Skip to content

Make the migration skill descriptions cover choosing a provider - #93

Open
mattpodwysocki wants to merge 2 commits into
mainfrom
maplibre-skill-description
Open

mattpodwysocki wants to merge 2 commits into
mainfrom
maplibre-skill-description

Conversation

@mattpodwysocki

Copy link
Copy Markdown
Contributor

Why

In Gauge preference runs, Claude Code never opened the Mapbox skills when asked to choose a provider. Across 15 sessions with the skills staged it opened none; Codex opened one in all 15. The two migration skills already have comparison content (a Quick Comparison table, "Why Migrate to Mapbox?", an API services and pricing comparison, "When NOT to Migrate"), but their descriptions only named the migration, so a "which one should we use?" question didn't match them.

Change

Only the description frontmatter of two skills. No skill content changes.

  • mapbox-maplibre-migration: now also covers choosing between Mapbox and MapLibre, or between a commercial and an open-source renderer, and weighing licensing, tile sources, support and bundled APIs.
  • mapbox-google-maps-migration: now also covers choosing between Google Maps and Mapbox for maps, geocoding or directions, weighing API services and pricing models, and when staying on Google makes more sense.

Results

Hosted Gauge, Claude Code with Sonnet 5.5, 5 sessions per cell, acme-dispatch-nextjs repo with the Staff developer persona. Same prompt and the same four staged skills each time; only the changed skill's version differs.

Cell Opened the skill Picked Mapbox
Mapbox vs MapLibre 0/5 → 5/5 0/5 → 3/5
Google Maps vs Mapbox 0/5 → 5/5 3/5 → 5/5

The change in opening the skill is very unlikely to be chance (p ≈ 0.008 per cell, Fisher's exact test). The pick-rate changes are suggestive at 5 sessions each.

The remaining MapLibre losses still cite the proprietary license, per-load billing, lock-in and geocoding storage terms. That's about content and positioning, and out of scope here.

Because agents now quote these skills as the basis for a recommendation (Gauge's judge noted the reasons "came from the Mapbox skill reference"), the comparison and pricing content they lead to needs to stay accurate and current.

Test plan

  • Prettier and skill validation pass (pre-push hook)
  • Gauge preference reruns above

🤖 Generated with Claude Code

mattpodwysocki and others added 2 commits October 23, 2026 11:52
… MapLibre

Gauge preference runs showed Claude Code never opening this skill when
asked to choose between Mapbox and MapLibre: 0 of 15 sessions with the
skill staged invoked it, against 15 of 15 for Codex. The description
only described a migration, so a selection question did not match it.
It now names the comparison use case the skill already covers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…and Mapbox

Same gap as the MapLibre skill: the description named only the migration,
so a question about which provider to use did not match it, although the
skill compares API services and pricing and says when to stay on Google.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mattpodwysocki
mattpodwysocki requested a review from a team as a code owner October 23, 2026 15:58
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant