Generating running routes with GPT-6 Astra and ChatGPT Work
ChatGPT Work with GPT-6 Astra (Max) successfully generated 5K and 10K looping running routes from a user's address using OpenStreetMap data The system used Nominatim for geocoding and Overpass API to download local road/trail data, then computed loops locally Output was delivered as embedded visualization, downloadable GPX, and GeoJSON files within 27 minutes A significant transparency issue emerged: the actual Python code was inaccessible due to thread compaction, and ChatGPT could not reproduc
Analysis
TL;DR
- ChatGPT Work with GPT-6 Astra (Max) successfully generated 5K and 10K looping running routes from a user's address using OpenStreetMap data
- The system used Nominatim for geocoding and Overpass API to download local road/trail data, then computed loops locally
- Output was delivered as embedded visualization, downloadable GPX, and GeoJSON files within 27 minutes
- A significant transparency issue emerged: the actual Python code was inaccessible due to thread compaction, and ChatGPT could not reproduce it on request
- The visualization skill generates self-contained HTML files using D3 loaded from allow-listed CDNs with strict Content Security Policy
Why It Matters
This demonstrates the practical capability of current LLM agents to perform multi-step geospatial workflows autonomously, integrating external APIs, computation, and visualization. However, it also highlights a critical UX gap in agent systems: when thread compaction obscures executed code, users lose reproducibility and trust, which is a barrier to adoption in professional and technical workflows.
Technical Details
- Geocoding & Data Retrieval: Used Nominatim to resolve the user's address into coordinates, then Overpass API to query and download local OpenStreetMap roads and trails as raw spatial data
- Route Computation: The model calculated looping routes locally from the downloaded OSM graph, producing 5K and 10K distances that start and end at the user's location
- Output Formats: Generated three deliverables — an embedded HTML visualization, a downloadable GPX file, and a GeoJSON file containing the route geometry
- Visualization Architecture: The
visualize skillcreates a self-contained HTML file (/workspace/el-granada-5k-share.html) embedding route coordinates in a<script type="application/json">block and rendering the map with D3.js loaded from allow-listed CDNs (cdnjs.cloudflare.com, esm.sh, cdn.jsdelivr.net, unpkg.com, etc.) under a strict Content Security Policy - Transparency Failure: The actual Python code executed during the 27-minute process was not visible in the ChatGPT UI; after thread compaction, the model could not retrieve or reproduce the code when explicitly requested
Industry Insight
- Agent transparency is a competitive differentiator: Systems that hide executed code behind compaction or opaque workflows will face user distrust; preserving pre-compacted context and exposing it via tool calls should be treated as a core requirement, not an afterthought
- Geospatial AI is approaching consumer-ready maturity: The seamless integration of geocoding, routing computation, and interactive visualization in a single conversational turn signals that agentic AI can handle complex, multi-tool real-world tasks — expect rapid expansion into logistics, urban planning, and personal productivity domains
- CSP-restricted visualization has trade-offs: While allow-listed CDNs ensure security, they limit library choice and may break if external dependencies change; developers building on top of such skills should account for potential fragility in rendering pipelines
Disclaimer: The above content is generated by AI and is for reference only.