deliverables for the clients, and spending more of my time there than on the actual work, right? Compiling the documents, creating the work papers.

Charlie (05:07): getting things stored properly, updating statuses, all that I’m trying to have agents drive. And then that way again I can come in and spend more time on those other places that I mentioned. But people may not want to get that far. maybe they

Rebecca (05:21): Yeah.

Charlie (05:22): want to say, well

Charlie (05:23): You know, I like doing the work. And maybe they like doing the work. They wanna be more involved in pieces of the work and they don’t want to fully, you know, agentify their firm. Well, that’s fine. Like again, I think

Rebecca (05:35): Yeah.

Charlie (05:35): people ’cause if, you know, someone is saying, Hey, I’m probably only gonna work five more years and then I wanna retire then maybe

Rebecca (05:42): Yeah.

Charlie (05:43): they don’t spend all this time of, you know, agentifying their firm or going fully AI native. But I still

Rebecca (05:48): Yeah. Their north star would be different,

Rebecca (05:51): right? Because

Charlie (05:52): Yeah, so their North Star

Charlie (05:53): looks a lot different than than what mine does. and so their goals are different. Thus the ideas that they can implement when it comes to AI will differ. And so it’s just gonna look vastly different. But what I was telling them is like get super, super clear on whatever that is, and then use AI to work through that and say, Okay

Charlie (06:13): I feel good about my vision. This is where I can see my firm in five years. how do we get there? What does that look like? What’s a six-month roadmap? What’s a one-year roadmap? What does all that look like? How should I be thinking about my employees? How should I be getting them set up and aligned with this North Star, with this vision that I have and us all working, you know, collectively towards it with implementing AI to the extent that it’s necessary to get to where they want to be.

Charlie (06:41): and so

Rebecca (06:41): Do you think that

Rebecca (06:42): we all need to have a backup plan too? Like just in case, you know, the price of you know using these models changes or the government shuts AI down or like I don’t know. do you think that

Rebecca (06:56): It is there something is there something that you could picture being too risky to go all in with right now that wouldn’t be reversible if you had to go back to the old way? is there anything we should be keeping in mind with that? should it be local things that we’re building so that it doesn’t run on, tokens or

Charlie (07:16): Well, think it’s definitely s wise to not put all your eggs in one basket, right? But

Rebecca (07:20): Mm-hmm.

Charlie (07:20): like this question came up in the in the chat, right? Because one way that I explain AI native firm is that if you remove AI, then the firm cannot operate. And whenever you

Rebecca (07:31): Mm.

Charlie (07:32): s when you explain it like that, this is the typical response is, Well that doesn’t even sound good because if AI ceased to exist, my firm cannot operate.

Rebecca (07:42): Right.

Charlie (07:43): But I

Charlie (07:43): responded with, Well, how’s it any different than like a firm owner? If a firm owner goes out, how much of the firm can really continue to operate for a period of time? Because for a lot of firm

Rebecca (07:54): That’s a too.

Charlie (07:56): owners, do a whole lot. And so if they go out, or if you have an important employee that goes on maternity leave or, you know, goes gets sick and is down for a while, like that’s a disruptor as well. Yeah, so

Rebecca (08:07): What a good point. Because people

Rebecca (08:09): are taking on that risk every day by being the bottleneck and you know, awful, horrible, traumatic things happen every day to people. And

Charlie (08:17): Yeah.

Rebecca (08:17): we’re actually living with a lot of risk in our firms. Just a different

Charlie (08:21): Yeah.

Rebecca (08:22): kind. Yeah, that’s a good point.

Charlie (08:24): Yeah, so I think you just gotta be adaptable. I mean, anything could happen, yes, right? But in my mind at least, if something like that were to happen, I’ll figure it out. I mean I mean yeah,

Rebecca (08:33): Yeah, you’re

Charlie (08:34): we I think a lot of us are are are obviously very, very smart in this profession in the sense

Rebecca (08:39): Yeah.

Charlie (08:40): that we can figure it out. I mean if we have to do a return by hand, I think we could get it done. that no one wants to do Yeah, no one wants

Rebecca (08:45): Yeah, we won’t forget how

Charlie (08:47): to do that, right?

Charlie (08:48): But we still have these fallbacks, which is the traditional way right now,

Rebecca (08:52): Right.

Charlie (08:53): which none of us like doing, which is why we’re trying to implement AI and to make these processes better. But,

Rebecca (08:59): Yeah.

Charlie (08:59): you know, the reality of that is if you do make change and that change changes

Charlie (09:05): then obviously you’re gonna be in a tough spot. But again, the key is being adaptable and being

Rebecca (09:10): Yeah.

Charlie (09:10): able to say, Well, that’s a risk I’m willing to take. Like for me it’s a risk that I’m willing to take. Because

Rebecca (09:15): Yeah.

Charlie (09:16): again, I have conviction in myself that if it doesn’t pay off, I’ll be fine. And I think that’s how

Rebecca (09:22): Mm-hmm.

Charlie (09:24): That’s kind of the mindset that I think a lot of people should have because we still have all the knowledge and the experience of how this work should be done that we can figure out a way to do it, kinda no matter what. But I think AI provides the opportunity to really do it in kind of the best way. this new this new technology and these new models are really kind of opening the doors to these new ways of shipping the work, delivering

Rebecca (09:47): Mm-hmm.

Charlie (09:48): the work, doing the work and all

Charlie (09:50): all parts of the firm I think it has a place to play. I think the key is determining how much you want AI to be involved in these different stages and then, you know, committing to it. Until you

Rebecca (10:01): Yeah.

Charlie (10:02): have strong conviction that warrants may be pivoting away from that. Like when I was telling them,

Rebecca (10:07): Mm-hmm.

