25 Claude Prompts for Finance Teams: Real Workflows for Cowork, Code, and FP&A

Every prompt in this library was used, demonstrated, or described in a live CFO Connect build session — in front of a combined audience of over 500 finance leaders. These are not theoretical examples. They are the actual prompts that built a working intercompany reconciliation file, a margin analysis report with AI-generated commentary, and a full three-way FP&A application with live Xero, QuickBooks, and Sage integrations.

Three sessions. Three experts. Twenty-five prompts you can copy, adapt, and use today.

Key Takeaways

• The single most important prompting skill in finance is using Claude Chat to write your Claude Cowork and Code prompts — not going straight to the tool

• Prompt quality determines output quality; the structure matters more than the length

• Every finance workflow prompt needs four elements: input format, output structure, calculation logic, and exception handling

• Screenshots beat paragraphs for communicating UI and layout changes in Claude Code

• A backlog file inside Claude Code replaces a product roadmap for long-running build projects

• These prompts were tested live on screen, reviewed by finance professionals, and used with real client data

What makes these prompts different from every other list?

These were tested live, on screen, in front of hundreds of CFOs — and reviewed by finance professionals.

Most Claude prompt libraries for finance are written by people who have never closed a month, reconciled an intercompany balance, or built a three-way model. The prompts in this article come directly from three CFO Connect live-build sessions:

• Christian Sanford, co-founder of Quantfy, which runs fully outsourced finance departments for companies between $5m and $200m in revenue — read the full Session 1 recap

• Sherilyn Kamga, founder of SmartPlans, which delivers AI-powered automation and strategic financial guidance to growth-stage businesses — read the full Session 2 recap

• Kevin Steel, fractional CFO who built a full production-grade FP&A application in Claude Code with zero prior coding experience in six weeks — read the full Session 3 recap

Each session was a live build. No pre-recorded outputs. No edited demos. The prompts below are what actually worked.

Section 1: Meta-Prompting — Use Claude to Write Your Claude Prompts

The highest-leverage prompting skill is not writing better prompts yourself. It is using Claude Chat to write them for you.

This was the single most consistent insight across all three sessions. Christian Sanford described it directly during Session 1:

I go to the chat function. I'm like, okay, I'm about to prompt Cowork to do this reconciliation. But if I was just to go into Cowork and say, 'Can you do this reconciliation?' — I don't think it would give me the output that I want. So I would go to Chat, and I would say, 'Help me create a prompt for Cowork to do this reconciliation.' And then it'll write a really thoughtful prompt in its own language effectively that you can then copy and paste into Cowork.

Sherilyn runs a dedicated Claude Chat project set up as an automation architect persona that reviews every prompt before she sends it to Cowork or Code:

I use an automation architect that will review my architecture and my prompt before I send it to Claude Code.

Prompt 1: The meta-prompt for Cowork

[Prompt] Help me create a prompt for Cowork to do this reconciliation. [Paste your reconciliation brief here.]

Use this in Claude Chat before opening Cowork. Paste the output directly into Cowork as your prompt.

Prompt 2: The automation architect persona

[Prompt] You are an automation architect. Before I send any prompt to Claude Code or Claude Cowork, I want you to review my architecture, identify any gaps or ambiguities, and rewrite it as a strong, precise prompt that includes the input format, expected output, and exception handling. Review this prompt: [paste your draft].

Set this up as a standing project in Claude Chat. Use it every time before running anything in Cowork or Code.

Prompt 3: The case study learning prompt

Christian revealed this was how every example in his live session was generated:

[Prompt] I'm preparing to learn how to automate [specific workflow] in Claude. Give me three practical use cases. For each one, provide: the exact prompt to use, realistic demo data I can test with, and the expected output. Walk me through each one step by step.

Use this to learn any new workflow by having Claude generate both the teaching material and the working example at the same time.

Prompt 4: The prompt revision prompt

Sherilyn, in Session 2: "It's okay to spend time prompting, but just don't do it by yourself."

