Curated's chat is a natural-language interface to a multi-agent reasoning system. Most of the time, you can just ask the question and the right thing happens. But there are patterns that consistently produce better answers.
Three things every good prompt has
- What — the geographic operation: find, count, compare, route, score, screen, rank
- Where — the location or scope: a city, a zip code, a trade area, a parcel, a region, a custom polygon
- Filter — the constraints: foot traffic over 10,000/day, opened after 2020, within 1 mile of a school, household income above $75k
A good prompt:
Find all coffee shops within a 5-minute walk of Penn Station, NYC, with foot traffic over 10,000 per day.
A weaker prompt:
Tell me about coffee shops near Penn Station.
Both will get an answer. The first will get an answer you can put in a deck.
Be explicit about geography
"Texas" is clearer than "the South." A specific zip code, county, parcel ID, or named place beats a region adjective. Where ambiguity matters, name the boundary explicitly:
- "...within Travis County, TX..."
- "...inside this drawn polygon..."* (then draw on the map)
- "...in the trade area we defined yesterday..."
Specify the time frame
Most data Curated draws from is timestamped. If you want 2024 numbers specifically, say so. If you want a comparison across years (2010 vs 2020 census), say that too. Vintage matters; specifying it removes one source of ambiguity from the answer.
Use follow-ups to refine
Don't try to cram everything into one prompt. Ask the broad question, then narrow. Curated keeps the conversation state — every follow-up builds on what came before:
Show me the top 20 candidate sites for a fitness studio in DFW.
Then:
Filter to only those with parking and a household income over $90k within 1 mile.
Then:
Compare the top 5 on competitor density.
When to use voice
Voice mode is best for:
- Short, specific prompts (under ~20 words)
- Hands-free contexts (field, in-car, walking)
- Live demos and presentations
Voice is less good for:
- Complex, multi-clause prompts
- Prompts with specific parameter values (numbers, dates, addresses with punctuation)
- Quiet meetings where the TTS output would be disruptive
Switch modes mid-conversation; session context is shared.
Cross-Space prompts
By default, a prompt runs against the current Space's data context. To run a prompt across Spaces (e.g., a portfolio-wide question), say so explicitly:
Across all my Spaces, find the artifacts that mention "grocery anchor."
Recovering from a wrong answer
If Curated gets it wrong:
- Tell it — "that's wrong. Use the Esri demographics layer instead of the Tapestry segments."
- Source-swap — "redo this with our internal CRM data."
- Re-scope — "narrow this to just the parcels we own."
- Report it — the thumbs-down button + report-an-issue capture flags the answer for our prompt-tuning work.
The model learns within a conversation but does not train on your data. Your corrections improve answers in this session; they don't leak into anyone else's tenant.
Related
- Reading Your First Answer (Sources and Citations)
- Refining Results with Follow-up Prompts
- Using Voice Mode