Charlie (10:07): like, develop your North Star but don’t get married to it. You know, if something happens

Rebecca (10:12): Is it?

Charlie (10:13): that, you know, causes you to change course, like be flexible. Change course. Yeah.

Rebecca (10:17): Be flexible. Yeah.

Charlie (10:20): and then maybe you have a a

Rebecca (10:20): That makes sense.

Charlie (10:21): your North Star shifts a little bit, but you’re still working towards this overall goal. because I

Rebecca (10:26): Yeah.

Charlie (10:26): mean anything could happen and at the pace that things are changing is just extremely rapid and

Rebecca (10:32): Yeah.

Charlie (10:32): it’s very difficult to even know what three months will look like, let alone three years. But again

Charlie (10:39): The clearer you can be about where you want to go, the easier it’ll be to make a pivot if there’s

Rebecca (10:44): Right.

Charlie (10:45): a reason to make one.

Rebecca (10:46): Yeah, that makes sense. And you mentioned the traditional way and exploring a different mindset. and you know, I have found it very interesting to watch certain firm owners that I’ve done coaching with where they they don’t come from an accounting background at all.

Rebecca (11:03): they come from a totally different industry and they for whatever reason end up with an accounting firm and they do things so differently that are not the traditional way at all and it’s fascinating to watch how they approach the same bookkeeping and tax work that everyone else is doing but just from a very different perspective and that makes me think about the AI stuff too and could we just kind of zoom out and just think about

Rebecca (11:27): Our goal and what we’re trying to do, and not worry about the way it’s always been done, and just have a conversation with our AI agent and say, every month I’ve got to do bookkeeping for a hundred clients, and every month, I get these documents from them, and the end product has to be a PL because they need to know these things, and I’ve talked to the client, and these are the things that they

Rebecca (11:53): Really wish they had answers to every month. And here’s everything you need to know about their business. And here’s the places that I collect the information, and here’s the folders where all the old stuff that we’ve done so far is

Rebecca (12:07): How could we do this for a hundred clients every month? You know, with these with my goal in mind is to, help my clients sleep better at night with answers about how their business is doing and to be compliant, for tax purposes, maintain the books for that purpose as well.

Rebecca (12:23): Go for it. And I’m not mentioning, we use QuickBooks. I’m not mentioning any of those things. I’m just saying this is the goal. I’m not telling you how I currently do it. I just am curious your take as my agent,

Rebecca (12:36): How could we accomplish this? to me, that’s really zooming out, not thinking traditional, not thinking of using any of the existing tools we have, and just staying focused on the goal. And I’m not saying this is the only way or the right way, but I find it’s very interesting to explore that. I’m curious what you think about that.

Charlie (12:55): Yeah, this kind of goes into the the ALFA call that we had on Wednesday, where I was I was explaining my workflow and how I approach some of this work and how I use these models and I was you know telling everybody I start with the goal.

Charlie (13:07): I my North Star on any project, any engagement that I do, it’s goal driven, and then I work backwards from there. And so if it’s a month in close, like I say, okay, you know, this is what we have to do, this is how typical deliverable works, these are some of the processes and steps that are taken, here’s the source documents, give it all this information and then really, have a conversation with it

Charlie (13:31): Here’s how the process works on a more granular level. They have QuickBooks. you know, we use double, we use Claude to do this, we use these various things, and here’s some of the frustrations that we’re having. And then

Rebecca (13:44): Mm.

Charlie (13:45): ask it like, hey, is there a better way to do this? If there was a better way to do this, what would it be? What would that look like?

Charlie (13:52): and really go deep into that, into that conversation with whatever AI you’re using and try and figure

Rebecca (13:57): What does it mean to go

Rebecca (13:58): deep? Like what what types of things can I say to the agent to help like not just get a quick answer and move on, to really dig into this?

Charlie (14:08): Yeah. One thing I say all the time is, help me flesh this out. And it seems to work for me at times. But basically tell it. Like I wanna get deep on this. I want to really consider all the options and everything that I’ve given you. I want you to you know, take your time, and if you haven’t already told it to

Charlie (14:28): explain things as a non technical So like I

Rebecca (14:30): Yeah.

Charlie (14:31): want the information to be d digestible. and also what I do a lot is tell it to kind of break it out into various pieces. So I’ll be like,

Rebecca (14:38): Mm-hmm.

Charlie (14:39): okay, well let’s start with this. Right? Let’s really dig into this part of it, whether it’s the fact that they are using QuickBooks or that we are using double or that we have stuff in Notion and Google Drive. Like, you know, the clearer picture that you can paint

Charlie (14:54): of whatever it is you’re working on to the model that you’re using, the better it’s gonna be able to really take that and digest it and then present other opportunities and ways that it could be done. And then whenever it does, like really pushing back on it or asking even

Rebecca (15:12): Thank you.

Charlie (15:12): for more clarity or more detail and asking why. You know, okay, if we did do that, you know, if we got them out of QuickBooks and instead they’re

Charlie (15:23): doing it this way, what does that mean for X, Y, and Z? What does that mean for my employees who are used to working in QuickBooks? What does that mean for, you know, the client that only connects their accounts in QuickBooks? And so like really considering all these different options and then just continue to narrow down as you kind of go through this conversation. And then you know what I’ll tell it a lot of times is whenever I’m working on something like that

Charlie (15:49): you know, I tell it to create like a roadmap. Okay, like so it’s like, okay, if we settle on we’re gonna try to go down this direction, A, how could we start without breaking everything? Right? Like

Rebecca (15:59): Yeah.

