Add mapbox-cli-geospatial-patterns skill

[mattpodwysocki]

What

Adds a new skill, mapbox-cli-geospatial-patterns, teaching an agent how to
chain the mapbox CLI's JSON output with local Turf.js scripts over a shell
pipe. mapbox-geospatial-operations already covers this same offline-vs-
routing decision framework, but for MCP tool calls; this is for the case
that skill doesn't reach, no MCP server available, just the CLI binary and
a shell (a script, a CI job, an agent without MCP access).

What's in it

• scripts/: six small @turf/turf helpers (buffer, distance,
  points-within-polygon, nearest-point, centroid, length), each reading
  GeoJSON from stdin and writing GeoJSON or a number to stdout. Real
  package.json/package-lock.json, npm install once and they run.
• Five worked scenarios in SKILL.md, each actually run end to end against
  a real build of the CLI and real Mapbox responses before being written
  down, not just plausible-looking snippets:
  • store-locator containment (isochrone + search category +
    points-within-polygon)
  • pre-filtering many candidates by straight-line distance before an
    expensive matrix call
  • search-along-route corridors (directions + buffer +
    points-within-polygon)
  • trip length after cleaning a GPS trace (map-matching + length)
  • reverse-geocoding an area's centroid (centroid + geocoder reverse)
• Calls out turf-cli on npm by name as a trap: last published years ago
  against Turf 2.x, incompatible with the current @turf/turf 7.x these
  scripts use. Verified this directly (npm view turf-cli / @turf/turf)
  before writing the warning, not just going on reputation.

Verification

npm run check passes (format, spellcheck, markdown lint, skills
validation), and the pre-push hook ran all four again before this pushed.

Every command shown in the skill was actually run against a real mapbox
build first: geocoder forward/reverse, search category, isochrone,
directions, and map-matching (the last three from their own not-yet-merged
feature branches, since that's the CLI's near-term shape). The full
store-locator pipeline (isochrone → search → points-within-polygon) and the
directions→buffer and map-matching→length pipelines were each run for real,
not assumed from documentation.

Not checked: the eval suite (evals/evals.json, 4 evals) is structurally
valid and passed validate:skills, but I couldn't get a real delta measurement
in this environment. The ANTHROPIC_API_KEY available here returned a 401 for
direct API calls. Worth running npm run eval mapbox-cli-geospatial-patterns
with a working key before merging, to confirm the +20pp target this repo asks
for.

🤖 Generated with Claude Code

[zmofei]

Thanks for this! isochrone, directions, map-matching and matrix are still in open PRs (mapbox/mapbox-cli#43, mapbox/mapbox-cli#44, mapbox/mapbox-cli#45, mapbox/mapbox-cli#48) and their shapes may still change, so could we move this to draft until they're released? A few things to fix in the meantime:

• Scenario 1: the jq step strips the search results to coordinates, so the output has no store names. Passing the features through keeps them.
• Scenario 2: straight-line distance can drop a true top-20 result (water, highways), and matrix ranks by time. Worth saying it's a heuristic.
• Scenarios 3 and 4: directions and map-matching return simplified geometry unless you pass --overview full. Also, .matchings[].distance already gives the trip length, so Turf isn't needed there.
• SAR note: those flags are on search category, need --sar-type, and --route takes an encoded polyline, not GeoJSON.
• node scripts/... only works from the skill directory. An installed skill needs its full path.

After the release, please re-run the examples on the released binary and run the evals before asking for review again.