generate-map

Turn unstructured text that names real-world places into a richly-annotated map link.

What this skill produces

A single message containing:

1. The geojson.io URL (clickable, opens straight to the map with markers)
2. A short summary of how many landmarks were plotted and any that couldn't be located

The GeoJSON itself is embedded in the URL — no file is saved unless the user asks.

Output Feature schema

Each landmark becomes one Point Feature with this exact shape:

{
  "type": "Feature",
  "properties": {
    "name": "Oakland Temple Hill Gardens",
    "description": "Serene temple gardens in the Oakland Hills with manicured blooms, reflecting pools, and sweeping Bay views. Best in spring evenings for golden light and cherry blossoms.",
    "address": "4770 Lincoln Ave, Oakland, CA",
    "hours": "Open daily: ~9 AM – 9 PM",
    "best_time": "Spring evenings for golden light + cherry blossoms",
    "maps_url": "https://www.google.com/maps/search/?api=1&query=Oakland%20Temple%20Hill%20Gardens%204770%20Lincoln%20Ave%20Oakland%20CA",
    "directions_url": "https://www.google.com/maps/dir/?api=1&destination=Oakland%20Temple%20Hill%20Gardens%204770%20Lincoln%20Ave%20Oakland%20CA",
    "reviews_url": "https://www.google.com/maps/search/?api=1&query=Oakland%20Temple%20Hill%20Gardens%204770%20Lincoln%20Ave%20Oakland%20CA",
    "marker-color": "#e377c2",
    "marker-size": "medium",
    "marker-symbol": "garden"
  },
  "geometry": {
    "type": "Point",
    "coordinates": [-122.1866, 37.8079]
  }
}

GeoJSON coordinates are [longitude, latitude] — easy to flip, double-check.

Workflow

1. Extract landmark names

Read the input and pull out every distinct real-world place the user is asking about. Don't include vague references ("a coffee shop downtown") — only nameable, geocodable places. If the text already gives a clear list, use it as-is. If it's prose, extract the proper nouns.

If the input has city/region context (e.g. "my Oakland day trip"), keep that context — it improves geocoding accuracy and disambiguates common names.

2. Geocode each landmark

Use scripts/geocode.py to look up coordinates and address via OpenStreetMap Nominatim. The script handles user-agent headers, the 1 req/sec rate limit, and returns [lng, lat, display_name] for each query.

python scripts/geocode.py "Oakland Temple Hill Gardens, Oakland, CA" "Lake Merritt, Oakland, CA"

Output is JSON to stdout, one object per query. Pass each landmark with whatever city/region context you have — "Lake Merritt" alone is ambiguous, "Lake Merritt, Oakland, CA" is not.

If a landmark can't be geocoded, note it and continue — don't fail the whole batch. Report which ones failed at the end.

3. Enrich each landmark via web search

For each successfully-geocoded landmark, use WebSearch (or WebFetch on a likely page) to ground the metadata fields. You're looking for:

• A 1–2 sentence evocative description (what makes this place worth visiting; sensory or experiential)
• hours (operating hours if applicable; "Open 24/7" for parks/viewpoints; "N/A" for natural features)
• best_time (when it's most worth visiting — golden hour, season, weekday vs. weekend)

Don't fabricate hours. If a quick search doesn't surface them, write "Hours unknown — check before visiting" rather than inventing. The description and best_time can lean on general knowledge if web search doesn't return useful results.

4. Pick marker styling

Infer the landmark category from name, address, and what the search returned. Map category → marker symbol + color. Use the Maki icon set — geojson.io renders these.

Common mappings:

If the input has a natural grouping (e.g. "morning stops" vs "afternoon stops", or "must-see" vs "optional"), use color to encode that grouping instead of category — it's usually more useful for the map's reader. Mention which encoding you chose in the summary.

marker-size is medium by default. Use large for headline stops if the user implies a hierarchy.

5. Build the URL

Use scripts/buildmapurl.py — it takes a FeatureCollection JSON on stdin and prints the geojson.io URL. It URL-encodes the JSON and prepends https://geojson.io/#data=data:application/json,.

echo '<feature-collection-json>' | python scripts/build_map_url.py

If the resulting URL is over ~6000 characters, geojson.io may struggle in some browsers. The script warns on stderr if you hit that. For very large maps (20+ landmarks with long descriptions), consider trimming descriptions or splitting into multiple maps — and tell the user what you did.

6. Present to the user

Reply with:

• The URL on its own line so it's clickable
• A 1–2 line summary: how many landmarks, any that failed to geocode, what the marker color encoding represents
• Don't dump the full GeoJSON in chat unless asked. It's already in the URL.

Why each piece matters

• Web-searching every landmark is slow but it's the difference between a map with rich, trustworthy hover-cards and a map with hallucinated hours that mislead the user when they actually try to visit. The user explicitly chose this tradeoff.
• Including city/region context in geocoding queries prevents Nominatim from picking the wrong "Lake Merritt" — there are multiple places with most landmark names in the world.
• Maki icons specifically because that's what geojson.io's renderer recognizes; arbitrary symbol names just become default pins.
• [lng, lat] order is the GeoJSON spec. Almost everyone gets this wrong on first try because Google Maps shows lat/lng. Re-check after building.

What not to do

• Don't save a .geojson file unless the user explicitly asks. The URL is the deliverable.
• Don't include landmarks you couldn't geocode as Features with [0, 0] coordinates — that drops a marker in the Atlantic. List them in the summary as "couldn't locate" instead.
• Don't fabricate operating hours, prices, or addresses. "Unknown" is fine.
• Don't add fields beyond the schema above. geojson.io ignores unknown fields but they bloat the URL.