Charlie (15:59): how could we try this on maybe one client? And if we did that, what would that look like? And then getting those detail, getting those steps, and then working through the steps. ’Cause I think that’s where

Rebecca (16:10): Yeah.

Charlie (16:10): a lot of people kind of get tripped up on is

Charlie (16:14): that you really have to see it from beginning to end whenever you wanna make a change

Rebecca (16:18): Anything.

Charlie (16:19): and kind of struggle through it, like really work through it, implement these various changes that it’s suggesting and then test it. And then provide

Rebecca (16:28): Yeah.

Charlie (16:29): feedback.

Charlie (16:30): Because the feedback part is so important when it when you’re working with AI. If you get a result and it’s not the way that you wanted it to be, it’s you know paramount that you tell it like, Hey, this is not this did not work out as expected. You know, why? Let’s dig into this. Let’s find out why this didn’t work out and figure out how we

Rebecca (16:50): Mm-hmm.

Charlie (16:51): could change it to make it better. And so

Charlie (16:53): you know, we had a good question from one of the members in in ALFA on that call, John, who was, asking, How would this even be done?

Charlie (17:01): You know, his company, his firm runs things a very certain way, and he’s trying to

Rebecca (17:06): Mm-hmm.

Charlie (17:06): think, you know, hey, how could we do this better? If you were to redesign a Charlie starting from scratch, how would you do it? And then, you know, I explain how for that specific question that he asked. But a lot of that was I would have a lot of conversations with my AI about,

Rebecca (17:22): Yeah.

Charlie (17:24): you know, how I’m thinking about it, why I’m thinking about it, help me, you know, flush

Charlie (17:29): this idea out and let’s figure out what it would look like in practice and you know just kind of start chipping away at it.

Rebecca (17:35): Yeah, and I think of it like you know, reimagine

Rebecca (17:40): my whole workflow for this service that we offer in our business. let’s take all of the, you know, existing structure and let’s just reimagine like if we were to do it a totally different way, what would that look like? And then please list in a nice chart for me what the pros and cons of this are so that I can really, you know, measure, you know, what if it’s worth doing it a different way. But I think we’re so in it. Like here’s a great example.

Rebecca (18:06): you know QuickBooks it’s like sometimes the bank feeds break and you gotta re-sync them and it and it’s like let’s zoom out like what are we trying to do here? We’re trying to get the information that’s on the bank statement onto the financials and reconciled right like that’s what we’re trying to do.

Rebecca (18:22): So if we’re looking at like maybe AI can help us like get those bank feeds to stop breaking or something, that’s like still being very zoomed in to the existing way. But if we were to reimagine and zoom out, we might be like, well, why not just get the bank statements directly I mean that there’s other things, but like, you know, zooming out

Rebecca (18:45): Allows

Rebecca (18:45): you to open your mind to what if I just found a way to get the bank statements and asked AI to generate financials with those bank statements and and give it context of all the other things it would need too. But I do think that like we’re rule followers and we are kind of black and white thinkers and what you know and that serves us well in our profession. But like thinking of that zoom out or that reimagining of like,

Rebecca (19:09): reimagine that for me. it’s just interesting to see what it comes up with when it’s it’s not like bogged down in this traditional way a firm does things. It’s like it’s such a different perspective.

Charlie (19:21): Yeah, it’s honestly my favorite part is like getting creative with the work.

Charlie (19:25): I mean, this is like one the first things I did. I was like, I don’t want to work in QuickBooks. I just despise QuickBooks. and it was the same thing. I was like, Well, what do I need QuickBooks for? Well, it’s got the bank feed. It’s like, okay, well, what does a client do in QuickBooks? Maybe the client does a lot in QuickBooks and they wanna work in QuickBooks. Maybe they send their bills, maybe they keep track of their ARAP, whatever it is they do, that’s fine. Are they looking at the numbers? Like I have I have a client

Charlie (19:51): that, you know, the only thing they do in QuickBooks or they don’t do it but their their admin does is like bill pay and stuff like that. Like they don’t they don’t really do much else in QuickBooks. and so I was like, well, I don’t want to work in QuickBooks and so I don’t work in QuickBooks. Right? And because, you know, what we need in QuickBooks, at least in in my case and for this client, was just the bank feed one of the transactions. Well you could you know you could set up your own

Charlie (20:16): software and you get a plaid connection, you could tell to go in and connect their accounts via plaid and then now I have, you know, the live bank feed Or you go the route of, hey, give me read only access to your bank accounts. I’ll go in. I could pull the the bank accounts and do whatever I want to with that. And I think now is like the first time really in history to where you could you just need the source.

Charlie (20:38): and then what you do with it, to get this work done, you can get creative with that part. And again, yeah.

Rebecca (20:43): A source and a goal, right? Like you need

Rebecca (20:45): the source and you need what what does the client want? What does the IRS want? This is the goal. Reimagine it for me. Tell like help me walk me through a different way and how can

Charlie (20:56): Yeah.

Rebecca (20:56): you help me to get there? And you know, and maybe you don’t want to adopt all of it, but

Charlie (21:02): Mm-hmm.

Rebecca (21:03): maybe it gives you some interesting, you know, perspective.

Charlie (21:06): Yeah, I think it’s great at pointing out things that you’re not thinking about or you might not be thinking about or you don’t even realize it’s possible.

Rebecca (21:12): Hmm. Yeah.

Charlie (21:14): Right. But again, I think it’s like the clear. It’s like you have the goal and then you get super granular of, you know, here’s the information, here’s where it lives, here’s how it flows, here’s what I do with it, here’s like how the client uses it, and like give all that to the AI and say, Okay, here’s how this process runs, here’s the

Charlie (21:34): the web of how the the information, the data flows, like how could

Rebecca (21:38): Yeah.

