# I Built an Entire AI Finance Team With Claude (Full Tutorial)

https://www.youtube.com/watch?v=BhKdOfKPSaY

[00:00] Using different prompts and chats for different financial outputs while trying to keep the context of your project aligned will break down fast.
[00:08] Now, the outputs don't tie together, the logic will drift, and the moment you need to reuse it, you're having to rebuild everything all over again.
[00:17] But, in this video, I'm going to show you how to build a complete AI finance team that works both independently and together on your project while maintaining full context at all time so that your output stay consistent across the entire workflow.
[00:33] Now, this doesn't replace your role, it automates the execution of work so that you can stay focused on decision-making.
[00:40] Now, as a quant and finance automation developer, I'm going to show you how to set up Claude correctly, how to build five financial roles such as revenue operations manager, financial control manager, FP&A manager, reporting strategy manager, and a workflow coordinator, and then show you how to run them both independently and connected system on all your
[01:02] Projects.
[01:04] Now, this allows you to turn from data into a fully structured executive report and presentation-ready output.
[01:12] Now, if you want the prompts and the instruction files and the datasets all from this video, they're free in my community, which is in the description below.
[01:20] Before we can build a finance team that actually works together, we need to set up the project.
[01:25] Now, what we are building here is a drive folder with different files inside that Claude read and then use to operate each role on the team.
[01:36] Now, the whole project lives in one main folder called finance team.
[01:40] Inside that folder, there are five subfolders, one for each of the four roles on the team and one called shared that holds the company-wide context that every role needs.
[01:52] Now, in a real finance team, each function works on its own data and then produces its own outputs, but everyone operates against the same business context, and that structure is
[02:02] What allows different roles to operate independently while still being able to stay consistent with each other.
[02:08] Now, inside the shared folder, there are three files that ground the entire system.
[02:13] The first is company context.
[02:16] The file name is company_context.md, and this describes the business.
[02:23] Now, in our case, it is a SaaS company called Flow Sync.
[02:26] Now, its pricing model, customer segments, and growth stage are all included.
[02:34] Every role refers back to this when it needs to interpret what the numbers actually mean.
[02:39] The second is an assumptions file, which is named assumptions.md.
[02:44] And this is where growth rates, churn targets, hiring plans, and cost behaviors live.
[02:51] Now, if you want to model a different scenario later on, you just change this file once, and then the next forecast will pick it up automatically.
[03:00] The third is the calendar file, which is called calendar.csv.
[03:03] And this gives every role a single source of truth for which months and quarters are being referenced.
[03:10] So, when one role talks about Q1, it matches exactly what every other role is using.
[03:17] Now, all of these role folders sit alongside the shared folder, and that is structured the exact same way.
[03:25] They have an instruction file that defines the role and two subfolders, one for inputs and one for outputs.
[03:32] Now, with the folders in place, I now connect the project to Claude.
[03:37] So, I open Claude CoWork, I click choose a different folder, and then I select the finance team folder as the working directory.
[03:44] And from this point on, Claude has access to every role folder, every input file, and the shared context, and that is what allows the team to operate as one connected system instead of four disconnected tasks.
[03:58] Now, with the project set up, we can now start building the team.
[04:03] Building the first role, revenue operations manager.
[04:06] To build a finance team that works together, you need to know how you build each role separately first before you can let them work together.
[04:14] So, the first role I'm going to build in Claude is the revenue operations manager.
[04:19] I'm starting here because this is the role that owns the revenue picture and everything downstream from that, including forecasting, reporting, presentations.
[04:29] It depends on having a clean view of how the business is generating money.
[04:35] Now, the role definition lives inside the revenue operations folder, which is revenue_ops.md.
[04:42] And this file is the entire role.
[04:45] It describes what the manager is responsible for, which input files it works with, what calculations to run, what the output should look like, and where the output needs to be saved.
[04:55] Now, once the file is written, then the role exists.
[04:58] The system reads the file and it executes.
[05:03] So, we do not need to re-explain what the manager does every single time that we run it.
[05:08] Points to three input files, which are customer data, subscription data, and pricing plans.
[05:14] So, these are customer_data.csv, subscription_data.csv, and pricing_plans.csv.
