Claude Cowork and Skill Packs: The Practitioner's Guide to AI-Powered Financial Workflows

You may think of AI as a chatbot. You type a question and it gives you an answer. That is useful, but it is limited. The chatbot does not know your data. It does not know your chart of accounts, your vendor list, or your reporting format. Every time you start a new conversation, you are starting from zero.

Claude Cowork works differently. Instead of chatting in a vacuum, you point it at a folder on your computer. That folder becomes its context. It reads your files, understands the structure of your data, and builds deliverables directly on your machine. Think of it less like a chatbot and more like handing a working folder to a sharp new analyst on their first day.

At the top of Claude's interface, you will see three modes. Chat is the traditional back-and-forth conversation. Code is a developer terminal. Cowork is the one I’ll be writing about today and tends to be the most practical for applied business and finance work. It can read spreadsheets, create new ones, and build presentations without you ever opening a terminal or writing a line of code. All you need is:

1. A Claude license (preferably Max or higher if you want to use it for more than a demo)

2. A secure folder on your computer to work out of

3. An understanding of your process and desired output

4. This article (or an understanding of the way Claude Cowork and similar agentic tools work)

Step 1: Load Your Skill Packs

Before touching any data, the first step is to install what Claude calls "skill packs." These are pre-built instruction sets, essentially playbooks, that tell the AI how to approach specific types of work. For example, its finance skill pack contains individual skills for variance analysis, reconciliations, journal entries, and SOX testing. Anthropic hired subject matter experts across accounting, finance, and audit to write these. The skill packs are available for free inside the plugin marketplace.

Here is how to install them:

• Open Claude Cowork and click the Plugins section

• Browse available packs (categories include Sales, Marketing, Data, Finance, and many others)

• Install the Finance plugin pack

• Confirm that skills like variance analysis, reconciliation, and journal entry appear in your workspace

It sounds technical, but in reality these skill packs are just Markdown files (plain text instruction files). You can open them, read exactly what they tell Claude to do, edit them, or write your own from scratch. If your team has a flux analysis format that leadership already expects, you can teach Claude to replicate that exact output. You are not locked into a generic template.

There are also public repositories of community-built skills on GitHub and other platforms. A word of caution there: anything downloaded externally should go through your security team before being used in an enterprise environment. My recommendation is to start with what Anthropic provides natively and build your own from there.

Step 2: Connect Cowork to Your Working Folder

Next, you select a working folder. This is the directory Claude will read from and write into.

Click the folder icon in Cowork, navigate to your working directory, and confirm access. Claude will display a warning that it now has permission to edit files in that folder. That warning is there for a reason: Cowork will edit your files in this folder.

When Cowork is running, any file that you are using as part of your Cowork workflow needs to be closed. It operates on the files in your folder directly, similar to how a colleague would need the file unlocked to make edits. It will edit these files, so treat this the same way you would treat handing files to a new team member. Use a working copy, not your source, when experimenting with Claude Cowork. Do not point it at your SharePoint audit repository. Use a dedicated working folder with version control. If something goes wrong, you want to be able to roll back to the original version.

Within the folder, I had a file labeled "GL Detail", that contained 1053 rows and and 27 columns of various GL detail; specifically, vendor data like costs, debits and credits, internal memos, dates, and names. When you run this exercise, make sure you add any necessary files to the folder you are working out of.

Step 3: Run the Variance Analysis

With the Finance skill pack loaded and the folder selected, I typed a single command using the backslash shortcut:

/variance-analysis

Claude asked a few clarifying questions about scope, format, and level of detail. I responded with something like:

I would like an advanced flux analysis in banking colors based on the GL data. Use reasonable classifications, explore the key drivers, and flag any major vendors causing concerns.

Then I let it run. In roughly three minutes, Claude built a complete Excel workbook from raw GL detail. It wrote the scripts, executed them, and produced the file.

What it produced:

• Executive Summary tab: Narrative explanations of the largest variances, written directly from the GL data

• Department Flux tab: Period-over-period variance by department with banking-style formatting

• Vendor Analysis tab: Concentration of spend by vendor, with flags for risk

• Monthly Trend Analysis tab: Visual trend data across periods

• Variance Analysis Waterfall Q1 to Q2: View of each major category by spend and high-level trends compared to the total spend, bridging Q1 & Q2.

Overall, the outputs are highly impressive; however it is important to note there is a risk with any AI generated explanations. Claude can tell you (with reasonable accuracy, and numbers you can check against sources), the % change or dollar change in an amount (like that outsourced services increased 27.4% quarter over quarter). It cannot tell you why; however, it can speculate (in other words, hallucinate), which may be dangerous. You need to review any justifications it makes, if it does make justifications, to ensure they are accurate. However, for finance use-cases, my experience is that Claude generally sticks to the facts and asks you to add in the color when it doesn’t know the answer.

Step 4: Iterate Like You Would with a Real Team Member

While reviewing the results, you can continue to enhance your workflows, as if you had a staff continuously improving the process. For example, I could type follow-up instructions directly into the same session like the following:

I would like the following improvements. One: Add a tab with reconciliation checks to make sure totals match. Two: Flag reversing entries and confirm they were accounted for. Three: Add an executive summary of strategic recommendations based on the findings. Four: Build a one-slide PowerPoint. Five: Forecast the remaining quarters based on historical trends.

Then Claude can go back to work while I perform my review. In this instance, within a matter of minutes (and before I completed my review), the new tabs were created, along with the PowerPoint summary. This is the part that changes how you think about and work with the tool when compared to a chatbot. You are not submitting a request and waiting. You are working in parallel. You give instructions the way you would talk to a capable team member sitting next to you, and the deliverables build while you move on to your next thought.

This is a completely new workflow that business leaders will need to get comfortable with to maximize their productivity. Instead of stopping at “is the file good enough”, we now have the capacity to improve the files, processes, and ultimately, our insights.

Step 5: Verify the Reconciliation (Trust, but Check)

Now many accountants will ask me at this point: "How are we going to ensure all of this data is correct? Won’t we need more processes, more time to review it, or different controls?" Accuracy and completeness of data is critical to anyone in a finance org, and Claude Cowork can actually help you perform these checks.

Instead of continuing to review the file myself, I asked Claude Cowork to create the review procedures I would normally perform. This includes reconciliations from the summary information to the GL data, tying out total debit and credit balances, and other routine checks. You still need to manually review these outputs, but you can have Cowork create a logical summary table to make this review easier, similar to a staff making an executive tab with all of the checks they performed, linked appropriately.

From this request, Claude built a reconciliation tab that cross-checked GL totals against the variance analysis output. What was especially nice was that Claude flagged the checks that failed. For example, it flagged that a vendor subtotal did not tie because certain GL entries had blank vendor fields, creating a gap between the vendor-level subtotal and the overall GL total. What impressed me most was that Claude did not just flag the failure, but also explained the root cause: entries without vendor data could not be allocated to the vendor-level rollup. That is a process insight rather than just a data exercise, similar to what a good senior accountant would catch and document.

In my review, I noticed that the source numbers were hard coded. I asked Claude to go back and link the reconciliation checks to the source GL data with cell references, which it easily did. It’s important to note many people give up at the point they see the hardcoded numbers; however, all you have to do is ask and it will perform what you want. The tools will not always default to your preferred level of traceability, so you have to be explicit.

Regardless of the tool you build, you need to build enhanced reconciliation and human-in-the-loop checks into every output. If you would not accept a workpaper from your staff without these, do not accept them from Claude.

Step 6: Executive Outputs & Summaries

At my request, Claude also generated a one-slide PowerPoint summarizing the analysis. The formatting needed some polish, but the content was sharp, including identifying real risks from the data: vendor concentration risk, unsustainable quarter-over-quarter cost growth, legal cost escalation, and departmental budget gaps with specific recommendations attached.

For a first pass with zero template input, that is a strong starting point. If you feed it your company's slide master or brand guidelines as part of the skill pack, the output gets materially closer to what your leadership expects to see in a deck.

Step 7: Enhanced Modeling Using Claude Excel - DCF & Forecasts

After the Cowork walkthrough, I switched to the Claude for Excel plugin to demonstrate a different mode of working. This is a separate add-in available from the Microsoft Store (search "Claude by Anthropic for Excel"). Fair warning: the store ratings are low, mostly because the installation process is not intuitive. Once it is running, the tool is excellent.

With the plugin active, I opened the forecast tab from our earlier analysis and typed:

Based on the forecast, build a DCF with a discount rate of 12%. Assume revenue growth stays consistent for five years, then steadies off to half the growth rate. Use the Gordon Growth Model to determine terminal value.

Unlike Cowork, the Excel plugin builds live inside your open workbook. You watch cells populate in real time: growth rates, WACC, terminal year projections, discount factors, and present values. It mirrors exactly what you would build by hand, just faster.

Take note the importance of domain expertise in working with these models. I had to know what a DCF model should look like. I had to know what discount rate was reasonable. I had to understand the Gordon Growth Model and when to apply it. Claude executed the build, but the judgment was mine. This is not a tool that replaces expertise. It amplifies it. A senior leader who understands the underlying finance can move at 10x speed. Someone without that foundation will produce a polished spreadsheet full of bad assumptions or pursuing the wrong goal.

The practical difference between Cowork and the Excel plugin area s follows:

• Cowork operates on your folder. Files you are working on must be closed. It builds in the background and creates new files. Best for multi-file, multi-format workflows that require enhanced context.

• Claude for Excel operates inside an open workbook. You watch it build live, cell by cell. Best for deep, single-workbook analysis where you want full visibility into the process, or with simpler files.

Step 8: Automation - Save Your Work as a Reusable Skill

This is the step that separates people who use AI once from people who build compounding value with it. After the full workflow, I told Claude:

Please build a skill and save it to the folder to replicate this process in the future.

Claude reviewed everything from the session, including the scripts, prompts, files, and outputs. It distilled the entire workflow into a markdown skill file. That file now lives in my working folder. The next time I drop fresh GL data into that same folder and type /flux-analysis, Claude runs the entire process exactly as we built it.

There is a meaningful difference between saving a prompt and saving a skill. A prompt is a single instruction. A skill is a complete set of indexed instructions that Claude loads before you even type your first message. It includes the formatting rules, the reconciliation checks, the output structure, and the model assumptions. Everything you refined during the session becomes the default behavior for next time.

Once you have that skill saved, two newer features start to make the picture even more compelling. Claude now supports scheduled tasks: you can tell it to run your flux analysis on the first of every month, against whatever files are in your working folder, and email you when it is complete. Your computer needs to stay awake for it to run, but the functionality is live and improving fast. For teams with repeatable monthly close steps, this is worth experimenting with today.

There is also a mobile dispatch feature that lets you text instructions to your Claude agent from your phone. Imagine you are in a meeting and you realize you need an updated analysis before you are back at your desk. You can send a message to kick it off remotely. Your machine needs to be on and connected, and this technology is still in its early stages, but the use case is not only obvious, but is already driving a real hardware trend: MacBook Airs and lightweight laptops are selling out at enterprise scale because teams are buying dedicated machines to run AI agents continuously in the background. We are quickly moving to a world where all employees have two computers: One for your daily work, one running your AI assistant. It sounds excessive until you see the output, which is exponential.

Summary: What Separates the Teams That Adopt from the Teams That Stall

I have worked with enough finance organizations at this point to see the patterns clearly. The teams that get value from these tools share a few common traits:

• They start with their existing workflow: If you already have a flux format leadership expects, teach Claude to replicate it. Do not start from a generic template. Reverse-engineer your best deliverable.

• They treat Claude like a capable staff member: It builds fast and formats well. It does not exercise judgment. Your job shifts from preparer to reviewer, and that review step is non-negotiable.

• They build reconciliation checks into every output: GL totals, debit and credit balances, vendor subtotals, and linked formulas versus hardcoded values. If you would not accept a workpaper from staff without these checks, do not accept them from Claude.

• They save and share their skills: Every session that produces a good output should end with "build a skill for this." That markdown file is your automation. It turns a 45-minute manual process into a single command.

• They use the right model for the work: Opus for complex, multi-step builds. Sonnet for fast answers. Haiku for quick lookups. Using a basic model for complex financial analysis produces polished-looking output with subpar accuracy.

• They take security seriously: Use embedded skill packs or skills you build yourself. Be cautious with third-party downloads from public repositories. Anything external should go through your security team first.

Conclusion: The Shift I Believe Most Leaders Are Missing

There is a pattern I keep seeing in conversations with CFOs, CAOs, and VPs of Finance. The instinct is to treat AI adoption like a technology implementation: build a roadmap, hire a vendor, run a pilot, and scale over 18 months. That playbook made sense for ERP implementations and BI platform rollouts. It does not apply here.

The organizations moving fastest are the ones investing in upskilling their existing people. Turning preparers into reviewers. Turning analysts into decision-makers. Treating AI competency as a core professional skill rather than an IT project on next quarter's roadmap.

The evidence is starting to show up in hiring strategy. In February, IBM announced it would triple its U.S. entry-level hiring in 2026, even for roles where AI can automate significant portions of the work. The rationale was not sentimental. IBM’s CHRO, Nickle LaMoreaux, framed it explicitly as a leadership pipeline issue: companies that slash junior hiring for short-term AI efficiency will face a shortage of mid-level managers and senior leaders three to five years from now. As she put it at Charter’s Leading with AI Summit, “The companies three to five years from now that are going to be the most successful are those companies that doubled down on entry-level hiring in this environment.” The jobs themselves are being redesigned. Software engineers spend less time on routine coding and more on customer interaction. HR staff spend more time intervening with chatbots than answering every question manually. The role changes, but it does not disappear.

A few weeks later, Reddit CEO Steve Huffman made a different but equally telling argument. He said Reddit would “go heavy on new grads, because they’re so much more AI native.” His reasoning was not about preserving a pipeline. It was about capability. “The kids coming out of college right now learned how to program with AI,” Huffman said. He noted that older workers tend to resist automating their craft, while younger hires do not carry that friction.

Two different companies but the same conclusion: the organizations investing in people right now, whether to build a leadership bench or to hire AI-fluent talent, are the ones positioning themselves to pull ahead. It's easy to mirror this strategy: invest in training your teams to be AI-ready, or in AI-ready resources. You need your people trained to take advantage of this new technology, and you need them practicing with real data on real workflows.

The tools are ready. The question is whether your team is.

Tool Stack

What I Used in This Walkthrough

• Claude Cowork (Opus 4.6): The primary workspace. Folder-based AI environment that reads your data, executes skill packs, builds Excel files and PowerPoints, and runs Python scripts in the background. Best for multi-file, multi-format business workflows.

• Claude for Excel Plugin: Excel add-in from the Microsoft Store. Builds live inside your open workbook. Used here to construct a DCF model from forecast data in real time. Installation is finicky. Performance once running is excellent.

• Finance Skill Pack (Built-in): Pre-built instruction sets for variance analysis, reconciliation, journal entries, SOX testing, and more. Installed in two clicks from the Cowork plugin marketplace. Free with any Claude subscription.

Products & Services I Recommend

Teams Building Advanced AI Solutions for Finance

If your organization is past the experimentation phase and needs domain-specific implementation, these are teams I would recommend talking to. If you are interested in training yourself or your organization, or an introduction to these teams, don’t hesitate to reach out to me.

Gaapsavvy - Angela Liu: Leader of the largest active community of enterprise accounting practitioners. Regularly discusses the intersection of cutting edge topics in AI, Technical Accounting, and Finance. We also co-host a podcast around these topics, where we covered building an AI Finance App using Antigravity which you can watch here: https://youtu.be/om3a-NLukHU

 TAbot - Harton Wong: Founded by a former Google AI engineer. Specializes in technical accounting and deal intelligence use cases for enterprise finance teams.

Lumera - Sowmya Ranganathan: Founded by a ex- OpenAI Controller. A live develop and deploy website for AI applications. It’s the first tool I’ve seen where you could build a new tool, workflow, or app for a business use case and deploy it at the same time with SOC2 compliance. 

That is it for this week. If this was useful, share it with someone on your team who is evaluating these tools or trying to figure out where to start.

If you or your team needs hands-on AI training tailored to your actual workflows and use cases, send me a message. I am happy to talk through what makes sense for your organization.

See you next week.

Devon