Prompting tips
How to get the best results from Felt AI — whether you're working with the AI assistant, building Extensions or custom popups, or writing SQL queries.
Felt AI shows up in a few places, and each one works a little differently. This page collects prompting tips by surface: the AI assistant, Extensions, Custom popups, and SQL queries.
Tips that apply everywhere
Use exact names. Generic references create ambiguity. Use actual layer names, attribute names, and values from your data.
✅ Good:
Filter the "permits" layer to show only records where permit_type is "residential" and status is "pending".
❌ Weak:
Show me the pending residential permits.
Handle edge cases. Say what should happen when data is missing or a filter returns nothing.
If assessed_value is null, display "Not available" instead of blank.
Describe the format you want. When output format matters, give an example.
Format dates as March 15, 2025 not 2025-03-15. Display currency as "$1,250,000" with commas and dollar sign.
Build incrementally. Complex requests built in one prompt usually need a lot of fine-tuning. Small, testable steps are easier to verify, easier to debug, and let you adjust direction based on what's working. When something isn't right, refine rather than starting over.
AI assistant
The AI assistant finds data, runs analysis, and styles your map from conversation — you describe what you're trying to learn or show, and it delivers results on the map.
Describe the goal, and let it work
Give the assistant your intent and constraints, not a step-by-step procedure. It can see your layers, connected data sources, and Felt's data libraries, and it will pick the data, analysis, and styling to match your goal.
✅ Good:
I'm evaluating flood exposure for our Houston portfolio. Add FEMA flood zones, identify which of our properties fall in high-risk zones, and style the result so risk levels are obvious at a glance.
❌ Weak:
Add a layer. Now filter it. Now change the colors.
Reference data with @
Type @ to point the assistant at exactly the layer or data source you mean — especially useful when names are similar or your workspace is large.
Compare @Census Tracts with @Store Locations and highlight tracts with no store within 2 miles.
Iterate in conversation
The assistant keeps the context of your conversation, so refine the way you'd direct a teammate:
Make the high-risk zones more prominent and mute everything else.
That classification hides the variation I care about — try quantiles instead.
Zoom to where the risk is concentrated.
Keep conversations focused
Start a New chat when you switch to an unrelated task. A conversation about one problem gets better results than one carrying leftover context from three different tasks.
Ask it to explain
If you want to understand or verify what it did:
Walk me through how you calculated that risk score.
Which data source did that layer come from?
Extensions
Extensions are interactive tools where users click, filter, and explore. In the Extensions editor, Felt AI drafts JS SDK code you can read, edit, and run. Think: "when the user does X, show them Y."
What to specify:
The trigger event (click, select, filter change)
What layers are involved
What should display in the panel
What visual feedback appears on the map
How to reset or clear previous results
✅ Good:
Create an interactive tool where I can click on any school to find nearby parcels. When I click a school:
Draw a 500-meter buffer around it
Spatially filter the parcels within that radius
Show a panel with the school name, count of nearby parcels, and a list of those parcels with their zoning type and assessed value
Clear previous results when I click a different school
❌ Weak:
Find parcels near schools.
For complex tools with multiple interactive features, number your requirements — each one is small and testable, and the AI follows the logic better than a run-on description.
When refining, check each layer of functionality — do clicks, filters, and selections behave correctly? Are calculations returning expected results? — and fix specifically:
The count isn't updating when I change the filter — can you check why?
Remove the chart, just keep the attribute summaries.
Custom popups
Custom popups display formatted information when users click a feature. Felt AI drafts the HTML, CSS, and JavaScript, which you can edit directly.
What to specify:
Exact attributes to show and their format
Any conditional styling or specific branding (color-code status values, show icons)
Links or images to include
✅ Good:
Create a popup for this layer showing:
Address as a bold heading
Price formatted as currency
Bedrooms and bathrooms as "3 bd | 2 ba"
Square footage with comma separator
A link to "View listing" using the listing_url attribute
❌ Weak:
Add a popup for properties.
Preview in both desktop and mobile modes before saving — layout requests like "keep it to two lines on mobile" work well.
SQL queries
SQL queries let you ask spatial questions of your cloud database. Felt AI translates your question into SQL you can inspect, edit, and run, and results become a map layer.
What to specify:
The spatial relationship you're looking for (within distance, intersects, contains)
Which tables to use
Any filters or conditions
What attributes to include in results
✅ Good:
Find all fire stations within 2 miles of areas where response times exceeded 10 minutes last year. Join with the incidents table and show the count of slow responses per station.
❌ Weak:
Show me fire station coverage.
If a query looks slow or returns too much, ask the AI to optimize it or add limits — and review the SQL it wrote; it's yours to tune.
Last updated
Was this helpful?