[05:23] Now, these three together give the role everything that it needs to understand the business.
[05:29] Customer data tells you who the customers are, subscription data tells you how they pay and how that changes over time, and pricing data defines what each plan should cost.
[05:40] Now, the role uses all three to calculate monthly recurring revenue, annual recurring revenue, churn rate, customer growth, and the breakdown of revenue across different plans.
[05:53] Now, to run the role, I'm in Claude CoWork with the finance team folder already connected as the working directory.
[06:01] I click new task, and I use a short prompt telling Claude to act as the revenue operations manager and follow the instructions inside the
[06:08] Revenue operations folder, which is revenue_ops.md.
[06:13] The prompt is intentionally short because all the actual logic is in the MD file.
[06:21] Now, once the task starts, Claude then reads the role definition, it scans the input folder, and it confirms it has all three files.
[06:27] It loads a dataset of the customers and the subscriptions, and then starts building the analysis.
[06:34] Now, because the file specifies exactly what the calculations that it needs to run are, Claude does not need to ask any clarifying questions.
[06:43] The output file, which is mrr_summary.csv, is saved in two places, its own output folder and the FP&A inputs folder because FP&A is the next role that will use this data.
[06:57] Now, every role saves its output in its own folder and to the inputs folder of whichever role needs it next.
[07:05] Now, that is what makes this team behave like a connected system.
[07:08] Instead of a chain of tasks where you manually have to copy files around.
[07:14] Now, with the first role live and its output already sitting where FP&A can pick it up later, we can now build the rest of the team.
[07:21] Building the rest of the team.
[07:23] The rest of the team includes three more roles, which are the financial control manager, the FP&A manager, and the reporting manager.
[07:29] Now, the good thing is that all of them follow the same pattern that we just used for revenue operations.
[07:35] I'm going to go through each one quickly, and if you want all the specific details and the complete breakdown, you can find them in the free community linked below.
[07:43] Now, let's start with the financial control manager.
[07:48] Its job is to make sure that all the financial activity in the business is accurate and properly organized.
[07:55] Now, while revenue operations focuses on what customers are paying, financial control focuses on what was actually recorded.
[08:02] So, it classifies every transaction as revenue, expense, or refund, and it validates revenue against the invoices, and then
[08:10] It aggregates everything into a monthly profit and loss view.
[08:15] The outputs, which are financial_summary.csv and expense_summary.csv, are saved to both its own output folder and to the inputs folder for the FP&A and reporting.
[08:29] To run it, I stay in the same Claude CoWork session, and I use a short prompt that tells Claude to act as the financial control manager and follow the role file.
[08:39] The system runs the validation, and it confirms there are no duplicates or gaps in the invoice references, and the financial summary shows total revenue, total expenses, and a net loss that signals a growth stage business operating in heavy investment mode.
[08:55] The FP&A manager, which is fpna_manager.md, is the third role, and this is where the handoff pattern really pays off because, unlike the first two roles, it does not work with raw business data.
[09:08] Its inputs are the outputs that the previous roles
[09:11] Produced, the MRR summary from revenue operations, the expense summary from financial control, and then the assumptions file from the shared folder.
[09:20] And that's why the FP&A had no input files at the very start.
[09:24] It uses the MRR summary from revenue operations, the expense summary from financial control, and the assumptions file from the shared folder, and then it projects the next 12 months.
[09:37] The output is saved into the FP&A outputs folder and the roles that need them next.
[09:42] The last role is reporting and strategy manager, which is reporting_manager.md.
[09:49] Now, its job is to take everything that the team has produced and turn it into something a leadership audience can actually use.
[09:58] The role file, which is reporting_manager.md, loads the financial summary, the forecast summary, and the company context.
[10:08] And it generates an executive report, which is executive_report.md.
[10:12] Structured into five sections.
[10:14] The report saves to the reporting outputs folder and the shared folder, so the whole system has access to it.
[10:22] Coordinating the team end to end.
[10:25] At this point, we have the whole team running, but it's not how the system is meant to operate.
[10:32] What you want is to give the team a single high-level objective and have every role execute in sequence without you triggering each one manually.
[10:44] Now, the way we do that in Claude is with a coordinator file, which is full_finance_workflow.md.
