Turn a messy spreadsheet into a live dashboard with Claude
Hey, it's Cindy 🌱 You commented DATA, so here is the exact prompt from the reel plus the whole cheat sheet: how to clean the file first so the charts are actually right.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
before you start
What you need (it is less than you think)
A Claude account and a file. That is the whole list. Claude builds the dashboard as an artifact, a live interactive panel that opens next to the chat and runs in your browser, and artifacts work on Free, Pro, Max, Team and Enterprise plans (what artifacts are). No Looker Studio, no Power BI licence, no one on the data team owing you a favour.
File types that work CSV, XLSX, JSON, and plain text. Google Sheets users: File, then Download, then Comma Separated Values.
Size limits Up to 500MB per file and up to 20 files per chat, so a normal export is nowhere near the ceiling (file limits).
How to attach it Drag the file into the chat box, or click the plus in the lower left and choose add files.
If the panel does not appear Open Settings, then Capabilities, and check that code execution and file creation is enabled. It is on by default on Free, Pro and Max.
Real data means real privacy calls. Strip customer names, emails, phone numbers and anything you would not paste into a shared doc before you upload. A dashboard of anonymised rows tells you exactly the same story.
step one
Clean the mess before you chart it
This is the step people skip, and it is the reason their first dashboard looks confident and is quietly wrong. Merged header rows, dates stored as text, a total row sitting in the middle of the data, three spellings of the same client name. Charts do not care, they will happily plot nonsense. Run this first and read what comes back.
🧹 audit and clean my file
I have attached a raw export. Before we build anything, audit it and tell me what is wrong with it.
Load the file and give me: the real column names and what each one appears to actually contain, the number of rows, and then every data problem you can find. Specifically look for dates stored as text or in mixed formats, numbers stored as text (currency symbols, commas, percent signs), duplicate rows, near-duplicate category names that are really the same thing spelled differently, blank cells and what they most likely mean, subtotal or total rows sitting inside the data, and any column that is a merged or repeated header rather than data.
Then propose a cleaned version: tell me the exact fixes you would make and which ones need a decision from me rather than a guess. Do not build any charts yet.
When Claude asks whether a blank cell means zero or means unknown, that question is worth more than the dashboard. Answer it properly, because it changes every average on the page.
step two
The exact dashboard prompt
Once the file is clean, this is the one from the reel, written out properly. It names the sections, forces the numbers to be filterable rather than baked in, and tells Claude to say so when the data cannot answer something instead of inventing a chart.
📊 build my interactive dashboard
Using the cleaned data, build me an interactive dashboard as an artifact that I can filter and click around in.
Context: this data is [what it is, for example monthly sales by rep and region], and the people looking at it care about [the one or two decisions it should inform].
Build it with: a row of headline numbers at the top (total, average, change versus the previous period, and the single most useful ratio for this data), then three or four charts that each answer a different question, then a sortable table of the underlying rows so anyone can check your working. Add working filters for [date range, category, region, whatever the key columns are] and make every number and chart on the page respond to those filters rather than being hardcoded.
Design it clean and readable: light background, one accent colour, clear axis labels with units, no chart junk. Under each chart write one plain sentence saying what it shows and what stands out.
If the data cannot honestly answer something, say so in the chat instead of building a chart that implies it can.
👩💻 Open a new chat and attach the file. Drag the CSV or XLSX straight into the message box. Give it its own chat so the context stays clean.
🧹 Run the audit prompt first. Read the problems it found, answer any question it asks you, and tell it to apply the fixes.
📊 Paste the dashboard prompt. Fill in the two square-bracket bits. The artifact panel opens on the right and builds in front of you.
🖱️ Click everything. Change a filter, sort the table, check one headline number against your own spreadsheet by hand. If that one number is right, the rest usually is.
🔁 Fix it in one sentence at a time. Use the cheat sheet below. Claude edits the artifact in place, so you are not starting over each time.
the cheat sheet
The follow-up lines that fix anything
You do not need a new prompt to change the dashboard. These are the sentences I actually type, one at a time, and each one edits the artifact where it stands.
"Make the pie chart a bar chart" Swaps the chart type in place. Bars beat pies almost every time once you have more than four categories.
"Add a filter for [column]" New dropdown or slider wired into every chart on the page, not just the one below it.
"Show change versus last month, not just totals" Turns a snapshot into something that tells you a direction.
"Sort the table by [column] descending and show the top 20" Stops a 4,000 row table from swallowing the page.
"Label the axes with units and add a one-line takeaway under each chart" The single fastest upgrade from student project to presentable.
"Explain what this chart is telling me, and what it is NOT telling me" The one that saves you in the meeting. It surfaces the caveats before someone else does.
Two more worth keeping in your back pocket: "make it work on a phone screen" and "make the colours colourblind-safe". Both take one line and both get noticed by whoever you send it to.
the other output
Want a real Excel file instead?
Sometimes the person you are sending this to does not want a link, they want a workbook they can poke at. Claude can build an actual XLSX with live formulas in the cells, not values pasted in as text, and file creation is available on every plan from Free through Enterprise (create and edit files with Claude).
📁 build the Excel version
Now build the same analysis as a real .xlsx workbook I can download and hand to someone else.
Structure it as: a Data tab holding the cleaned rows, a Calculations tab, and a Summary tab at the front with the headline numbers and charts. Every number on the Summary tab must be a live formula that references the Data tab, not a pasted value, so that when someone replaces the raw data next month the whole workbook updates. Use named ranges or a proper table so the formulas keep working if the number of rows changes.
Format it to be readable: bold headers with a frozen top row, currency and percentage formatting where it belongs, and a short notes cell at the top of the Summary tab saying what the data is, what period it covers and any assumption you made while cleaning it.
Tell me which formulas you used for the headline numbers so I can check them.
keep it alive
Sharing it and refreshing it next month
🔗 Publish the artifact to share it. Open the artifact and hit Publish to get a link anyone can open in a browser, no Claude account needed on their end (publish and share artifacts). Check what is on the page before you send it, since a published link is public to anyone who has it.
📅 Rebuild it monthly in one message. Next month, open the same chat, attach the new export and say: "same dashboard, new data, keep the layout and filters identical and tell me what changed versus last month." You get continuity plus a diff.
📌 Make it repeatable properly. Once the layout has settled, ask Claude to write the whole thing up as a skill so you stop re-explaining the format. In the app that is Customize, then Skills, then Add, then Upload a skill, and it works on every plan including Free (creating custom skills).
🧾 Keep the raw file. The artifact is built from the data you gave it, it does not stay connected to your source. Keep the export that made it so you can prove any number later.
The honest limits. This is not a live database connection, it is a snapshot of the file you uploaded, so nothing updates on its own. Very large files get sampled or summarised rather than fully loaded, so if the row count Claude reports does not match your file, say so and ask it to process the whole thing. And always hand-check one headline number before anyone else sees the page.