[Prompt] Here is a prompt I've drafted for [task]. Revise it to make it stronger. Make sure it includes: a clear description of the input data, the exact output format required, the logic or calculations involved, and instructions for handling missing or inconsistent data. My draft: [paste your prompt]

Section 2: Financial Model Audit Prompts (Excel)

Use the Claude Excel add-in to find errors, add explanatory comments, and check balance sheet integrity — replacing the analyst-to-associate review cycle.

Christian demonstrated this live in Session 1 with a broken three-statement model. The balance sheet was not tying out:

"Find all of the errors in this document, highlight yellow, and then put note comments on each of the errors as to what they are. You can see it's highlighted all of the errors, put comments on each one and why there are errors."

Prompt 5: The model audit prompt

[Prompt] Find all errors in this document. Highlight each error cell in yellow and add a comment on the cell explaining what the error is and why it exists.

Claude will scan every tab, identify formula errors, hard-codes, and broken references, highlight them, and add explanatory comments — without you locating them manually.

Prompt 6: The checking functions requirement

Christian was explicit: "When you're doing your prompt, say: 'Make sure to include checking functions.' Add that level of responsibility into the actual workflow that's coming from AI. Because if you're having to manually do all of those checks, it kind of defeats the value."

[Prompt] [Your main prompt here.] Additionally, include a reconciliation tab with checking functions that verify the key outputs. Flag any cells or calculations that do not check out. Add a source of truth citation to each output row.

Add this instruction to every prompt that produces a financial file. It turns Claude's output into a self-reviewing package.

Prompt 7: The balance sheet tie-out prompt

[Prompt] This three-statement model has a balance sheet that does not tie out. Work through every tab systematically. Identify where the discrepancy originates, explain what is causing it, fix it, and confirm the balance sheet checks out. Do not change any assumptions or hard-coded inputs without flagging it first.

Section 3: Intercompany Reconciliation and Close Prompts

Use Claude Cowork to compress a multi-day intercompany allocation process into a structured, controller-ready file.

Christian walked through a live build in Session 1: four entities, three currencies, expenses allocated by percentage of revenue, headcount, and direct charges. His method: draft the prompt in Chat first, then transfer to Cowork with the invoice data attached.

Prompt 8: The intercompany reconciliation starter (Chat — first step)

[Prompt] I'm looking to create a reconciliation file for intercompany expenses. Here is our shared services invoice: [paste invoice data]. We have [number] entities in [number] currencies. Allocation methodologies vary by expense line: some are percentage of revenue, some are headcount-based, some are direct charges. Help me build a prompt I can use in Cowork to automate this process and produce a journal entry upload file.

Run this in Chat. Use the output as your Cowork prompt.

Prompt 9: The journal entry output specification

[Prompt] The output should include: a journal entry upload sheet with entity names, GL account codes, debit and credit amounts, currency, line item descriptions, and allocation methodology per line. Include a separate checking tab that verifies debits equal credits per entity and flags any discrepancies. Cite the data source for each output row.

Add this to any reconciliation prompt. It makes the file immediately reviewable by a controller without a separate preparation step.

Prompt 10: The missing data routing prompt (Cowork)

Sherilyn demonstrated this live in Session 2 during her margin analysis workflow. When the pipeline encountered an unrecognised client, it did not stop:

It didn't stop. It just ran with what it had, and it flagged what was missing. This isn't a process that will break when something is missing. It tells you what's missing, and it keeps on moving.

[Prompt] If any required input is missing or a record cannot be matched, do not stop the workflow. Flag what is missing, continue processing everything else, and send a Slack message to [name or role] explaining exactly what information is needed and why. Once the reply is received, update the relevant mapping and regenerate the affected output.

This turns a fragile workflow into a resilient one. The loop closes itself.

Section 4: Invoice Processing and Workflow Automation Prompts

The same prompt structure works whether you build the workflow in Zapier or Claude Code. The difference is execution speed versus flexibility.