Charlie (21:38): we reimagine this? I do this all the time, I’ll say, How could this be 10X better? That’s a very common thing that I say is like,

Rebecca (21:44): Ooh, that’s a good question.

Charlie (21:45): how could this be 10X better? And then I read the response and a lot of times I’ll say, Well I don’t I don’t I don’t like that. I don’t or I don’t agree with that. But that piece is very interesting.

Charlie (21:56): That one I hadn’t thought about. Let’s explore that more. Like let’s let’s flesh that part out. It’s like one thing of, you know, maybe ten different things that it said, but it’s like, wow, that that’s actually has that could have some legs. Let’s let’s look into that a bit more and start

Rebecca (22:09): Yeah.

Charlie (22:09): thinking about how how that would work. And then you start going down this

Rebecca (22:12): And that’s like the human part

Rebecca (22:14): and the professional judgment part, like, you know, and the reviewing part. Like you you as you know, an experienced CPA, like you you’re using your brain and and all of your experience and all of your training to still have input in this process and to know what makes sense and what doesn’t. So there’s still a place for you, you know, in this.

Charlie (22:35): Yeah.

Rebecca (22:36): yeah, and I I think it’s interesting, you know,

Rebecca (22:40): People ask us a lot, I would love to see how you work in these things. And, you know, I don’t think I think this is kind of awkward for us because we never want it to sound like what we think we have it figured out or we do it the right way and you do it the wrong way. But at the same time, if sharing how we do something that and how it’s been successful for us could help someone, well then it’s worth, you know, being vulnerable and telling you guys how we do something.

Rebecca (23:06): Something. So I’ll preface this with this isn’t necessarily the right way, but this has been working very well for both of us, we both believe that our agents

Rebecca (23:17): have enough knowledge that

Rebecca (23:19): We just tell it what our goal is and where the source stuff is you know some parameters, some sort of guardrails or harnessing or whatever the buzzwords are right now. But like, you know, we give it some structure, but we really empower our agents to present us with their solution. In an alternative way would be to create a skill for everything you do in your firm and have

Rebecca (23:45): you know, a skill that each employee is running to do this task, you know, to work within the parameters of of still like how the firm has always run. And there isn’t a right way or a wrong way, but I do think people would be interested to hear a little bit more It just not intuitive to me. I wouldn’t think to use a

Rebecca (24:02): so without sounding like we’re better than anyone here like because that’s not what we think I think people would be interested to hear your take on that.

Charlie (24:12): I mean I definitely have skills that I’ve created on my machine.

Charlie (24:16): ’cause I’ve always taken that that goal oriented approach. and just explain what I’m trying to accomplish in very detailed. And I guess you people would say, Well you could just that could just be a skill, you could just turn that into a skill. but to me it was all about you know the concept of the harness. Right? So getting the right environment set

Rebecca (24:32): Explain that.

Charlie (24:33): up.

Rebecca (24:33): I still don’t fully understand that buzzword.

Charlie (24:36): So the harness is just basically if you’re gonna run a model, right, the environment that it works in. So let’s say you are running Claude and you’re in the desktop app, okay, and you’re trying to do a month in close. And you just launch Claude just regularly in the desktop app. You’re not pointing it at anything specific and you say, Hey, I need to do this month-in close on

Charlie (25:01): John Doe L L C whatever it may be. Well

Charlie (25:04): That harness that it’s now in, which is just Claude Desktop app, and it has tools that it can call, it can use. If you have MCP and QuickBooks, it can go into QuickBooks. But if you don’t really give it anything else, it has to figure out what it’s supposed to do. So if you say I need to do this month in close for John Doe LLC, it’ll be like, okay, month in close, John Doe LLC, that’s a business. I see I have a MCP connection to QuickBooks, let me go to

Charlie (25:32): QuickBooks and see if I can find John Doe LLC. Okay, I see it. Let me read all this information about John Doe. So it has to figure out what is what it’s doing and like what it’s gonna do, it has to figure out about the client. So that that right there means the harness isn’t optimized. Now, if you were to instead point Claude at a specific John Doe LLC folder.

Charlie (25:58): And inside that folder, you have a Claude MD about John Doe LLC. You have instructions MD about John Doe LLC about what you’re gonna do. You have prior month-in closes in there about John Doe LLC that the model can reference. It still has you know the MCP connections, it still has all these tool capabilities, but now you’ve opened it up inside this harness that has all this additional context, and you could basically at that point say, okay.

Charlie (26:26): Let’s do the month in close for John Doe LLC. And it can say, okay, I know exactly who John Doe LLC is because it had it can read the Claude MD, it can read these instructions. If you do have skills in there, it can run skills because it knows which skills to run for this client because it’s all laid out into the context that’s in this folder that you now have opened it up in. Right? and so that’s like the

Rebecca (26:45): Hmm. Okay. Can we talk about folders really quick?

Charlie (26:50): okay.

Rebecca (26:51): So okay, one way to have folders could be just like literally on your computer, like in File Explorer or whatever it’s called, like just, you know, or

Charlie (27:01): Yeah.

Rebecca (27:01): you know, folders and subfolders within them, right? That’s one way. Another way could be like a cloud based folder structure, like let’s say Google Drive or something

Charlie (27:10): Mm-hmm.

Rebecca (27:10): where you know there’s a folder for each client and subfolders within it, you know, and you could create these like MD files or ask Cloud to help.

Rebecca (27:20): figure out how to make an MD file where you just put all that but then there’s also inside of Claude and Chat GPT that like projects area where you could create a project for a client let’s say and then you can put like it there’s like an instruction section that you like a little box that you can say in there and you could put like some source documents there and have it it I think that’s kind of a harness right like that that

Charlie (27:45): Mm-hmm.

