PEARC26 Cursor Workshop — Weather Patterns

Hands-on workshop using Cursor to explore US weather patterns. Work is organized into modules; Modules 1–2 are ready now, with later modules to follow.

Workshop plan

Module	Focus	Status
1	Map wind across the CONUS from a provided weather station dataset	Ready
2	Investigate a severe weather event and trace it via station-metadata MCP	Ready
3	(coming later)	Planned

---

Prerequisites

Complete this setup before starting Module 1.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Cartopy may need system libraries (GEOS/PROJ) on some platforms. On macOS with Homebrew:

brew install geos proj

The first wind-map run downloads Natural Earth basemap shapefiles into Cartopy's data directory (requires network).

---

Module 1 — Wind map

Plot meteorological wind barbs across the contiguous United States using the provided dataset in data/ (~500 real CONUS stations with a week of hourly readings). See data/README.md for schema and units.

The dataset spans one week of hourly readings (2025-06-09 through 2025-06-15). Each run plots a single timestamp. For example, Wednesday noon UTC:

python scripts/plot_wind_map.py --timestamp 2025-06-11T12:00:00Z

Writes output/wind_map.png.

---

Module 2 — Investigate a severe weather event

On the evening of Friday, 13 June 2025, a strong storm cell moved through central Oklahoma. Start by mapping winds around that time — for example:

python scripts/plot_wind_map.py --timestamp 2025-06-13T22:00:00Z

Study the map, then use Cursor to inspect the underlying hourly readings.

Explore the hourly CSV

Important: Read these instructions yourself in the editor. During Ask/Agent investigation, attach only the data files you want analyzed (e.g. the CSV below). Do not @README.md — it will leak the exercise framing into the model’s context.

1. Ask mode + @File the dataset

Switch the chat to Ask mode, attach only @data/weather_hourly.csv, and run an open-ended quality check — for example:

@data/weather_hourly.csv Do a quality check of this weather dataset. Flag any physically impossible or suspicious values.

Try a few follow-ups if the first pass is too broad (e.g. focus on wind, or on a specific evening window).

2. Narrow the location

If you find a suspicious station_id, you can @data/stations.csv (or @data/README.md for units) for coordinates and naming. For hardware/software details, use the station-metadata MCP in the next section — do not dig through metadata JSON files by hand.

3. Ask Cursor to quantify

Have Cursor summarize ranges, count impossible values, or walk the hours around the storm for the affected station — whatever helps you explain why the map looked wrong.

Station metadata MCP (root-cause context)

Each station has hardware/software metadata (platform, sensors, firmware language and typed fields). That catalog is not meant to be browsed as raw files during the exercise — query it through a local MCP server so the agent can pull details on demand.

Start / connect the MCP server

Switch back to Agent mode for this step (Ask mode cannot call MCP tools).

1. This repo ships a project MCP config at .cursor/mcp.json that launches:

.venv/bin/python scripts/station_metadata_mcp.py

2. In Cursor: Settings → MCP. Confirm station-metadata appears and shows a green/connected status. If needed, click refresh/restart after creating the venv.

3. In an Agent chat, verify tools such as get_station, get_firmware, list_stations, search_stations, and list_firmware are available.

Discovery path

After you have a suspicious station_id from the hourly CSV (still in Agent mode):

1. Ask the agent to look up that station via the station-metadata MCP (not by opening JSON files in the editor).
2. Follow the station’s firmware_id with get_firmware.
3. Inspect how that firmware stores wind speed (language, integer width, signedness, range) and connect it to the impossible reading you found.

Example prompts (attach the CSV only if needed; prefer MCP for metadata):

Using the station-metadata MCP, call get_station for the station_id with the impossible wind value. What hardware and software is it running?

Using the station-metadata MCP, call get_firmware for that station’s firmware_id. How is wind speed stored, and could that explain a reading of -127?

Fix the wind map

Once you understand the bad reading, harden scripts/plot_wind_map.py so one corrupted station cannot distort the CONUS plot.

1. Plan the fix

Switch to Plan mode and ask Cursor to propose an approach — for example:

@scripts/plot_wind_map.py We found a physically impossible negative wind speed caused by legacy signed int8 overflow at one station. Plan a fix in this plotting script so invalid wind values are handled safely (without silently inventing weather). Keep valid stations plotting as before.

Review the plan before coding (what counts as invalid, whether to drop vs. flag, logging/messaging, etc.).

2. Implement the fix

Switch to Agent mode and have Cursor implement the agreed plan in scripts/plot_wind_map.py.

3. Re-run the map

Regenerate the storm-time map and confirm it no longer blows up on the bad value:

python scripts/plot_wind_map.py --timestamp 2025-06-13T22:00:00Z

Compare output/wind_map.png to your earlier plot. The Oklahoma storm cell should still appear; the impossible barb should not dominate the map.