Claude Finance Agents: The 23-Agent Financial Modelling Pack
Claude accelerates financial modeling when the assignment is split into evidence, assumptions, schedules, integration, challenge, and communication. It should not be asked to source data, choose assumptions, build, audit itself, and write the recommendation in one prompt. This guide provides the operating manual; the companion pack provides 23 bounded Markdown agents with explicit jobs, controls, and handoffs. They work natively as project subagents in Claude Code. In Claude Projects, upload selected files and use them as reusable role instructions rather than native subagents.
The six-stage workflow
A controlled model is a sequence of artifacts and approvals, not one impressive response. Predictable build stages are workflows; agents are used as bounded specialists. The main Claude conversation holds the work order, routes the specialists, freezes versions, records unresolved items, and stops at each human gate. This separation prevents a builder from quietly choosing the assumptions it needs or an auditor from reviewing only its own narration. Every handoff identifies the model version, assumption version, artifacts, controls run, locked items, and unresolved questions. That creates a traceable path from the first source to the final recommendation.
Define the workbook, conventions, accounting treatments, interfaces, timeline, and review gates before formula work begins.
Use narrow owners for revenue, costs, schedules, statements, and integration, all working from one approved assumption version.
Recast history, define the driver framework, and maintain a versioned assumption register that builders are not allowed to rewrite.
Apply DCF, comparable companies, M&A, or LBO methods only after the integrated base model passes its first independent audit.
Challenge circularity, scenarios, sensitivities, formula behavior, evidence coverage, and the complete workbook end state.
Extract only approved outputs, then create a repeatable reporting package and an Oria-ready decision story.
Prompt Claude with a work order
“Build me a DCF” leaves the date, evidence, accounting basis, units, forecast ownership, terminal method, tests, and permissions unstated. A work order makes the assignment reviewable and gives the coordinator explicit stop conditions.
<objective>Build a review-ready three-statement model and DCF.</objective>
<decision_use>Internal investment committee discussion.</decision_use>
<as_of_date>2026-06-30</as_of_date>
<sources>Filed statements first; approved data feed second.</sources>
<conventions>USD millions; FY2022-FY2030; no hidden plugs.</conventions>
<assumptions>Use approved register v1.3; label new requests [UNRESOLVED].</assumptions>
<checks>Balance sheet, cash bridge, roll-forwards, provenance, sensitivities.</checks>
<stop_conditions>Missing primary statements; failed balance; unapproved terminal value.</stop_conditions>Five habits improve every consequential run: specify the decision and as-of date; name the source hierarchy; state periods, units, and accounting basis; define the deliverables and deterministic tests; and expose uncertainty as an assumption or unresolved item. For a broader foundation, see the ultimate guide to Claude for finance.
Choose and sequence the 23 agents
Start with Set Up, then let Budgeting & Forecasting prepare a driver snapshot for approval. Build schedules, assemble statements, integrate, and run Model Checks before Value. After scenarios and sensitivities, run Model Checks again. Outputs & Automation begins only after final approval, then prepares the handoff to Oria. Schedule builders may work in parallel only after architecture, interfaces, and assumptions are frozen.
The six stage labels are the capability backbone, but the safe execution route deliberately resolves forward assumptions before builders use them. Revenue, costs, working capital, capex, debt, and tax can then advance in parallel against fixed interfaces. The three statements should be assembled in sequence before agent 04 integrates them. Agent 22 stays independent of the builders and runs at least twice: once before valuation and again after material scenario or valuation changes.
Set Up
Model Setup
Defines structure, interfaces, sequence, scenarios, controls, and gates.
Use when: New, extended, repurposed, or opaque models.
Excel Best Practices
Sets formula, input, sign, unit, link, and version conventions.
Use when: Before Build or when workbook standards are inconsistent.
Accounting Refresher
Creates the source ledger and maps accounting treatments into model lines.
Use when: Reporting changes or judgments affect comparability.
Build
Three-Statement Model
Integrates approved schedules, statements, and cash and debt links.
Use when: After statement handoffs and approved refreshes.
Revenue Drivers
Builds segment revenue from approved operating drivers.
Use when: Operating drivers can replace a top-line plug.
Operating Costs
Builds COGS, headcount, opex, margins, and EBIT.
Use when: Fixed-variable or operating-leverage logic is needed.
Working Capital
Builds operating balances, drivers, and cash effects.
Use when: Cash conversion matters.
Capex & D&A
Builds capex, PP&E, D&A, and intangible roll-forwards.
Use when: Asset intensity affects results.
Debt Schedule
Builds debt, interest, liquidity, maturities, and covenants.
Use when: A model is financed or liquidity-sensitive.
