Tutorials
Claude Opus 5.5: Check Marketing Charts Before Turning Data Into a Claim
Give Claude Opus 5.5 the source table as well as the chart.
Give Claude Opus 5.5 the source table as well as the chart. Ask it to verify values, denominators, dates and visual encoding before improving the design. A chart can be beautifully aligned and still tell the wrong story. This workflow is for campaign reports and product presentations where the reader must be able to trace a claim back to supplied data.
Choose the check that matches the error
| Suspicious result | Check | Useful correction |
|---|---|---|
| Taller bar has a smaller label | Source values versus plotted values | Rebuild from the verified table |
| A small improvement looks enormous | Axis origin and range | Use a zero baseline for magnitude bars |
| Rates change after grouping | Numerators and denominators | Recalculate from summed counts |
| Mobile labels collide | Longest label at narrow width | Change layout without changing data |
Datawrapper explains the zero-baseline rule for bars. For a trend or a small difference, choose an appropriate alternative chart and make its scale explicit. A dramatic-looking shape is not a substitute for a clear comparison.
A worked example with unequal denominators
The following campaign numbers are invented for teaching; they are not Panelly results or an Opus test.
| Creative | Clicks | Impressions | Click-through rate |
|---|---|---|---|
| A | 80 | 4,000 | 2.0% |
| B | 30 | 1,000 | 3.0% |
| Combined | 110 | 5,000 | 2.2% |
The combined rate is 110 ÷ 5,000 = 2.2%, not the simple average of 2% and 3%, which is 2.5%. B has a higher observed rate, while A produced more clicks. “B is the winner” still needs comparable audience, placement, time and a suitable experiment. Neither conversion performance nor statistical confidence appears in this table, so neither belongs in the chart’s headline.
A useful title is “B had a higher observed click-through rate in this sample.” Put the measurement dates, denominator and source nearby. If the records are not comparable, show them as separate observations rather than ranking them. Missing results should stay missing instead of becoming zero-height bars.
A complete chart-review prompt
Review [source table filename] and [chart or report filename]. The source table is authoritative; the chart may contain errors.
List every plotted series, unit, date window, filter and denominator. Mark missing information as unknown.
Recalculate rates from their underlying counts. Never average percentages without checking weights. Do not infer conversions, significance or causation from click counts.
Compare each displayed label and plotted value with the source. Check bar baselines, axis ranges, omitted categories and missing values.
Propose a truthful headline and the smallest design change that makes the comparison readable. Preserve the data and source notes.
Check the longest labels at 390-pixel and 1280-pixel widths, plus the final static export. Return a discrepancy table with source row, observed issue, corrected calculation and acceptance check.
If code or rendering tools are available, run the calculations and render checks and record their outputs. Otherwise provide reproducible calculations and mark rendering untested. Deliver the corrected source and chart separately.Run numeric checks before tuning color. If the source table and exported chart disagree, stop the design pass and resolve the discrepancy. Keep the original data file unchanged and put derived calculations in a separate table so reviewers can reproduce them.
Test narrow layouts with the difficult labels
Use the longest category, the largest value and the missing-value case as fixtures. A horizontal layout may give long names more space; shortening names should not erase distinctions. If the chart is interactive, ensure a static export still includes the information needed to read it without hover.
For Chart.js specifically, responsive sizing depends on the chart’s container; consult its responsive configuration rather than repeatedly changing the canvas dimensions. A September 9 community question reports label and dimension problems in a different library. It is evidence of the task, not proof that switching libraries fixes it.
Questions before publishing
Can the model recover exact data from a screenshot?
Treat that as transcription needing verification. Small labels, overlapping marks and rounded values can make the original numbers unrecoverable. Request the source file when the claim depends on precision.
Can a generated image contain the final chart?
It can be a concept illustration. For a factual chart, preserve deterministic plotting from verified data and inspect the final export. Do not let image generation silently redraw the numbers.
How does this connect to an advertisement?
After approving the claim, use Panelly Studio to develop the accompanying product visual. Keep the measured chart separate from illustrative imagery. Use the ad-copy revision workflow to align the headline with what the data actually supports.
