Custom Insight Elements

Sometimes, none of the ready-made insight element types quite fit what you’re picturing. That’s what Custom is for: describe the visualization you want in plain language, and Caplena’s AI builds it for you — layout, colors, labels, and all. No chart-building menus, no fiddling with settings: just tell it what you want to see.

​

Where to find it

When adding an insight element to any report section, look for Custom at the top of the Generate category in the side panel. You’ll see a blank canvas with a prompt field: “Describe what you want to visualize and AI will build it for you — the sky is the limit.”

​

Why use Custom instead of a preset

The preset elements (NPS score, topic frequency, driver analysis, and so on) cover the most common report needs. Reach for Custom when you want:

• A specific chart type not covered by a preset — a bar chart or a map are just two examples; you can also ask for a pie or donut chart, a butterfly chart, a world or regional map to show results per region, a stacked bar chart, and more
• Precise styling — a particular color palette, a shared scale across charts, labels in a specific place. Note that a lot of this is also achievable through a custom theme applied to a whole report, so Custom isn’t your only option if you want consistent styling everywhere
• Multiple charts in one widget — one per segment, side by side, without building each manually. A single Custom insight element can contain several charts, so you stay within the same insight-element-with-widgets framework rather than needing separate elements
• A one-off visualization for a specific stakeholder ask that doesn’t need to become a permanent report fixture

​

Writing a good prompt

The more specific your prompt, the closer the first result lands to what you had in mind. A strong prompt usually mentions:

1. Chart type — donut, bar, stacked bar, map, line, radar, etc.
2. What’s being measured — which columns, scores, or topics (you can reference several at once).
3. How it should be split — by segment, country, verification status, time period. Keep in mind that any filters or segments already applied to the report also apply here — Custom builds on top of that filtered view, not the full unfiltered dataset.
4. Styling cues — color scheme, exact labels, title, legend placement, rounding.

Adjustments like these are optional, not a requirement — if you don’t mention a scale, color, or legend placement up front, Caplena’s AI will just pick a sensible default. You can always fine-tune afterward with a simple follow-up prompt, like “legend at the bottom” or “use shades of blue as colors.” Below are real examples built from a demo airline-reviews project, including a couple of “before and after” prompts that show how a small amount of extra detail turns a so-so result into a genuinely useful one.

​

Example 1: Grouped bar chart

Prompt: “Grouped bar chart comparing average Overall Rating for the top airlines by review volume, with bars split by Verified (true/false), sorted by overall average descending. Show the exact value above each bar.” This produced a clean, titled chart (“Avg Overall Rating by Airline: Verified vs Not Verified”) with a legend, exact values labeled above every bar, and airlines sorted from highest to lowest rating — ready to use as-is.

​

Example 2: World map

Prompt: “World map showing average Overall Rating by Territory, colored on a red-to-green scale.” This produced a full world map (titled “Average Overall Rating by Territory”) shaded from red (lowest average) through yellow to green (highest), with only territories present in the data colored in and the rest left grey — a fast way to spot regional patterns without building a table first.

​

Example 3: Donut charts, before and after

First prompt: “Two side-by-side donut charts comparing the share of Recommended = Yes for Verified vs Not Verified reviews.” The first result had two real problems: the percentage labels were cut off mid-number (“yes: 29…” instead of the full value), and both slices used shades of blue, making “yes” and “no” hard to tell apart at a glance. Refined prompt: “Two side-by-side donut charts titled ‘Recommendation Rate: Verified vs Not Verified’. Left chart: Verified reviews, right chart: Not Verified reviews. Each shows the share of Recommended = Yes (green) vs No (red). Show full percentage labels with one decimal place, not truncated. Add a legend below each chart.” The refined version fixed both issues: full, precise percentages (e.g. “yes: 30.0%”, “no: 73.9%”) and a clear green-vs-red color split that makes the comparison readable immediately.