Tax Schedule
Builds current, deferred, effective-rate, and cash tax.
Use when: Tax timing or attributes matter.
Income Statement
Assembles earnings and diluted-share and EPS outputs.
Use when: Operating, debt, and tax handoffs are ready.
Balance Sheet
Assembles assets, liabilities, equity, and retained earnings.
Use when: Schedules and opening balances reconcile.
Cash Flow Statement
Assembles cash flows and reconciles ending cash.
Use when: Statement movements are traceable.
Forecast
Budgeting & Forecasting
Calibrates drivers and controls sourced, owned, versioned assumptions.
Use when: Base forecasts, budgets, updates, and cases.
Value
DCF Valuation
Builds discounting, terminal value, and the EV-to-equity bridge.
Use when: Audited cash flow and approved valuation assumptions exist.
Comparable Companies
Normalizes peers and applies consistent trading multiples.
Use when: Approved peers and current sourced data exist.
M&A Model
Structures sources and uses, purchase accounting, synergies, and accretion or dilution.
Use when: An acquisition case is authorized.
LBO Model
Structures entry, exit, debt paydown, and sponsor MOIC and IRR.
Use when: Leveraged-return analysis is authorized.
Stress-Test
Model Circularity
Implements approved circularity and tests convergence and switches.
Use when: Genuine cash, debt, or interest feedback loops exist.
Scenario Analysis
Turns risks into driver-based cases and reverse stresses.
Use when: Testing resilience, liquidity, covenants, or thesis risk.
Sensitivity Analysis
Builds sensitivities and tests base ties and direction.
Use when: Uncertain inputs drive the outcome.
Model Checks
Audits formulas, links, provenance, controls, and workbook end state.
Use when: After integration or changes and before communication.
Communicate
Outputs & Automation
Packages approved outputs and an Oria-ready decision brief.
Use when: Only after qualified Gate 3 approval.
Install all 23 if the project spans the full model, but route only the smallest set needed for the current work order. Each file in the companion pack contains a bounded method, evidence rules, stop conditions, output contract, and standard handoff. The main conversation remains the coordinator; the agents are leaf specialists and do not delegate to one another.
Turn repeated methods into Skills
A work order briefs one consequential run. An agent owns one bounded role and output contract. A Skill records a repeatable firm method, such as source hierarchy, model conventions, valuation policy, model audit, or slide storyline. Record a Skill when the method recurs, its exceptions can be described, and quality improves when the same checks are applied. A useful Skill defines source precedence, period and unit normalization, judgment points, escalation triggers, deterministic tests, and what done means. Test normal and failure cases, remove engagement-specific or sensitive material, and revise the method when reviewers find a recurring defect. Screen-recorded Skill creation is available only on supported Claude plans and surfaces; a written Skill remains the portable option. The institutional-grade finance playbook explains the wider operating model.
Controls before communication
A model is ready for decision use only when the workbook end state passes evidence and behavior checks: consistent formulas, no hidden plugs, balanced statements, complete roll-forwards, linked scenario drivers, sensitivity base-cell reconciliation, directional behavior, source coverage, and a versioned assumptions register. A balanced balance sheet is one control, not proof that the economics, periods, signs, or sources are correct. Reviewers should inspect the workbook and deterministic tests, not accept an agent's statement that its task is complete.
Gate 1A
Approve the canonical source ledger, history, architecture, conventions, and accounting treatments.
Gate 1B
Approve the versioned assumption snapshot before builders create forecast formulas.
Gate 2
Audit and approve the integrated base model before any valuation agent begins.
Gate 3
Approve valuation, scenario behavior, intended use, confidentiality, and distribution.
No agent executes trades, posts ledger entries, changes a production system of record, or distributes external work without explicit authorization and human review.
From model to Oria deck
Agent 23 extracts approved outputs and writes a page-by-page decision spine: question, historical engine, base case, value drivers, downside, liquidity, valuation range, and decisions. Each slide specification contains an answer-first headline, evidence, the chart or table source, implication, and footnote. Oria is AI for complex slides. It turns the approved analytical spine into polished, editable slides without rebuilding the model inside the deck. The model remains the source of truth; Oria makes its decision logic usable in the room. When an assumption changes, the team refreshes and rechecks the model first, then updates the linked story so the deck never drifts from the approved analysis.
Frequently asked questions
Are these autonomous financial advisers?
No. They create draft analyst work product for review.
Where do the files work?
Natively as Markdown subagents in Claude Code; as reusable instructions in Projects; adapted definitions in API or managed-agent systems.
Can I install all 23 at once?
Yes, but route only the specialists needed for the work order.
Does Oria build the model?
The pack governs model work in Claude; Oria turns approved analysis into complex slides.