Rebecca (27:45): project so which one’s better like

Rebecca (27:48): Like what what’s the difference there? Like what should we do

Charlie (27:51): Yeah.

Rebecca (27:51): if we’re if we’re excited to build this now?

Charlie (27:54): Yeah, and I think this this definitely depends on how your team works.

Charlie (27:58): Right, if you have

Rebecca (27:58): Okay.

Charlie (27:58): a team. Like for me, I’m solo, so I don’t have any employees. So all of when I say folders, all my folders live on my local machine. Okay, so my computer, you know, your desktop is a folder. And then if you put a folder on your desktop, that’s a subfolder of your desktop. And so the way that I haven’t broken out and I kind of showed this in the in the ALFA call is that I have these like three main directories, three main folders that I work in.

Charlie (28:25): And within those folders, I have a bunch of subfolders and even subfolders and subfolders. But I have it set up, the harness set up, so that if I point Claude at any one of these folders, it has a Claude MD in there. It has a README, it has agent MD. It has all these context files. So that if I were to point Claude at one of these folders, it knows the contents of this folder, I know how this folder is operated or what’s typically done in this folder.

Charlie (28:53): And so for the sense of the the analogy that I was getting about John Doe LLC, like if you

Rebecca (28:58): Mm-hmm.

Charlie (28:59): set up, you know, if you have a practice folder and you have a client subfolder, and then you have John Doe LLC as a you know subfolder of the client subfolder in there, like that’s that harness because the practice main folder has a Claude MD that says, hey, this is practice folder, it holds all clients and their client documents, it has maybe firm documents.

Charlie (29:21): It has other things related to the practice. And then you go into the client folder, and it’s got a Claude MD. Hey, this is the client directory. It has all the subfolders are clients, and this is kind of how you should operate when you’re in this folder. And then again, in John Doe, you can have like all these folders of like here’s information on the client, here’s prior month end closes, here’s you know prior tax return, here’s all this information on John Doe, right? And then here’s the typical engagement, and you can get you know you can get.

Charlie (29:49): pretty specific with it so that again you could come in here and open up Claude in John Doe and say hey let’s run the month and close and it has all this context you have the harness set up right to where it could say okay I know exactly what I do for John Doe when I’m in this folder and I can see that you know maybe August hasn’t been closed yet. I see only through July so let’s start the the August close and it knows exactly what to do.

Charlie (30:15): because it has all this context.

Rebecca (30:15): I have two questions.

Charlie (30:17): Okay?

Rebecca (30:17): So, okay. the first question is

Rebecca (30:20): Are you just like smarter than the rest of us and you know how to make like a Claude MD and an agent MD folder or document in the folder? Or do you literally just like use Whisperflow and talk to Claude and say, this is the process, this is everything, like make a Claude MD document and put it in this folder? That’s question one. And question two is what about maintenance? Like do we have to remember when something changes with the client to tell it that? How does that work?

Charlie (30:50): So the first one you can definitely have Claude set up your Claude MD.

Rebecca (30:53): Okay.

Charlie (30:53): and I would if I’m you know, or for anyone out there, I would definitely say I’d go into Claude right now and say, Do I have any Claude MD set up on my computer? I mean if you work on your local computer, because again if we go back real quick to the the idea of the projects, right, you could do a very similar thing within the projects.

Charlie (31:16): To to me, I’ve just found it better and easier to do it the way that I do it on my local machine. Cause I mean I work in the terminal, I work via CLI. You could get pretty close, I think doing it this way may be slightly better, but I don’t know for sure because I don’t work in the projects in the desktop app.

Rebecca (31:32): I wonder if collaborating

Rebecca (31:33): with other like teammates would be harder or easier one way or the other.

Charlie (31:38): Well, that’s for sure.

Charlie (31:39): I mean the good thing about projects is you can share projects with someone else, right? So you

Rebecca (31:43): Okay. Okay.

Charlie (31:44): could work pretty much jointly on a project within the desktop app if you have a team member you you could invite them to this project and y’all could work together on it. but it’s a very similar concept in the sense like

Rebecca (31:54): Okay.

Charlie (31:54): you set up a project for John Doe and then you you give it all this context, you upload it into project files, you set instructions on there, and then you would start working through

Charlie (32:04): through whatever you need to do, it’s very similar structure. it’s very similar idea. and then the Claude MD, yeah, those are markdown files. Dot MD is a markdown file. and it’s the first thing Claude reads whenever it goes into any directory, wherever you point it, it looks for these Claude MD files. Because it’s like

Rebecca (32:23): Okay.

Charlie (32:25): you know, overall kind of instructions of like where it is. It gets oriented whenever you point it there.

Rebecca (32:30): Mm-hmm.

Charlie (32:30): And so I would if it were me, I mean first thing

Charlie (32:33): Either way you’re working, whether you’re working in projects or you’re working like I’m working, I would like take the time to say, can you audit my setup? Like here’s what I’m trying to do. Am I doing

Rebecca (32:41): Mm. The goal.

Charlie (32:44): it right? Right, goal. Right? And better yet, is it set up for the agent to do it right?

Rebecca (32:53): Hmm.

Charlie (32:54): Okay, so like do you have all the information you need? Do you have these context.md files? Is there a Claude MD file? Is there clear instructions for you? Is it too much? Is there too much instructions? whatever it may be, I would just straight up ask it and say, Okay, if it’s not optimized for you to work in this environment that I have set up to try to accomplish this goal, how do we get it set up right? I mean this is the one thing I think

Charlie (33:18): no one told me this but I just kind of it was a bit intuitive to me when I first started, it was like I want to use you as best as I can and I want you to have the best output. So what does that look like for you? Like Claude, what

Rebecca (33:31): Hmm.