​

Example 4: Radar chart

Prompt: “Radar chart comparing Seat Comfort, Cabin Staff Service, Food & Beverages, Ground Service, Value For Money, Inflight Entertainment, and Wifi & Connectivity for Air Canada vs Air France. Round all values to 2 decimal places in both the chart and tooltips.” This produced a clean two-airline comparison across all seven service metrics at once, with a legend and every value rounded to 2 decimal places. A radar chart like this is a fast way to spot an airline’s relative strengths and weaknesses across several metrics simultaneously — something that would otherwise take several separate bar charts to show.

​

Example 5: From addresses to a city-level map

Location often matters as much as the metric itself. Store visits, service points, lounges, delivery hubs, whenever your data ties back to a real place, seeing it on a map tells you things a table can’t: whether a problem is spread evenly or concentrated in one region, or whether performance actually tracks geography at all. The catch: Custom can’t show a city-level map on its own, and it needs coordinates, not addresses, to plot anything. Two small problems, one fix, let’s walk through it. Take airport lounges, for example. Every review comes tied to one specific spot, not “Emirates” in the abstract, but this lounge, this terminal: “LHR – Emirates Lounge T3,” “CDG – Qatar Premium Lounge.” Averaged across the whole network, sentiment might look perfectly fine, but is Frankfurt actually pulling its weight, or is it quietly dragging the average down while Amsterdam carries the brand? That’s exactly the kind of question a map answers at a glance and a spreadsheet buries. Turn addresses into coordinates. Before writing any prompt, add a Smart Column that geocodes your address column:

1. Go to Smart columns → New column → Formula, and set Output column type to Text.
2. Use the built-in address_to_coordinates filter — either write it directly:
   {{ Source (Column value) | address_to_coordinates }}
   
   or click Generate with AI and describe it in plain language, e.g. “Convert the address in the Source column into coordinates.”
3. Compute preview to confirm you’re getting a [longitude, latitude] array. Save and apply to all rows.

Prompt for the map: Prompt: “Create a world map showing lounge locations. Use the ‘Coordinates (from Source)’ column for positioning. Represent each lounge as a labeled marker, with a colored ring around each marker indicating net sentiment. Zoom the map to show the area(s) where the data points are.” At global scale, isolated lounges read cleanly right away, one ring, one label, sentiment obvious. Europe was the real test: five lounges packed within a few hundred kilometers, all competing for the same patch of map. Refine dense clusters. A follow-up like “zoom in on the European cluster and spread out overlapping labels, but keep markers at their true geographic position” untangles the labels without faking the geography — two lounges at the same airport still show up clustered together, just readably so. Double check the tooltips. In one pass, this map’s tooltips showed unresolved placeholders (like {@[2]}) instead of real values. A direct follow-up — “tooltips show placeholder text instead of the actual net sentiment and mention values, please fix” — resolved it. Worth a quick hover-check before treating any Custom map as final.

​

Refining an existing Custom element

You don’t need to delete and start over if a Custom element isn’t quite right. Below the chart, use the “Refine your chart…” input field to describe what to change, the same “name the specific flaw” approach from the examples above works well here too (e.g. “round all values to 2 decimal places”, “use green for yes and red for no”).

​

Tips for better results

• Name the exact columns or segments if your project has several similarly-named ones — the AI works from your project’s structure, and precision here avoids guesswork.
• Ask for value labels up front (“show the exact value above each bar”, “full percentage labels, not truncated”) — this was the single biggest fix across our test prompts.
• Say if you want one chart or several — “one chart per segment” is a very different ask from “one combined chart with all segments shown together.”
• Specify colors when meaning matters — e.g. green for “yes” and red for “no” reads instantly; two shades of the same color doesn’t.
• Mention the scale explicitly if comparability matters — e.g. “on a shared scale” ensures multiple charts use the same color range instead of each auto-scaling to its own data.
• Ask for rounding if precision isn’t the point — raw unrounded decimals in tooltips or labels look unpolished; say “round to 2 decimal places” up front.
• Iterate by naming the specific flaw rather than rewriting the whole prompt — small, targeted edits get you there fast.
• Use Custom for exploration, then swap to a preset element later if the chart turns out to be something you’ll want to keep updating long-term with built-in configuration options.

​

Related

• For the full list of ready-made elements, see Insight Elements – General Overview & Settings.
• Example gallery report showing every example live.