[10:52] And it tells Claude to operate as a finance workflow coordinator instead of acting as a single role.
[11:00] Now, its job is to manage the entire sequence that we've built so far.
[11:04] So, the coordinator enforces the order strictly.
[11:07] It cannot skip a role.
[11:11] It cannot proceed before the previous output is saved.
[11:14] And it
[11:13] Cannot reinvent the logic.
[11:16] Now, without that strictness, Claude would start improvising, or it would shortcut validation, and the whole point of having defined roles would just break down.
[11:27] Now, before running the coordinator, we also update the input data because this is where we want to see the team respond to new information.
[11:35] Now, the assumptions file gets switched to a more aggressive growth scenario with higher growth, lower churn, and an expanded hiring plan.
[11:45] The subscription data gets refreshed with stronger recent performance.
[11:50] So, this is the kind of update a finance team gets in real life.
[11:53] And it lets us see how the team responds to new information.
[11:58] Now, to run the team, I open a fresh Claude co-work session in the finance team folder.
[12:03] Because fresh sessions, they matter because we do not want leftover context from the individual role runs.
[12:11] So, I use a prompt telling Claude to act as the
[12:14] Finance workflow coordinator, and then follow the instructions in the coordinator file, which is full_finance_workflow.md.
[12:24] Now, Claude does not jump straight into calculations.
[12:26] It reads the coordinator file, it then scans the folder structure, and then it loads every role's instruction file, and it checks the available data sets.
[12:37] Instead of acting as one single role, Claude is mapping out the full workflow before executing.
[12:44] And only after mapping the complete workflow, does it then begin executing.
[12:49] Revenue operations runs first and gives us the complete view of MRR, customer base, as well as growth and sales plans with the updated subscription data.
[13:01] Then, financial control runs and gives us the average monthly expenses along with a detailed expense breakdown.
[13:08] FP&A then runs third, projecting MRR and ARR, and then reporting runs last and synthesizes.
[13:14] Everything into an executive report.
[13:18] After all four roles complete, Claude runs a final verification to confirm every output exists in its required location.
[13:28] So, from a single high-level instruction, the entire team has executed.
[13:34] And that is the difference between using Claude as a tool and using it as a coordinated finance system.
[13:40] Now, the team has done its work, but there's still one thing missing.
[13:45] From analysis to presentation, the output is still spread across multiple files.
[13:50] And to make any of this useful in a meeting, for example, we'd need to pull it all into one place.
[13:56] So, this is where the PPT builder file, which is PPT_builder.md, comes in.
[14:04] It sits at the top of the finance team folder alongside the coordinator.
[14:08] Now, its job is to take everything that the team already produced and to take it and turn it into a structured presentation.
[14:13] Now, the file
[14:16] Specifies which inputs to pull from and what each slide should contain.
[14:21] I stay in the same coordinator session.
[14:23] The prompt tells Claude to act as the finance presentation builder and follow the instructions in PPT builder, which is PPT_builder.md.
[14:33] Claude then loads the presentation file, and the PPTX skill pulls every output from the team, and then builds a deck in one pass.
[14:44] Once the slides are generated, the system converts them to images and runs a visual check for layout, spacing, contrast, and alignment.
[14:52] If something looks wrong, it fixes it before saving.
[14:54] And the check is what makes the deck come out of this presentation ready instead of needing a manual cleanup pass.
[15:03] Now, the deck saves in two formats, a PPTX file in the reporting outputs folder as the final deliverable, and a markdown version alongside it for easy editing later on.
[15:15] Now, this is the full system.
[15:15] Five role files, a shared.
[15:18] Folder, a coordinator that runs the team end to end, and a single prompt.
[15:25] And then, a presentation builder that turns the entire output into a deck.
[15:30] Now, once the structure is in place, you don't have to rebuild any of it.
[15:35] If new data comes in, the same team runs against it, and then the output just updates.
[15:40] So, your role here shifts when the system is running.
[15:46] You are no longer the one extracting data, or building the model, or formatting the slides.
[15:51] You become the person who's looking at the executive report and deciding whether the recommendations are right for the business.
[15:59] The team handles the execution.
[16:02] It's your judgment that makes the output worth presenting.
[16:04] See you next time.