Sherilyn built the same invoice processing workflow two ways in Session 2 — Zapier first, then Claude Code. Her core insight:

Prompting is the universal skill that you want to develop. Every time AI is in the loop, the quality of your outputs starts with the quality of your prompt. Give it relevant context, work piece by piece, don't ask too many things at once.

Prompt 11: The invoice extraction prompt

[Prompt] Extract the following fields from this invoice: vendor name, vendor email, invoice number, invoice date, due date, line items (description, quantity, unit price, total), subtotal, tax amount, and total amount due. Return the output as structured data. If any field is missing or illegible, flag it with a null value and a note explaining what is missing.

Works as a Claude API step inside Zapier, or as the core extraction logic in a Claude Code script.

Prompt 12: The vendor validation prompt

[Prompt] Check whether the sender email address [email] appears in this approved vendor list: [paste list]. If it matches, return "approved" and the vendor name. If it does not match, return "rejected" and stop. If it is a close but inexact match, flag it for manual review with an explanation of the discrepancy.

Prompt 13: The workflow architecture prompt

[Prompt] I want to automate this finance workflow: [describe the process — what triggers it, what data comes in, what the output should be, who reviews it, and what happens when something is missing]. Before building anything, map this as a step-by-step workflow. Identify any gaps, edge cases, or data dependencies I should resolve before automating.

Run this in Chat before opening Cowork or Code. It surfaces problems before they become debugging sessions.

Section 5: Reporting and Commentary Prompts

Use Claude Code to transform raw data into formatted, shareable reports with AI-written commentary — in a single automated run.

Sherilyn demonstrated this live in Session 2: a flat CSV export from a time tracker went in; a formatted HTML report with margin analysis by client, month-on-month variance, and AI-generated narrative came out:

"You have that AI-generated commentary explaining why the margin moved. You also have different tabs to see month-over-month variance and team utilisation, and basically any KPI you need based on the CSV."

Prompt 14: The margin analysis report prompt

[Prompt] Read this CSV file: [attach file]. Calculate gross margin by client and by team member. Identify the top three margin movers versus last month and explain the variance. Generate a formatted HTML report with the following sections: executive summary, margin by client table, margin by team member table, month-on-month variance analysis, and a written commentary explaining the key movements. Flag any client or category that could not be matched and note what data is missing.

Prompt 15: The AI commentary generation prompt

[Prompt] Based on this financial data: [attach or paste], write a one-page management commentary covering: revenue performance versus prior period and budget, gross margin movements and key drivers, EBITDA bridge, and the top two or three items requiring management attention. Write in a professional tone suitable for a board pack. If a comparison is not possible due to missing data, say so explicitly — do not estimate.

Prompt 16: The iterative report-building prompt (Claude Code)

Kevin built his entire reporting suite using this pattern in Session 3:

[Prompt] Add a new report section called [section name]. It should display [specific metrics or KPIs]. The data comes from [source]. Show it as [chart type, table, or narrative]. It should appear in both the individual company view and the consolidated group view. Match the existing report styling exactly.

Section 6: Building Finance Apps with Claude Code

You do not need coding experience to build a production-grade finance application in Claude Code. You need the right starting prompt, an iterative method, and the discipline to act as QA at every step.

Kevin Steele built Inflective Intelligence — a full three-way FP&A application now in live use with paying clients — starting with a single prompt. He had zero prior coding experience. Read the full account of how he did it in Session 3.

"A lot of people, you might see tips online. Like, context is everything. Give Claude this big massive context. Here, I'll spend two months designing my grand vision of this, then I'll paste it in Claude, and Claude will build the whole thing. I didn't take that approach. I literally went step by step."

Prompt 17: Kevin's exact first prompt

This is verbatim. It started a six-week build:

[Prompt] I'm looking to build an app which has a profit and loss account with drivers where I can update to populate. Eventually, I want to move on to balance sheet and then cash flow, so a full three way forecast. But I want to take it step by step and start with the P&L.

Claude will recommend a tech stack and ask clarifying questions. Answer them. Approve the stack. Type "continue." That is the entire first session.

