Skip to content

Instantly share code, notes, and snippets.

@mattpodwysocki
Created April 17, 2026 20:38
Show Gist options
  • Select an option

  • Save mattpodwysocki/0a198cff0c39cf2dc3f2dc7c02906ba9 to your computer and use it in GitHub Desktop.

Select an option

Save mattpodwysocki/0a198cff0c39cf2dc3f2dc7c02906ba9 to your computer and use it in GitHub Desktop.
AI at Mapbox presentation plan - Monday demo

AI at Mapbox — 20-Minute Demo Presentation

Date: Monday
Format: Demo-heavy, 20 minutes
Apps: Pre-built before the presentation


Structure

Segment Time
Skills intro + skills.sh discovery 2 min
Store finder app walkthrough 3 min
MCP servers intro 2 min
Trip planner app walkthrough 8 min
Where this is going / call to action 3 min
Buffer 2 min

Segment 1: Skills (2 min)

  • What are skills? Markdown files that teach agents Mapbox-specific patterns
  • Live: open skills.sh, search for Mapbox, show the skill list
  • Show a SKILL.md briefly — "this is what the agent reads"
  • Point to npx skills add mapbox as the install path

Segment 2: Store Finder App Walkthrough (3 min)

App: Pre-built store locator using Mapbox skills

Interesting moments to highlight:

  • Show the SKILL.md that guided the build — make it concrete
  • Point out a specific correct pattern the agent got right because of the skill:
    • Token security (string resource / Info.plist, not hardcoded)
    • Annotation manager created once and reused, not recreated on each update
  • The contrast: "without the skill, agents get this wrong consistently"

Talking point: Skills close the gap between what the LLM knows about Mapbox and what's actually correct today.


Segment 3: MCP Servers Intro (2 min)

  • Three servers: Mapbox (Search, Directions, Isochrone, etc.), DevKit (styles, tokens), Docs (live documentation)
  • Together they give the agent live Mapbox data + style control + up-to-date docs in one session
  • Skills tell the agent how to use Mapbox. MCP servers give it live access to Mapbox.

Segment 4: Trip Planner App Walkthrough (8 min)

App: Pre-built trip planner using all three MCP servers

Interesting moments to highlight:

  1. Grounding in action

    • Ask: "What restaurants are near the hotel?" without MCP → hallucinated answer
    • Ask same question with Mapbox MCP → live data, citations, neighborhood name from reverse geocode
    • "This is what grounding means — anchoring the response in live data instead of training data"
  2. ground_location_tool

    • Show the single tool call replacing reverse geocode + category search + isochrone
    • Show the structured response with citations
  3. DevKit live style swap

    • Toggle the map style live — show the map update in real time
    • "The agent controls the style, not just the data"
  4. Docs MCP

    • Show an SDK question answered inline with current documentation
    • "It's not guessing from training data — it's reading the actual docs"

Segment 5: Where This Is Going (3 min)

  • Skills coverage gaps: testing, Flutter, React Native — issues filed, contributions welcome
  • Evals: we measure skill quality with a before/after delta — show the +37pp improvement example
  • Grounding as a Mapbox product direction: ground_location_tool is a proof of concept for a grounding API
  • Call to action: try npx skills add mapbox, file issues, contribute skills

Apps to Build

  • Store finder — uses mapbox-store-locator-patterns skill, search + markers + directions
  • Trip planner — uses Mapbox MCP (search, directions, isochrone), DevKit (styles), Docs MCP

Open Questions

  • Which platform for store finder? (web / iOS / Android / React Native)
  • Trip planner stack? (web most demo-friendly)
  • Record backup video of each demo in case of live failure
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment