Getting Good at Claude: A Finance Professional's Playbook

In Part 1 of this series, I introduced Claude- what it is, how it's structured across three modes (Chat, Cowork, and Claude Code), and where it fits in the day-to-day work of a finance professional. If you haven't read that piece yet, I'd encourage you to start there. It provides the foundation for everything that follows.

This article is about getting good at it.

I have a sneaky suspicion that most people who try Claude - or any AI tool - have a similar experience. They type something in, get a response that's decent but not quite right, and walk away thinking the tool isn't ready. Or, they ask it to complete a task, don’t get back exactly what they need, and think “I could’ve done that better myself.”  What they don't realize is that the quality of the output is almost entirely a function of the quality of the input. AI isn't magic. It's a skill. And like any skill, the difference between mediocre results and exceptional ones comes down to technique, practice, and understanding how the tool actually works. 

Pro-tip: you can actually create new skills in Claude that are tailor made for your own situation.

In this article I outline the principles, features, and habits that helped me go from a casual user to using Claude as a genuine part of my workflows. These aren't theoretical tips. They're drawn from real use cases and all the frustration, lessons and triumphs in between.

The Art of the Prompt

A prompt is simply the instruction you give Claude. It can be a question, a request, a set of parameters, or a combination of all three. The quality of your prompt is the single biggest lever you have over the quality of your output. This is true whether you're drafting a variance narrative, building a financial model, or asking Claude to review a contract.

Here's the fundamental principle: Claude doesn't know what you haven't told it. 

It has broad knowledge of accounting standards, financial analysis, business writing, and general best practices. What it doesn't have, unless you provide it, is context about your company, your audience, your preferences, and the specific outcome you're looking for. The more of that context you include in your prompt, the better the result.

Consider the difference between these two requests:

Weak prompt: "Write a variance analysis for this P&L."

Strong prompt: "I'm the CFO of a $40M professional services firm. I've uploaded our March P&L with a comparison to budget. Write a variance analysis narrative for our board of directors. Focus on variances greater than $50K or 10% of budget. Use a professional but accessible tone because our board includes non-financial members. Organize by revenue first, then operating expenses, then below-the-line items. Flag anything that requires a management action item."

The first prompt will give you a generic paragraph. The second will give you something close to what you'd actually send to your board. The difference isn't that Claude is smarter in the second scenario - it's that you gave it enough information to do the job well.

Five Principles for Better Prompts

1. State your role and your audience. 

Claude adjusts its tone, depth, and emphasis based on who is writing and who is reading. "I'm a controller preparing this for our external auditors" produces different output than "I'm a CFO presenting this to a non-financial board." Both are valid but they lead to very different results. Always tell Claude who you are and who the output is for.

2. Be specific about format and structure. 

If you want a table, say so. If you want bullet points, say so. If you want a two-page memo with an executive summary followed by detailed analysis, describe that structure. Claude is remarkably good at following formatting instructions, but it can't read your mind about how you want the output organized. Specificity here saves you the most editing time not to mention frustration!

3. Set boundaries and constraints. 

Tell Claude what to include and what to leave out. "Focus only on operating expenses" is a constraint. "Keep this under 500 words" is a constraint. "Don't include recommendations, just present the data" is a constraint. Without boundaries, Claude will try to be comprehensive, which often means giving you more than you need. Constraints sharpen the output. 

4. Provide examples when possible. 

If you have a previous variance narrative that matches the tone and format you want, paste it in and say "Use this as a style guide." If you have a template for how your board memos are structured, share it. Claude learns from examples within the conversation faster than from abstract descriptions. One good example is worth a paragraph of instructions.

Pro-tip: go to Settings/Cowork/Global Instructions and input rules you want Claude to always follow; these rules and parameters will apply universally across all of the work you give it without you having to restate your preferences every time. 

5. Iterate, don't start over. 

Instead of treating each interaction as a standalone event, keep prompting in the same conversation. Claude maintains context throughout a conversation. If the first response is close but not quite right, tell it what to change: "Make the tone more direct," "Add a section on cash flow impact," "Shorten the executive summary." Iterating on a good first draft is almost always faster than rewriting your prompt from scratch.