Prompt 18: Answering Claude's clarifying questions

When Kevin answered Claude's follow-up questions after Prompt 17, these were his exact responses:

[Prompt] Forecast horizon: 12 months to start. Driver types: Start with the basic ones you recommend. Actuals vs forecast: Yes, include actuals — they are good for testing. Business type: Services business. Use demo data. A 4 million revenue marketing agency.

Specific answers produce better outputs than vague ones. A named business type and revenue figure gives Claude enough to generate realistic seed data immediately.

Prompt 19: The screenshot iteration method

Kevin described this as one of his most effective techniques throughout the Session 3 build:

Take a screenshot of what's there, and then highlight. I would circle using the snipping tool. I would paste that into Claude Code, and then I would say, 'Hey, April and May — I don't like the look of this.' Screenshots and pasting them into Claude is a really, really good way because it's a lot easier to show Claude an image than it is to type out a long three, four paragraph trying to explain.

[Prompt] [Paste screenshot with the issue circled.] In the area I have highlighted: [describe the specific issue — e.g., "the font is too large", "these figures should display as percentages", "this column should not show negative values"]. Please fix this and confirm what was changed.

Prompt 20: The backlog file setup

[Prompt] Create a backlog file in this project. Every time I mention a new feature, improvement, or idea during our sessions, add it to the backlog automatically with a brief description and an estimated complexity: low, medium, or high. When I ask to review the backlog, show me the full list sorted by what you think is the highest priority based on our recent work together, with a rough time estimate for each item.

Kevin: "That's been one of the most helpful things for me — having that backlog, and Claude cross-referencing and going, 'Oh actually, we're working on this — do you know what, we could maybe knock off that backlog item at the same time because it's quite interconnected."

Prompt 21: The backlog review and planning prompt

[Prompt] Show me the current backlog. Based on our work over the last few sessions, what do you think should be the top priority? Are there any backlog items that could be completed alongside the feature we are currently building because they are closely related?

Prompt 22: The security check prompt

Kevin ran this before every new technical component was built in Session 3:

[Prompt] Before we build this feature, is it safe and secure for what we are building? What are the edge case risks? What should I flag or be aware of? Do not overwrite any existing files or packages without telling me first.

Section 7: QA and Accuracy Prompts

Claude handles the mechanics. You handle the financial judgment. These prompts build verification into every output.

Kevin spent seven hours getting the prepayments module right during Session 3 — not because Claude was incapable, but because the accounting treatment required a finance professional to verify each iteration:

You know the impact on corporation tax. You know the impact on sales tax. You know the impact on retained earnings. You know the impact on cash. So you can just manually verify with the demo company. Here's what it was before. I make the change. Is this what I expect to happen?

Prompt 23: The three-way model verification prompt

[Prompt] I am going to test this scenario: [describe the transaction — e.g., "a prepaid expense of £12,000 covering 12 months, with cash paid over 10 months, starting April"]. Before I enter the data, tell me exactly what I should expect to see in the P&L, balance sheet, and cash flow for each affected month. I will compare your prediction against the actual output. If they do not match, we will identify where the logic has broken down.

Force a prediction before you test. Discrepancies between Claude's predicted output and actual output pinpoint exactly where the logic is wrong.

Prompt 24: The human-in-the-loop output structure prompt

Christian's framework from Session 1 for every automated file sent to a controller:

You wanna have the package structured in a way where you're getting the output, but you also have the supporting documentation, the sources of truth, and also have the checking functions.

[Prompt] Structure the output as follows: a main results tab, a separate checking tab that verifies the key totals, and a sources tab that lists the data source and methodology for each output row. The file should be self-explanatory to a controller who has not seen the underlying data. Highlight any cells requiring manual review in amber.

Prompt 25: The workflow blind-spot check prompt

[Prompt] Before we finalise this workflow, identify the three most likely failure points. For each one, tell me: what would cause it to fail, what the output would look like if it failed silently, and what checking function or safeguard I should add to catch it. Then add those safeguards to the workflow.