Charlie (33:32): like what do you how do you need to be set up so that you can do what I’m trying to do in the best way possible? Yeah, I mean I just asked it.

Rebecca (33:37): You just asked.

Rebecca (33:40): Genius.

Charlie (33:40): and it was like

Charlie (33:40): based on how you’re working, like I’m building apps, we have T C R we have AI labs, we have real firms, we have all this stuff going on, and I still have my own practice. And it’s like, Okay, well, you have all this stuff going on, you should probably separate out. They should have their own Claude MDs, they should have different README’s, you should be able to do certain things, you know, differently depending on which directory you’re in. and you know, it kind of got me all

Charlie (34:04): set up and I mean it took a while. It took several iterations before it evolved to to kind of what it is today to where, you know, it’s working fortunately very well. But again, it’s it’s just this that kind of journey of goal and then figuring out like, okay, suggestions from Claude. let’s implement that. Let’s test it. And then let’s iterate it if we need to. Let’s make it better. Let’s keep making it better and better. And I think

Rebecca (34:27): So I know yesterday

Rebecca (34:28): you had a a client meeting, like a lunch that you went to it with to it with a client, and I’m guessing that there was some new information about, you know, the scope of the engagement, like what what they’re looking for. So what do you do? Do you run back to your computer and say like we gotta update the Claude MD on this client? Is that how this works?

Charlie (34:49): Well, yeah, pretty much. So I came back from that lunch with these two clients

Rebecca (34:52): Okay.

Charlie (34:53): and I literally sat down at my desk and said, Okay, just had lunch with, you know, these two clients. Here’s some of the things we talked about. and it did it naturally. It said, Okay, updating Claude D or actually client D. So like you can have a client D. This is

Rebecca (35:09): Mm-hmm.

Charlie (35:10): going back a little bit to the John Doe LLC, like there’ll be a client MD for John Doe LLC. And that basically explains who this client is, you know,

Rebecca (35:18): Mm-hmm.

Charlie (35:19): anything important, things to consider. And so Claude just updated that client MD for me. Yeah.

Rebecca (35:25): Just like you would tell an assistant, like, hey, I just had

Rebecca (35:28): a meeting, like here’s the follow-ups of what we need to do. Yeah.

Charlie (35:31): Yep, yep. I mean it’s

Charlie (35:32): I think that’s such a big point is that if people can start thinking of these models as like employees.

Charlie (35:39): you do very similar type stuff of like you gotta give the employee all the information it needs to do the job, you gotta

Rebecca (35:45): Yeah.

Charlie (35:45): get it set up, you gotta get the right environment, you gotta get it access to the tools, it needs a QuickBooks

Rebecca (35:49): Mm-hmm.

Charlie (35:50): login, it needs Google Drive login, it needs all this information.

Rebecca (35:53): Yeah, same

Charlie (35:54): and it needs

Rebecca (35:54): thing.

Charlie (35:55): to get up to speed on where the client stands. And I think these agents just work a little bit differently. Obviously they read a bunch of

Rebecca (36:01): Mm-hmm. Yeah.

Charlie (36:04): text and mark down, but they still need access to the same information. And so

Rebecca (36:07): Okay.

Charlie (36:09): yeah, that’s kinda how I would think about it.

Rebecca (36:10): So if that meeting had

Rebecca (36:11): been like a Zoom meeting instead, could you have just handed it the transcript and said l and say like figure out what needs to be updated?

Charlie (36:21): Yeah,

Charlie (36:22): I mean same concept.

Charlie (36:24): You could say just go pull the transcript, from the recent conversation we had and you know, there’s gonna be probably a lot of maybe a lot of noise in the transcript

Rebecca (36:32): Mm-hmm.

Charlie (36:33): in the sense like a lot of information you don’t really need, but there could be some good information there that that drives a change to the client MD or whatever it may be. And so it’s again the persistence of updating and keeping things in sync.

Rebecca (36:47): Yeah.

Charlie (36:48): and

Charlie (36:48): having that like real time information available and making sure stale information doesn’t continue to carry forward across the engagement.

Rebecca (36:56): Okay.

Charlie (36:57): And that does take some work. I mean you can try to automate that part as well. Like, you know, anytime you have a call with a client, there’s a transcript, like maybe you send it you have an automation that sends it somewhere and you can set up a cron job to where this gets pulled, you know, whether it’s daily,

Rebecca (37:11): Yeah.

Charlie (37:12): weekly, whatever it may be, and it goes and updates the right places. you can definitely get to that point.

Rebecca (37:16): Yeah.

Charlie (37:16): I don’t do that. it’s just easy for me to just say, Hey go.

Charlie (37:20): And you know, take this transcript or here’s the updates and just make the changes. But I can imagine

Rebecca (37:24): Mm-hmm.

Charlie (37:24): if you’re working at scale and you’re dealing with hundreds of thousands of clients, it would be worth setting up some of these cron jobs and having a lot of that automated. But that’s that’s

Rebecca (37:33): Yeah.

Charlie (37:34): that’s where you go back to the goal and say, you know, I’m

Rebecca (37:36): Right, it’s a goal.

Charlie (37:38): having trouble keeping all my clients updated, and here’s how it works. I have a call, it goes here, and it gets stored here, and then I have to update this. Right now it’s done manually. How do

Charlie (37:48): we fix this? How do we make this ten X better? And then it tells you and and

Rebecca (37:51): Yeah. Yeah.