Pro-tip: there are limits to how much you can iterate effectively. I found that the longer the conversation goes on, the less sharp the output becomes akin to Claude getting tired of looking through the data over and over. 

Projects: Giving Claude a Memory

I’ve recently started using Projects to organize the various build-outs and conversations I have floating in Claude. A project is a persistent workspace where you can store reference documents, set custom instructions, and have multiple conversations that all share the same context. For example, I have a project dedicated called “Calculators” and another one for P&L variances. Think of it as giving Claude a briefing book about your needs before every conversation starts; all of the relevant materials (artifacts, in Claude lingo) live in the project and are accessible in a single organized folder. 

Here's why this matters: without a project, every new conversation begins from zero. You have to re-explain your company, your reporting structure, your preferences, and your standards every time. With a project, you upload those reference materials once, and Claude has access to them across every conversation within that project.

For a finance team, this is transformational. Consider setting up a project with your chart of accounts, your standard reporting templates, your company's key metrics and how they're defined, your board memo format, and a brief description of your business model and organizational structure. Once those documents are in the project, every conversation you have within it starts with that context already loaded. You don't have to explain that your company is a SaaS business with three product lines and a fiscal year ending in March; Claude already knows.

How to Structure a Finance Project

The most effective finance projects I've built follow a simple structure:

Company context document: A one-to-two page summary that includes your company name, industry, size (revenue, headcount), business model, fiscal year, organizational structure (divisions, departments, cost centers), and any terminology specific to your business. This is the document that prevents Claude from making generic assumptions.

Chart of accounts or reporting structure: Upload your COA or a summary of how your financial statements are organized. This allows Claude to reference your actual account names and groupings when analyzing data or drafting narratives, rather than using generic categories.

Style and formatting preferences: A short document that describes how you want output formatted. Do your board memos have a specific structure? Do your variance narratives follow a particular flow? Do you prefer tables or prose? Capturing these preferences once means Claude applies them consistently.

Key metrics and definitions: If your business tracks specific KPIs (ARR, net revenue retention, EBITDA with specific addbacks, custom profitability metrics) define them here. This ensures Claude uses your definitions, not textbook ones, when referencing these terms.

Custom instructions: Every project allows you to set a system-level instruction that applies to all conversations within it. Use this to establish standing rules: "Always use GAAP terminology," "Default to a professional but conversational tone," "When analyzing variances, focus on items exceeding $25K or 5% of budget."

Setting up a project takes thirty minutes. The time it saves over the following months is exponential.

Markdown Files: Your Secret Weapon

If projects are Claude's briefing book, markdown files are the individual chapters. A markdown file is simply a plain-text document with lightweight formatting (headers, bullet points, bold text) that Claude reads exceptionally well. You don't need to know how to code to create one. If you can write an outline, you can write a markdown file.

Why markdown instead of Word or PDF? Two reasons. First, markdown is plain text, which means Claude processes it faster and more accurately than formatted documents. There's no hidden formatting, no embedded objects, no conversion artifacts, just clean text. Second, markdown files are small, which means you can include more of them in a project without hitting token limits (your usage limits).

Here's a practical example. Suppose your company has a standard set of EBITDA addbacks that you apply every month for covenant compliance. Instead of re-explaining those addbacks in every conversation, you create a markdown file:

# EBITDA Addback Definitions — [Company Name]

## Standard Addbacks
- **One-time legal fees:** Litigation costs outside normal course of business
- **Severance:** Termination-related payments exceeding $10K per occurrence
- **M&A transaction costs:** Legal, advisory, and due diligence fees related to acquisitions
- **Non-cash stock compensation:** All equity-based compensation expense

## Reporting Rules
- All addbacks require supporting documentation
- Threshold for inclusion: $5K minimum per item
- Addbacks must be approved by CFO before inclusion in covenant package

The same approach works for close checklists, revenue recognition policies, intercompany elimination rules, or any other recurring reference material. Pro-tip: anything you find yourself explaining to Claude more than once should become a markdown file in your project.

Features That Matter for Finance

Beyond prompting and projects, several Claude features deserve attention from finance professionals:

File Uploads and Analysis

Claude can read and analyze uploaded documents - PDFs, spreadsheets, images, and text files. This is where the tool becomes immediately practical. Upload a P&L and ask for variance commentary. Upload a lease agreement and ask for a summary of financial terms. Upload a set of invoices and ask Claude to identify duplicates or anomalies. The ability to work directly with your documents, rather than requiring you to copy and paste content, is what makes Claude feel like a working tool rather than a chatbot.

Artifacts

As I mentioned in Part 1, artifacts are interactive outputs that Claude builds inside the conversation - calculators, dashboards, tables, and functional tools. For finance, the applications are compelling: a debt service coverage ratio calculator, a working capital model, a scenario analysis tool, an amortization schedule builder. These aren't static outputs - they're functional applications you can interact with, adjust, and share. I built and published a PE waterfall calculator this way, and it required zero coding knowledge. If you can describe the logic, Claude can build the tool.

Web Search

Claude can search the web during a conversation, which means it can pull current information when you need it. Researching a new accounting standard? Claude can look up the latest guidance. Analyzing a potential vendor? It can pull recent news and financial information. Preparing for a board meeting and need current market context? Web search fills the gaps that Claude's training data might not cover. This is especially useful for fast-moving regulatory or market topics where yesterday's information may already be outdated.

Memory

Claude can remember information across conversations - your preferences, your recurring tasks, the way you like things formatted. Over time, this means Claude gets better at anticipating what you need. It remembers that you prefer tables over prose for variance analysis, that your board memos always start with a cash position summary, or that your company's fiscal year ends in September. Memory turns Claude from a tool you configure every time into one that learns your workflow.

Habits That Separate Power Users

The techniques above are important, but the real difference between casual and effective use comes down to habits - the patterns you build into how you work with Claude every day.

Start with your hardest task. 

Most people begin with easy questions to "test" the tool. That's backwards. The value of Claude shows up most clearly on the tasks that consume the most time or require the most synthesis. Your first real use case should be something that currently takes you an hour or more - a board narrative, a contract review, a financial analysis. That's where you'll see the ROI immediately and build conviction to keep going.

Build a prompt library. 

When you write a prompt that produces great output, save it. Keep a running document - a simple text file or note - of your best prompts, organized by task type. Over time, this becomes your personal playbook. When month-end rolls around, you don't start from scratch - you pull up your proven variance narrative prompt, swap in the current data, and go. Part 3 of this series will give you a head start on this library.

Use Claude as a reviewer, not just a drafter. 

Most people think of AI as a writing tool. It's equally valuable as a reviewing tool. Paste in your own draft of a board memo and ask Claude to identify gaps, suggest improvements, or flag anything that might confuse a non-financial reader. Upload a contract you've already reviewed and ask if you missed anything in the financial terms. Using Claude as a second set of eyes catches things you'd miss on your own and strengthens the work you've already done. 

Set up before you need it. 

Don't wait until the middle of the month-end to start building your project and uploading reference documents. Set up your Claude project during a quiet week. Upload your chart of accounts, your reporting templates, your standard definitions. Write your custom instructions. Do the infrastructure work when you have time, so that when the pressure hits, the tool is ready.

Pro-tip: when building anything new, tell Claude what you need, then ask it to ask you clarifying questions; it will come back with a handful of things you probably haven’t thought of that are important to the task. 

Know when to stop. 

Claude will keep iterating as long as you ask it to. But at some point, you're past the point of diminishing returns. A good rule of thumb: if you've iterated three or four times and the output is close but still not exactly right, it's faster to make the final edits yourself than to keep prompting. Claude gets you most of the way there. Your expertise closes the gap. That's the partnership.

What's Next

In Part 3, I'll share the prompts themselves - specific, ready-to-use templates for common finance and accounting tasks. Each one comes with an explanation of why it's structured the way it is, so you're not just copying a recipe - you're learning the principle behind it.

If Part 1 was about understanding what Claude is, and Part 2 is about how to use it well, Part 3 is the toolkit you take with you.

The finance professionals who will lead in the next five years aren't the ones who avoided AI. They're the ones who learned it early, used it intentionally, and stayed in control of the judgment that matters. The skill isn't hard to build. The best time to start was last month. The second best time is now.