The prompting framework behind every prompt in this library

Sherilyn described the structure underpinning all high-quality finance prompts in Session 2:

An important requirement is that you need to describe your data structure and your business logic as well as the expected outputs with real precision. You remain the architect here.

Every prompt above follows this four-part structure:

1. Context — what the data is, where it comes from, what business process it belongs to

2. Input format — the exact structure of the data you are providing

3. Output specification — the exact format, fields, and structure you need back

4. Exception handling — what Claude should do when data is missing, inconsistent, or ambiguous

When any of these four elements is missing, output quality drops. When all four are present, Claude consistently produces outputs that are review-ready rather than rework-required.

FAQ: Claude Prompts for Finance Teams

What are the best Claude Cowork prompts for finance? The best Cowork prompts follow a four-part structure: context, input format, output specification, and exception handling. Always write your Cowork prompt in Claude Chat first — use Prompt 1 or 2 from this library to generate a stronger version before pasting it into Cowork. See the full Cowork workflow breakdown in our Session 2 recap.

What are the best Claude Code prompts for FP&A? Start with a single, specific feature rather than a complete application. Kevin Steele's exact first prompt (Prompt 17) produced a working P&L with drivers in 15 minutes. Build from there iteratively. Use screenshots (Prompt 19) for UI changes and a backlog file (Prompt 20) to manage ongoing development. Read the full Session 3 build account for the complete six-week playbook.

How do I stop Claude from making errors in financial outputs? Add checking functions and exception handling to every prompt (Prompt 6). Ask Claude to predict the expected output before running it, then verify against the actual result (Prompt 23). Treat every output as a version one that requires QA, not a final deliverable.

Should I use Claude Chat, Cowork, or Code for finance prompts? Start every prompt in Claude Chat, even if the task will run in Cowork or Code. Chat is the best environment for refining and testing prompts before deploying them. Use Cowork for multi-step autonomous workflows. Use Code when you need a custom tool, script, or application. See the full breakdown of all three tools in our Session 1 recap.

How long does it take to write a good finance prompt? Sherilyn's answer: "I might spend more time prompting than actually building the script with Claude Code. Because if the prompt is strong and it already has all the requirements, the output is well-detailed, then Claude Code is gonna get it. That part might take you a few minutes only." Time spent on a prompt is leverage, not overhead.

Can I build a finance application in Claude Code with no coding experience? Yes. Kevin Steele did it in six weeks using the iterative method in this article. His first prompt is Prompt 17 above. The application is now in live use with paying clients. Read the full story here.

What is the backlog file method in Claude Code? A backlog file is a running list of features and ideas maintained inside your Claude Code project, set up with Prompt 20. Claude updates it automatically as you mention new ideas, cross-references it during related work, and helps you prioritise. Kevin credits it as one of the two most useful practices throughout his entire build.

Closing thought: the prompt is the skill

Each of the three experts in this series came from a different angle. Christian runs outsourced finance departments. Sherilyn builds automation systems for growth-stage businesses. Kevin is a fractional CFO who built his own software from scratch.

All three said the same thing about prompting.

Christian: "The biggest guardrail is prompting still, and the biggest solution is to just use the tool to prompt."

Sherilyn: "Prompting is the universal skill. Every time AI is in the loop, the quality of your outputs starts with the quality of your prompt."

Kevin: "I go step by step. I'm doing QA on every single step. My development speed just got faster and faster because I started understanding things."

The finance professionals who move fastest with Claude are not the ones who read the most about it. They are the ones who write the first prompt, check the output against their own financial knowledge, and iterate.

The 25 prompts above are a starting point. Adapt every one of them to your own data, your own process, and your own team.

Want to see these prompts in action across full live builds?

• Session 1: Claude for Finance Teams — the complete toolkit overview, with Christian Sanford

• Session 2: Claude Code, Cowork, and Zapier for finance automation, with Sherilyn Kamga

• Session 3: How a fractional CFO built a full FP&A app in Claude Code in 6 weeks, with Kevin Steele