04 · Spreadsheet to visual report
At a glance
| Solves | A few thousand rows that need to become charts and conclusions someone can read |
| Needs | An Excel or CSV file; the xlsx Skill |
| Takes | 5–15 minutes depending on data size and chart count |
| Risk | Low for the source data — but high for correctness of the result |
The situation
Sales sends you 2026H1-sales-detail.xlsx, 8,000 rows. Your boss wants to know by tomorrow which product lines are growing, which are declining, and why.
You open it and find 23 fields, six of them abbreviations, two duplicated column names, and one column entirely empty.
Preparation
Install the xlsx Skill (part of the office document suite) and put the file in the workspace.
A spreadsheet is the input most easily misread
Abbreviated field names, merged cells, multiple header rows, hidden columns, text mixed into numeric columns, inconsistent date formats — any one of them produces a result that looks entirely presentable and is entirely wrong.
And a wrong chart is more dangerous than no chart: it looks like evidence.
Which is why this playbook must be two rounds.
The prompt
Round one: confirm it read correctly
Read 2026H1-sales-detail.xlsx in the workspace and answer only these four:
1. What fields does it have, and what do you understand each to mean?
2. How many rows? Are there multi-level headers or merged cells?
3. Which fields have missing values, and how many each?
4. Is there anything clearly anomalous — negative amounts, future dates,
duplicate rows?
Answer only those four. Do not start analysing. Do not produce any files.Read the answer carefully, especially point one. If it reads GMV_ex as "GMV excluding tax" when internally you mean "GMV after removing fraudulent orders", every conclusion downstream is wrong.
Correct anything that is off:
Two corrections:
- GMV_ex is GMV after removing anomalous orders, not excluding tax
- In the channel field, D means direct, not distributor
Also ignore rows where status is not "confirmed" — those are unsettled orders.Round two: analyse
Correct. Now:
Aggregate revenue and margin by month and product line, producing:
- A stacked bar chart: monthly revenue by product line
- A line chart: margin trend by product line
Then write a conclusion of under 250 words naming the two changes most worth
attention. For each: what happened, how large, and what might explain it.
Where the data does not support an explanation, say "cannot be determined
from the data".
Output as an HTML page with interactive charts (values on hover), saved to
the workspace as H1-sales-analysis.html.Why it is written that way
Two rounds is the entire point. Thirty seconds confirming field understanding beats three rounds of rework. This is small steps in its most valuable form.
Round one asks about missing and anomalous values. Those two numbers tell you how much to trust the conclusions. Two thousand of eight thousand rows missing an amount means every total needs discounting.
"Where the data does not support an explanation, say cannot be determined." The biggest distortion in analysis work lives here — a model is very good at inventing a plausible reason for a decline ("likely seasonal effects") that the data does not support at all.
Interactive HTML rather than static images. HTML opens in the Browser view on the right, hover reveals exact values so you can spot-check, and it is shareable by nature.
A 250-word conclusion cap. Uncapped, you get an essay. Capped, it has to choose what matters.
Acceptance checklist
- [ ] The field meanings from round one match your understanding
- [ ] Hand-check the total for two or three months
- [ ] Axes, units and legends are labelled
- [ ] The conclusion points at changes that genuinely exist in the data
- [ ] Unsupported speculation is actually marked "cannot be determined"
- [ ] The HTML opens in the Browser view and the charts respond to hover
- [ ] How missing and anomalous values were handled is stated in the report
Hand-check one number
Pick a month, SUMIF the total yourself in Excel, and compare it to the chart.
It takes a minute, and it is the line between a report you can present and one you cannot — especially the first time you use a new spreadsheet structure.
Variations
No data yet — gather it first
Gather market information on AI office tools in China over the past year,
focusing on user scale, use cases and representative products, as structured data.
Requirements:
- Note the source and publication date for every data point
- Distinguish "officially disclosed" from "third-party estimate"
- Leave fields blank where no reliable data exists — do not fill with
industry rules of thumb
Output as CSV first. I will decide about charts after reviewing it.This needs the web-search Skill. CSV before charts, for the same reason as above: when data is unreliable, a chart just packages the unreliability more attractively.
As a Word report
From that analysis, produce a Word report:
1. Executive summary (half a page, conclusions and recommendations only)
2. Data notes (definitions, time range, how missing values were handled)
3. Analysis by product line, one section each, with charts
4. Conclusions and recommendations
5. Appendix: the full data table
Save as H1-sales-analysis.docx.A dashboard that keeps updating
If this analysis recurs monthly, do not rerun it from scratch:
Split this into two files:
1. sales-data.csv — data only: month / product line / revenue / cost / orders
2. sales-dashboard.html — reads sales-data.csv and renders charts and cards
From now on I only update the CSV and the page follows.That is the Library's MD / CSV / HTML split applied to a single case. Next month is a few new rows.
Merging multiple tables
There are 12 monthly sales files in the workspace with supposedly identical
field structures.
Check first: are the field names identical across all of them? Does any month
have extra or missing columns? Answer only that — do not merge yet.Inconsistent fields is the number-one cause of botched merges.
Let it suggest the questions
When you are not sure what to look at:
Read this sales data and tell me: if you were our business lead, what three
questions would you most want answered, and why?
Do not analyse yet — questions and reasoning only.FAQ
It cannot read the Excel file.
Confirm the xlsx Skill is installed. Failing that, try a model with document support, or save as CSV and retry.
Tens of thousands of rows — will it blow the context?
No — the xlsx Skill reads the sheet with code rather than stuffing it into context. But analysis results and intermediate output do enter context, so do not ask it to print all the raw data.
For genuinely huge files see model tiers, but first consider whether you can aggregate before analysing.
Can I control the chart colours?
Yes, just say so: Use our brand colours #5B47E0 and #14B8A6 on a light background.
Or show it: Match the chart style in this screenshot.
The numbers are wrong.
Check round one's field understanding first — eight times in ten that is the problem. The other two are usually numbers stored as text, or a definition mismatch (does "revenue" include returns?).
State the definition and rerun: Revenue means net of returns; returns are in the refund column.