Making maps with AI

The 2026/27 theme of the Coding Café is AI in research software development, and the #30DayMapChallenge is a good place to try it: a small, self-contained task, with a result you can see. AI assistants and coding agents are genuinely good at map code. They know the GeoPandas and sf APIs better than most of us, and they will happily write the forty lines of matplotlib styling you did not want to look up.

They are also confidently wrong in ways that are specific to maps, and a wrong map looks just as convincing as a right one. This page is about getting the first without the second.

What goes wrong¶

These are the failures we see again and again, roughly in order of how much damage they do.

None of these produce an error. The script runs, the PNG looks good, and the map is wrong.

Prompting so it does not happen¶

Give the assistant the same information you would give a new colleague. A prompt that works well for map code has five parts:

1. The data, with its path, format and CRS: “data/eindhoven_cafes.geojson, points, EPSG:4326”. If it needs to download something, name the source.

2. The CRS to work in: “Reproject everything to EPSG:28992 before plotting or measuring.”

3. What the map should show, as a statement: “Show that cafés with a terrace cluster in the city centre.”

4. What must be on the image: title, legend with units, credit line with “© OpenStreetMap contributors”, the CRS.

5. What to do when stuck: “If the data is missing or a download fails, stop and tell me. Do not invent data, not even as a placeholder.”

An example:

Using Python with geopandas and matplotlib, make a map from data/eindhoven_cafes.geojson
(points, EPSG:4326) and data/eindhoven_boundary.geojson (polygon, EPSG:4326).
Reproject both to EPSG:28992 first.

Show the cafés that have outdoor_seating=yes in red and all others in grey, on the boundary
in light grey. Title: "Where to sit outside with a coffee in Eindhoven". Add a legend and a
credit line: "Data: © OpenStreetMap contributors (ODbL) | EPSG:28992".

Save to maps/day01-<me>/map.png at 7x7 inches, 200 dpi, white background.
Use only the data in these files. If something is missing, stop and ask me.

For a coding agent working in a repository, put the rules in an instruction file (AGENTS.md, CLAUDE.md or .github/copilot-instructions.md) so you do not have to repeat them in every prompt. The starter repo’s AGENTS.md is a ready-made example to copy.

Checking the result¶

Never accept a map because it looks right. These checks take a minute and catch every failure in the table above. The code below runs them on the Day 1 data; put the same lines at the end of any script an AI wrote for you.

import geopandas as gpd

cafes = gpd.read_file("data/eindhoven_cafes.geojson")
boundary = gpd.read_file("data/eindhoven_boundary.geojson")

# 1. Is the data what you think it is? Count and look at a few rows.
print(len(cafes), "features;", cafes.geom_type.unique().tolist())

# 2. Is it where you think it is? The Netherlands is roughly 3-7.3° E, 50.7-53.6° N.
xmin, ymin, xmax, ymax = cafes.to_crs(4326).total_bounds
assert 3 < xmin and xmax < 7.3 and 50.7 < ymin and ymax < 53.6, "not in NL: lat/lon swapped?"

# 3. Are all layers in the same, projected CRS before measuring?
cafes, boundary = cafes.to_crs(28992), boundary.to_crs(28992)
assert cafes.crs == boundary.crs and cafes.crs.is_projected

# 4. Do the numbers make sense? Eindhoven is about 89 km².
print(round(boundary.area.iloc[0] / 1e6, 1), "km²")

# 5. Is every point inside the area it is supposed to be in?
print(cafes.within(boundary.union_all()).mean() * 100, "% of cafés inside the boundary")

91 features; ['Point']
89.4 km²
100.0 % of cafés inside the boundary

Then, by eye:

• Compare with a real map. Open the same area at https://www.openstreetmap.org or in Overpass Turbo. Are the points where the cafés actually are?

• Read the code for any hard-coded data. A list of coordinates or names typed into the script is a red flag unless you put it there.

• Check the legend against the data. Do the class edges match describe()? Does the colour bar run the right way?

• Check the credits. Every source named, with its licence.

If the assistant made a mistake, tell it what was wrong rather than asking it to “fix the map”. “The points are in the Indian Ocean: you swapped latitude and longitude in points_from_xy” gets a correct answer; “it looks wrong” gets a different guess.

Being open about it¶

The challenge has one rule about AI: every map’s README says, in one line, what the AI did and what you did.

AI disclosure: AI wrote the plotting code, I chose the data and the projection.

AI disclosure: I wrote the code; an assistant helped me debug a CRS error.

AI disclosure: No AI used.

This is not a confession. It makes AI use something the group can learn from: which tasks it helped with, and which it made harder. Whatever the AI produced, you are responsible for checking that the data and the map are real and correct.

Day 10: Prompting only¶

Day 10 asks for a map made entirely through an AI chat or agent: you may not edit the code yourself. It is a good test of everything on this page. Save your prompts (for example in prompts.md next to the map), and note how many rounds it took and what you had to correct. The data still has to be real: “prompting only” means the AI writes the code, not the facts.

Day 27: New tool¶

A good day to try an AI coding agent if you have only used chat so far, or the other way round. Compare the result with a map you made yourself earlier in the month.