Charlie (37:53): then you you work through implementing kind of whatever it tells you and then

Rebecca (37:56): And you still have to do

Rebecca (37:57): this without, like in a regular firm, we still have to, after we have a meeting with the client, figure out like, okay, we need to add more projects, you know, that more we have to send them an engagement line for this new service, we have to their pricing needs to change probably because we have to do this or they just started a new business. Like this isn’t different than what we already have to do, right?

Charlie (38:20): Yeah, I mean we we’ve

Charlie (38:21): been managing context and information, the whole career. I mean it’s a large part of our job. client

Rebecca (38:26): Yeah.

Charlie (38:27): will call you and tell you something, you’ll write it down on your notepad and figure out, okay, where do I store this actually information? how do I keep it up to date? Because, come tax time

Charlie (38:37): Right, if they told you something back in October, you know, they had a child or whatever it may be that, you know, impacts the return and you didn’t capture that context, then you know, it’s a problem. I think, the great thing about AI it’s a lot easier to I think to capture that context and carry

Rebecca (38:53): Right.

Charlie (38:53): it forward. it makes it a lot less friction, to do

Rebecca (38:56): Yeah.

Charlie (38:56): that. And so

Charlie (38:57): But again, iron out your own process, like figuring out like what is your process and how could it be better and where does AI fit into it? And then I think once you’ve gotten to a point to where kind of going full circle here, to where you have implemented maybe put some of these band-aids on a lot of places that are that are bleeding in your firm and you got

Rebecca (39:15): Mm-hmm.

Charlie (39:16): it working better to frees up some of your time, then you can start thinking about, okay, well, how could we remove that band aid

Charlie (39:23): by implementing maybe a more agentic or more of an AI role into this process to where we no longer have to like kinda have it this way, which is slightly

Rebecca (39:33): Right.

Charlie (39:34): better, but again going back to how do we make a ten X better. and then

Rebecca (39:37): Yes.

Charlie (39:38): what would that look like and and working through that?

Rebecca (39:40): You’re probably not gonna get to 10x with the band-aid, but you’ll get a couple X’s and you’ll and that’ll

Charlie (39:46): Yeah.

Rebecca (39:47): be great, like because things you’ll have some more capacity, you know, the bleeding will stop. but 10x probably would involve really just reimagining things, I feel like.

Charlie (39:58): Yeah, I think so. but you know, you have to start somewhere, right? And you have to be realistic

Rebecca (40:02): Yes, yeah.

Charlie (40:04): with your current workload, your current clients, your employees, and kind of take these steps that you

Rebecca (40:09): Yeah.

Charlie (40:10): know, hopefully if you want to get there, if you want to ten X and you want to really make these changes, then you know, again, figuring out that path.

Rebecca (40:18): Mm-hmm.

Rebecca (40:19): Okay, so

Rebecca (40:20): The goal thing is like a huge takeaway for me. Like just tell your agent what your goal is, what you wish could happen, what you wish, like what would be so amazing if this could be, and then have it help you get there with a new perspective. I love that. And then what do you think, kind of closing this out, what should we go do now? Like we just heard all of this, like what’s one thing we maybe could go try?

Rebecca (40:47): Like what’s one conversation we should go have with our agent after this call?

Charlie (40:52): I would kinda do this, you know, a little soft audit of hey

Charlie (40:56): Wherever you work, I mean if you’re in the chat, if you’re if you’re using heavy projects, you know, if you’re in Cloud Code, in the terminal, wherever you may be, I would, you know, start there and say, Hey, I’ve been working here, this is what we’ve been doing, here’s like some of the frustrations that I’ve had. You know, maybe you repeat yourself a lot, maybe you have to like constantly provide information to it or whatever it may be.

Charlie (41:20): and say, how could it be Ted X? How could this be better?

Charlie (41:24): I don’t feel like I’m fully utilizing your capabilities, right? I hear all these amazing things that people are doing crazy things with AI and I haven’t gotten that light bulb moment. Or even if you

Rebecca (41:34): Yeah.

Charlie (41:34): are pretty advanced and you’re like, okay, I feel like I’m pretty advanced, but still I think there’s room to get better. What would that look like

Rebecca (41:42): Yeah.

Charlie (41:43): for me? You know, how could we be doing this in a better way? What do

Rebecca (41:46): Right.

Charlie (41:47): you need, talking to the AI, what would you need in order to maybe do this better?

Charlie (41:52): to do this faster so that you can take on more and not have to rely on me so much.

Charlie (41:58): And this is going back to really letting these models, you know, run for longer periods of time without human intervention. You know, how do we get to that point? Because, you know, right now I do feel like there’s a lot of people are still doing a lot of prompting. You know, the agent runs for

Rebecca (42:13): Mm-hmm.

Charlie (42:13): maybe a little bit and it comes back and say, okay, what next? you know, and again,

Rebecca (42:17): Yeah.

Charlie (42:17): this goes back to setting the goals. I mean, if you set a goal for an agent and you really, really dial in the steps that are needed.

Charlie (42:24): to get to that goal. then you could let the agent work towards that goal and give it that autonomy, give it that freedom. So I would

Rebecca (42:31): So that’s when we

Rebecca (42:32): hear people say, I woke up and this was built, you know, overnight for me. It’s because they really took the time to explain what the finish line would look like and you know, had that dialogue and then let it just go on its own.

Charlie (42:49): Yeah,

Charlie (42:49): a lot of people are really focused on task right now, like hey we gotta get this one task done, like here’s how we do this task, okay, now we have this task to do. Again,

Rebecca (42:58): Yeah.

Charlie (42:59): all working towards this one goal.

Rebecca (43:01): Right.

