Route Planning & Storytelling for Sulu

Our focus was Sulu, an existing Yeti built platform that allows tour providers to easily create custom self-guided tours. One of the platform’s most painful friction points has been the route-building process. Currently, users must jump into external tools like GeoJSON editors to build routes, export files in XML or KML formats, and manually upload them into the Sulu back office—a tedious and inefficient workflow.

We saw an opportunity to simplify this dramatically using AI. Our goal was to build a smarter, in-browser route builder that not only allowed users to draw, edit, and export KML-based routes easily, but also layered on rich, generative storytelling powered by AI. From this single concept, the experiment quickly branched into multiple directions—each exploring different possibilities for smarter, faster, and more engaging tour creation.

The Experiment

We began by standing up a basic web app using Yurt, our internal monorepo tool. From there, we used Cursor—a coding environment powered by AI—to rapidly build out a map-based UI using Mapbox. Users could drop pins, drag route lines, select transport modes (walking, cycling, driving), and export a completed route as a KML file, streamlining what was previously a multi-step external process.

Once the foundational route builder was in place, we explored three key feature directions:

• 1. Discover Interesting Routes
  This feature introduced a search modal with a bounding box to surface nearby points of interest (POIs). Using the Mapbox Search API, it pulled in places based on category and relevance, then layered in data from Wikipedia and Unsplash to show descriptions and imagery. A fallback system ensured a relevant image was always shown—from Mapbox, Wikipedia, or Unsplash.

• 2. AI-Powered Storytelling
  This feature allowed users to select interest categories—like local history, food, or cultural trivia—and generate narrative scripts for specific segments of the route. Prompts sent to OpenAI generated context-aware, spoken-style content designed to serve as audio tour scripts. However, we ran into a recurring issue: the AI often ignored bounding box constraints and referenced unrelated or distant locations. After extensive prompt tweaking, we achieved higher relevancy, though geography remains a known weakness in LLMs.

• 3. Aggregated Location Context
  This feature surfaced enriched contextual information about stops along a saved route. Using the Wikipedia API, it provided brief descriptions, tags (like “park,” “museum,” “landmark”), and geo-location previews to add educational or cultural value to each point. The idea was to automatically populate a sidebar feed with layered insights as users mapped their tours.

Throughout the process, we experimented with other AI tools—Gemini 2.5 and Windsurf—but ultimately returned to Cursor and ChatGPT for their familiarity and effectiveness. One major constraint we encountered was Cursor’s default behavior of writing all logic into a single bloated App.tsx file. Without modularity or architectural guidance, our feature branches diverged quickly, making code sharing and merging essentially impossible. Each team member ended up working independently on their own version of the app, limiting our ability to consolidate progress into a unified product.

What We Learned

The experiment demonstrated just how fast and far AI can take us in early-stage product development. Cursor, in particular, excelled at bootstrapping UI, integrating APIs, and responding to specific prompts—getting us to functional prototypes in hours rather than days or weeks. We were able to layer on smart, location-aware features and generate rich storytelling content with minimal manual coding.

At the same time, the project surfaced key limitations of prompt-based development workflows:

• AI lacks architectural discipline. Without predefined structure, the generated code quickly becomes messy and unmanageable.
• Real collaboration is still difficult. Cursor and tools like Lovable are not designed for multi-user workflows, making parallel development and merging difficult.
• Geographic reasoning is weak. Even with highly specific prompts, generative models struggled to stay within defined map boundaries or relevance zones.
• AI can easily “vibe” into complexity. Without tight constraints, small changes often spiraled into bloated logic or confusing abstractions.

Despite those limitations, the output was impressive—and promising. Each branch independently delivered a valuable, usable feature. In future iterations, we’d pair tools like CodeGuide with Cursor to define architecture, scope, and modular components before building. That way, AI output could be better structured and easier to maintain or combine.

Perhaps most exciting is the broader implication: with tools like these, we could offer clients or internal teams working prototypes of advanced features in a matter of days—not weeks. We could also use this process to validate product concepts early, generate investment-ready demos, or build functional MVPs at a fraction of the traditional cost.