Transcription
Claude released a new feature called Claude Design, but most finance professionals will completely underestimate what it actually changes. And right now, most finance work still breaks in the same place. You can get analysis out of Claude, but the layer between the raw data and a meeting still has to be built manually. The structuring, the formatting, the narrative, the actual outputs you present. Now, that's where most of the time goes up until now, but Claude Design compresses that entire layer.
So, as a quant and finance automation developer, I'm going to show you three full workflows built inside Claude Design from raw financial data. These are going to be a variance analysis tool that runs as a reusable application, a board presentation generated end-to-end with a working narrative, and a multi-tab finance dashboard that mirrors how real management reporting packs are structured. And each one is a real artifact that you can export and use in a meeting the same day.
Now, if you want the prompts and the data sets from this video, they're in my free community, which is linked in the description below. Now, let's get into it.
Now, before I start building, I want to be clear about what Claude Design actually is because it's different from Claude Co-work and different from the standard chat. Co-work is the agentic execution mode where Claude plans and runs multi-step tasks across files and tools. Design is a visual workspace where Claude builds artifacts like slide decks or prototypes and interactive tools beside the conversation in real time. So, you describe what you want, and then you see it appear, and then you refine it through conversation.
Now, where this matters for finance is the middle layer that we talked about. That middle layer of work where it's not transactional processing, it's not the system of record, but everything between the raw data and the meeting. Now, variance analysis, board reporting, management dashboards, scenario modeling, those kind of things. So, this is the layer where most finance professionals spend the bulk of their time on structuring and formatting rather than thinking. And that's exactly the layer that Claude Design compresses.
To access it, you go to Claude Design, which is claude.ai/design, and it is at the moment only browser, not on the desktop app. And it's available on the paid plans. Now, once you're in, the workspace has three sections: a chat panel on the left for instructions and uploads, a design files area in the middle, and a preview canvas on the right where the artifact renders as it is built. Now, that is the environment everything we're about to build runs inside.
Financial board deck. So, the first workflow that we're going to build is a full board presentation from a single Excel file. Now, the reason I'm starting here is that this is the workflow where most finance teams lose the most hours. The analysis is usually done, and the numbers exist, but what takes time is turning them into a structured story that holds up in front of leadership. And that's the part that we want to remove.
So, inside Claude Design, again, claude.ai/design, I select the slide deck option from the left panel, and then name the project Financial Board Deck. I attach the data set, which is a multi-tab Excel file from a power and utilities company with a real reporting structure. I then tell Claude to act as a senior finance professional preparing a presentation for leadership to analyze the file and build a clean executive-friendly deck that highlights what matters most for decision-making. I'm not telling it which slides to create or what to put on each one. The point is to see whether it can make those decisions itself.
Now, what happens next is the part that matters. Claude does not jump straight to slides. It first reads the data and then forms an interpretation of what the data set represents. Now, you can see it identify the company structure, the reporting period, and the key performance areas before it touches the design. Then, it builds a narrative outline. Slide titles get written as messages and each slide already communicates a point rather than just naming a section, which is exactly how board decks are structured in practice.
Now, once the narrative is set, it defines the visual direction like layout, typography, color palette, and density. And after that, the build starts. So, you watch the slides assemble in the preview panel. Now, most importantly, Claude reviews its own output as it goes. Now, one of the slides it detects content overflow, it fixes the layout, and then it continues. Now, this self-correction loop is what separates a usual deck from raw output. Now, by the time it finishes, the slides are visually consistent, the spacing is controlled, and nothing looks like it was generated and then just abandoned.
Now, once it's ready, I go to the share option in the top right, and I select export, and then I choose PowerPoint format with universal font so that the file renders consistently across devices. Now, the download opens directly in PowerPoint as a fully editable deck. With it open, I can walk you through what Claude actually builds. The final deck opens with a slide that sets the tone immediately. It is minimal, structured like a real executive presentation, and frames the discussion without loading the first page with data. The second slide gives a full business snapshot showing financial and operational indicators together against plan. From there, the deck moves into what drove performance. The later slides go into depth where it counts for a board audience. A liquidity section explains how cash position moved over the period rather than just showing the closing balance. An operational slide connects things like outages and efficiency directly to financial outcomes. There is a people slide that flags workforce gaps affecting future performance, a capital and covenant section covering debt coverage and sensitivity, and a risk section where issues are ranked with context and suggested actions. The deck closes with future priorities and a summary slide that leaves room for discussion. The presentation ends with a summary slide that reinforces the main message and leaves room for discussion. The structure is intact, the visuals translate cleanly, and it's ready to present.
Variance analysis. The second workflow is a variance analysis app, but it is structured very different from the board deck. The board deck is a one-time output. This one is a reusable system. So, the goal is to build something that takes any messy financial file, applies proper variance methodology, and returns a structured tool that you can use again next month with a different data set, and get the same quality of output.
To set this up properly, I prepared three different files representing three different industries, different formats, and different naming conventions. One is a CSV, two are Excel files with multiple tabs. Now, real finance data is rarely consistent across periods, and definitely not across companies. So, a system that only works on a perfectly structured input doesn't really have a real-world use.
Now, before going into a Claude design, I prepare the analysis logic separately. So, I take the various analysis methodology from the Claude Finance plugin, the one that defines materiality thresholds, driver decomposition, and narrative structure for various commentary. Claude design does not have access to that plugin directly, so I copy the full logic into a markdown file, which is variance-analysis.md, and attach it as the instruction layer. Now, this is what turns the output from generic AI commentary into something that follows actual a methodology.
Now, inside Claude design, I select prototype. I choose high fidelity for the design quality. I name it variance analysis system, and then I just click create. Now, I attach the markdown file along with the first data set. Then, I send a structured prompt that defines what we are actually building. Since the prompt is quite long, it is pasted as a text file, and it instructs Claude to apply the methodology from the markdown file strictly. The prompt here, it instructs Claude to construct a reusable application. It must accept messy financial files, detect line items and comparison columns automatically, and then apply the methodology from the markdown file. The output has to be a clean, interactive interface with a header summary, a top drivers panel, a sortable variance table, a click-to-explain rows and proper materiality filtering.
Now, the key piece in the prompt is the instruction to make logical assumptions when the data is ambiguous and proceed without asking questions, which is what allows the system to run end-to-end.
Now, Claude reads both files and immediately starts thinking out loud on the left panel. It confirms it has read the markdown methodology and the data file. It identifies issues in the data set, duplicates, missing values, inconsistent labels, and walks through how it plans to clean them. Then, it defines the design direction. It starts parsing the data into a structured format and builds the application piece by piece. The data parser, the variance calculation logic, the explanation engine, the interface components all get assembled in sequence in the design files panel.
Now, the first version of the app comes out clean. The summary section at the top gives an immediate read on overall performance, and below that, the top drivers panel splits favorable and unfavorable variances so that the eye goes straight to what really matters. Now, the main table lists every line item with sortable columns and color coding for direction, and there is materiality filtering built in so that you can focus only on variances above a particular threshold. And then clicking on any row reveals a short explanation, what changed, the quantified impact, and the likely driver. The methodology from the markdown file is visibly carrying through.
Now, here is where the test gets real. I upload a second file with a completely different structure to see if the system actually generalizes. Now, the good news is the analysis still works. Line items get detected, variances get calculated, and drivers get identified. The bad news is the KPI cards at the top break because the second file does not have the same revenue and gross profit fields that the first one had. So, the system is too dependent on fixed financial categories.
Now, rather than rebuilding the app, I just tell Claude to remove the pre-defined KPI cards entirely and replace them with a dynamic summary that works for any data set focused on total variance, count of material items, and top drivers. I also tell it to make the comparison column detection fully dynamic, so it does not break on naming variations like budget, plan, forecast, or prior. Claude then rebuilds the summary layer. It makes the data handling more tolerant of incomplete inputs, and the second version comes out much more robust.
Now, to make the app portable, I open the share menu, I select export, and then I choose the HTML option. Now, the whole application downloads as a single standalone file. I open it in my browser, and it runs exactly the way it did inside Claude Design. Except, it no longer needs Claude.
Now, to prove the system actually works as a reusable app, I upload the third file directly into the exported HTML. Now, it processes the new data instantly, the summary updates, the drivers refresh, the table populates, and the materiality threshold still adjusts in real time. So, we have built a one-time deliverable and reusable tool.
Finance dashboard. The third workflow is a finance dashboard. We have done plenty of dashboards on this channel before, but this one is worth showing for a specific reason. It gets built in the same environment as the other two artifacts, but the result looks like a real management reporting pack rather than a single screen visualization.
In Claude design, I select prototype. I set it to high fidelity. I name it finance dashboard, and then I click create. I attach the Excel file, which is a multi-section investment banking data set covering trading, deals, clients, and risk. And the prompt I give is lighter than the variance analysis one. I tell Claude to analyze the file, identify the most important metrics and trends, to ignore irrelevant parts, structure the dashboard logically, and use appropriate charts and summaries.
Now, the intent is to see how much of the structuring and prioritization the system can do on its own.
Now, what happens before any visual work is the part I want to point out. Claude reads the data and immediately starts validating it. It notices that some variance values look unusual and pauses to investigate. It checks whether months are missing or whether the scaling is off. Then, it identifies that part of the data set has no actual values for certain months, and then it correctly interprets those as forecast periods rather than treating them as errors. It then rechecks the date range. It realizes that the data covered a shorter time frame than it first assumed, and then it aligns all calculations to that period. This is the kind of contextual reading that usually happens manually before you even open a visualization tool. Only after the data is fully understood does it move into building. It defines the design direction first and then it constructs the dashboard in modular sections. So, KPIs at the top, trend charts in the middle, detailed tables at the bottom, and tab based navigation between four views, which are overview, trading, deals, and clients and risk. That refines the visuals as it goes and it corrects scaling mismatches before finalizing.
Exporting works the same way as the variance app. Share menu, export to standalone HTML, download, and then open in a browser.
Now, with the dashboard running locally, I can walk you through what Claude actually built. So, the final dashboard is structured the way a real investment bank would build a half-year management pack. The overview tab shows H1 revenue at 966 million against a slight underperformance to budget, net income at 370.6 million with a 38.4% margin, trading P&L up 23.8% at 101.6 million, and a 29 billion deal pipeline. There is a written narrative embedded directly in the view explaining where performance came from and which divisions were above or below plan.
Now, the trading tab moves into desk level detail. 43.5 million from fixed income with a 100% win rate. Well done. Weaker performance in equities, regional breakdown by APAC, US, and Europe, and then a trade blotter with strategy notes. Now, the deals tab covers pipeline by sector with fees, a deal roster including closed wins and losses, and the clients tab brings in concentration risk, tier level, growth rate, and total portfolio exposure.
So, that is the full picture of what Claude design does for finance work. Now, the shift this represents is not just speed. It is that the distance between having data and being able to explain what that data actually means is now much, much smaller. The system handles the structuring, the calculation, and the presentation layer. So, your job is to review what it produces, to apply the judgment that only comes from knowing the business, and to make the decisions that no automated system can make for you.
Now, remember, if you want the prompts and the data sets from this video, they're in the free community, which is linked in the description below. And I'll see you next time.