Example prompts¶
The model decides when to call cluster_points based on what you ask. These
prompts reliably trigger a tool call and get a useful explanation back.
Good prompts¶
Specific about the input shape
Here's a GeoJSON FeatureCollection of recent rhino sightings in Kruger. Use
cluster_pointswitheps = 3000andminPts = 4to find sighting clusters, then summarise each cluster in plain English.
Asks for analysis on top of the tool output
Cluster these delivery drop-points and tell me which cluster is likely to need a second driver — that's the one with more than 20 stops within a 2 km radius.
Lets the model pick parameters
Cluster these GPS pings into camps. Pick reasonable DBSCAN parameters for "human-scale walking distance", run
cluster_points, and explain your parameter choice.
Prompts that go sideways¶
Vague geometry
"Cluster this CSV." — the tool needs GeoJSON. Convert first, or ask the model to do the conversion.
Too-loose eps
"Cluster with
eps = 100000." — at 100 km you'll get one giant cluster. Start small (1 km) and widen until the output stops merging.
Trying to cluster polygons
cluster_pointsonly consumesPointfeatures. Centroid your polygons first if you must — or open an issue asking forcluster_polygons.
Tips for getting the most out of the model¶
- Tell the model why you're clustering. "Find delivery routes" leads to
different
eps/minPtschoices than "find rare-species hotspots". - Ask it to render the output. "Now produce a Leaflet snippet that shows these polygons on top of an OSM basemap."
- Stack tools. Once we add
convex_hullorbufferyou can chain them: cluster, then hull, then buffer.
Want to ship a prompt with your team?
Drop it in the /users/prompts/ folder of our examples
repo (PRs
welcome) so others can reuse it.