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

Core idea / problem addressed

• If you try to use multiple prompts/chats for different financial outputs while keeping “context” manually, logic drifts and you end up rebuilding when you reuse it.
• Solution: make Claude operate as a multi-role system where every role reads shared context and writes outputs to the next role, preserving consistency across the whole workflow.

---

System design (Claude project structure)

Top-level folder

A main folder named finance team contains:

• shared/ (company-wide context used by all roles)
• Role subfolders (each with its own instruction file + inputs/ + outputs/)

---

Shared context files (grounding the system)

1. company_context.md

   • Describes the business (example given: SaaS company “Flow Sync”)
   • Includes pricing model, customer segments, growth stage

2. assumptions.md

   • Growth rates, churn targets, hiring plans, cost behaviors
   • Changing it once updates future forecasts automatically (scenario switching)

3. calendar.csv

   • Single source of truth for which months/quarters correspond to references like Q1

Why this prevents drift

Roles run independently on their own required inputs, but all interpret numbers using the same shared business context and the same calendar definitions.

---

The 5-role “AI finance team” (role instruction files + what they do)

1) Revenue Operations Manager

• Instruction file: revenue_ops.md
• Inputs:
  • customer_data.csv
  • subscription_data.csv
  • pricing_plans.csv
• Calculates:
  • MRR, ARR
  • churn rate
  • customer growth
  • revenue breakdown across plans
• Output:
  • mrr_summary.csv
• Saved to:
  • its own role output folder
  • FP&A inputs folder (handoff)

---

2) Financial Control Manager

• Instruction file: financial_control_manager (referenced via “role file”; outputs named below)
• Focus:
  • Ensures accounting/recorded activity accuracy (vs. customer payment view)
• Actions:
  • Classifies transactions as revenue / expense / refund
  • Validates revenue against invoices
  • Builds monthly profit and loss view
• Outputs:
  • financial_summary.csv
  • expense_summary.csv
• Saved to:
  • its output folder
  • FP&A and reporting inputs folders
• Validation checks include confirming no duplicates or gaps in invoice references.

---

3) FP&A Manager

• Instruction file: fpna_manager.md
• Key property:
  • Does not use raw business data directly
  • Uses outputs from prior roles + shared assumptions
• Inputs:
  • mrr_summary.csv (from Revenue Ops)
  • expense_summary.csv (from Financial Control)
  • assumptions.md (from shared)
• Output:
  • 12-month projections (MRR and ARR)
• Saved into:
  • FP&A outputs
  • wherever downstream roles need it

---

4) Reporting & Strategy Manager

• Instruction file: reporting_manager.md
• Inputs:
  • financial summary
  • forecast summary
  • company context
• Output:
  • executive_report.md
  • Structured into five sections
• Saved to:
  • reporting outputs folder
  • shared folder (so the system can reuse it)

---

5) Workflow Coordinator (end-to-end orchestrator)

• Instruction file: full_finance_workflow.md
• Purpose:
  • Runs roles in strict sequence as one workflow from a single high-level objective
• Enforces:
  • cannot skip a role
  • cannot proceed until required outputs are saved
  • cannot “reinvent logic” (prevents skipping/shortcutting)

---

Running the system (example “scenario update”)

Before running the coordinator:

• Update the inputs to mimic real finance changes:
  • switch assumptions.md to an aggressive growth scenario
    • higher growth, lower churn, expanded hiring
  • refresh subscription_data.csv with stronger recent performance
• Use a fresh Claude session in the finance team folder to avoid leftover context.
• Coordinator maps out the workflow, loads role instruction files, checks available datasets, then executes:
  1. Revenue Ops → MRR/customer base/growth
  2. Financial Control → expenses + breakdown
  3. FP&A → projections (MRR/ARR)
  4. Reporting → executive report synthesis
• A final verification confirms every output exists in required locations.

---

Presentation builder (turn analysis into a deck)

• File: PPT_builder.md (sits alongside the coordinator)
• Inputs:
  • pulls outputs produced by the finance team
• Output:
  • Builds a PPTX deck in one pass
  • Also saves a Markdown version for later editing
• Quality steps:
  • converts slides to images
  • runs visual checks for:
    • layout, spacing, contrast, alignment
  • automatically fixes issues before saving
• Result: “presentation-ready” output without a manual cleanup pass.

---

What this tutorial emphasizes (features/benefits)

• Consistency across workflow via shared context + role-scoped instruction files
• Reusability: once structure is built, new data/assumptions update outputs automatically
• Automation of execution, while humans focus on judgment from the executive report (not rebuilding models/pipelines manually)
• Uses folder + instruction-driven execution as a coordinated system, not disconnected chat runs.

---

Main speakers / sources

• Main speaker/source: The creator/instructor of the tutorial (speaking throughout; references “in my community, linked below” and “see you next time”).
• System components referenced: Claude (Claude CoWork) and the tutorial’s provided files like:
  • company_context.md
  • assumptions.md
  • revenue_ops.md
  • fpna_manager.md
  • reporting_manager.md
  • full_finance_workflow.md
  • PPT_builder.md