Charlie (43:02): And

Charlie (43:02): but there’s still there’s there’s task, task, task, task. I think you could kinda say, let’s just think about the goal more, break it out, get really, really clear on what you’re supposed to do to reach this goal, and then say, Okay, do you have do you know what you need to do? Do you do you have everything that you need to do to reach this goal? Yeah.

Rebecca (43:18): Just like you would with an employee. Yeah.

Charlie (43:20): If you don’t, let me know. Like what’s still needed.

Charlie (43:24): And or is this the right environment for it? Should we have should we be in Claude Code? I’m in chat

Rebecca (43:28): Yeah.

Charlie (43:28): right now. Should we go to Claude Code? Likely. okay, we’re in Claude Code. Like, should I be in projects? Should you work in this project or should I give you a local folder on my machine? And should we get that local folder set up, right? The file and folder architecture of the Claude MD, the client MD, project MD, like set all these context files up so that you don’t have to ask me for questions.

Charlie (43:49): you you have all the

Rebecca (43:50): Yeah.

Charlie (43:50): information you need. and even going further, like if there’s something that maybe you don’t know, how does the agent handle that? Where does that go? So that when you do come

Rebecca (44:00): Mm-hmm.

Charlie (44:01): in, you can see the assumptions the agent made, the items that were marked for review. Right, you can just come right in, sit right down and be like, okay, you just ran for seven hours. What did you do? Without being bombarded

Rebecca (44:12): Yeah.

Charlie (44:13): with a whole bunch of information. So

Charlie (44:15): Again, this takes time of like getting the proper

Rebecca (44:17): Mm-hmm.

Charlie (44:18): setup, but that is the first thing that I would do. I tell people that all the time is like really get deep on your setup, on your harness, on how you’re

Rebecca (44:25): Yeah. Yeah.

Charlie (44:27): working with Claude and ask it.

Charlie (44:30): Am I using you the right way? And if not, like how? Like tell me. I want to learn how to use you better and let’s actually do it. Like let’s get it set up. if it does say that there’s some room for opportunity, like okay, well let’s let’s get it set up. What does that look like? Help me work through it or just do it.

Rebecca (44:45): I just got an idea. What if people took the transcript from this call that we’re having right now? and they just gave it to their Claude or ChatGPT and said, like, I want to do that. Would that get them started? Like if people feel like overwhelmed by what you even just said that they should go do? Like is that kind of a shortcut?

Charlie (45:05): Yeah, if

Charlie (45:07): You could definitely do that. I would say like focus in on this this one part.

Charlie (45:11): again I think the key is an open mind when you’re working with these models. A. but B, if it were me and I was listening to this and

Charlie (45:19): you know maybe I wasn’t as as advanced, I would for sure do that. I would take the transcript and say, hey, he talked a lot about setup, harness, Claude md, client md, folder, versus projects. can you go and take the transcript and really focus in on that part and explain it to me in layman’s terms. And also

Charlie (45:40): apply it to how I’ve been working with you, right?

Rebecca (45:42): Yeah.

Charlie (45:43): And say

Charlie (45:44): you know, try to give me some areas of improvement or opportunities or things that we could change and just lay it all out for me. I mean create an

Rebecca (45:49): Yeah.

Charlie (45:50): HTML file, turn it into a visual if you learn that

Rebecca (45:52): Yeah.

Charlie (45:53): way. explain this concept of a harness to me as a visual. I mean I’ve done that many a times. I’ve had my own harness

Rebecca (45:59): Maybe.

Charlie (46:00): visualized to me by Claude so I can try to understand it, you know, exactly how it works ’cause it can get pretty complex.

Rebecca (46:07): Yeah, because

Rebecca (46:07): you know you’re a visual learner, so you ask for it in a mode that makes sense for you. And I even think you could ask it, here’s the transcript, break it down into the different topics that Rebecca and Charlie talked about, and then like let’s go through each of those together and see how we can optimize our setup, but let’s do it like one at a time or something, you know. there’s a lot

Charlie (46:28): I did this.

Rebecca (46:29): we could do.

Charlie (46:29): Yeah. I did

Charlie (46:30): that this this morning. So I went on a run this morning. you know, I ran for like an hour and I talked to Chat GPT the whole time. And I about four or five various topics. And then as soon as I got back to my desk, I said, Okay, can you break out this this thread that we had into the various topics that we talked about? And

Rebecca (46:48): Mm.

Charlie (46:49): then like there was one that I really wanted to focus on. And so I took that part of it and I was like, Okay, I went to Codex and I was like, Okay, let’s let’s

Charlie (46:58): flesh this out. Let’s dig into this more. I want to get deeper on this. but yeah, it’s a great way of like separating out again like some of the noise from like your signal. So figuring out like what’s your signal, what you want to focus on, and then really focusing on it and going deeper into it.

Rebecca (47:15): so cool. Okay. I hope people do that and that they tell us about how it went. I’m really curious to hear people experimenting with this stuff. And yeah just ask Claude.

Charlie (47:26): I’m always trying to learn, very open to to learning new things and I think that’s like kind of the mindset you really need to have when it comes to

Rebecca (47:33): Yeah.

Charlie (47:34): this is just saying, I mean these models are very smart. I’m I’m not claiming to be smarter than these models.

Rebecca (47:40): Yeah.

Charlie (47:41): I think again but like having this mindset of like okay it’s a tool. It can help me do my job better, it can help me accomplish things, it can help me formulate my thoughts and my ideas. just really thinking about it as a partner and kinda treating it that way, I think that’ll it’ll just kinda go a long way.

Rebecca (47:58): Yeah, I love that. Cool. Alright, I can’t wait to hear what people end up doing with this transcript.

Charlie (48:05): Yeah, yeah, that’ll

Charlie (48:06): be interesting. But all right. We’ll we’ll see ya next week.