AI for GIS: 5 Prompts That Replace Hours of Work

I lost most of an afternoon once to a column of addresses that would not standardize. Some had “ST” where I wanted “Street,” some wanted “Saint,” and a few coordinates were cut off at the wrong digit. I wrote a regex from scratch, then rewrote it twice. The work was tedious, the kind that drains a day and teaches you nothing.

That task is one of five choke points that show up in almost every geospatial workflow. A CSV with bad formatting. A projection that will not align. A GDAL error that reads like noise. These are the moments that stop you cold, and they are the exact problems AI is built to solve.

AI models are language models. They understand text. When the problem is text, text patterns, or text-shaped errors, a model can move through it fast. Below are the five fixes I use, the named tools for each, and the prompt structure that makes them reliable. None of this is about replacing you. It replaces the manual grind so you can spend your hours on architecture and analysis instead.

Why AI for GIS works on these specific tasks

Most GIS pain is not analytical. It is plumbing. Cleaning, reading documentation, decoding errors, aligning coordinate systems, converting formats. These are text problems wearing a spatial costume, which is why an AI GIS workflow handles them so well.

The one thing AI cannot do yet is design the system. It will not tell you what to build or how the pieces fit. That part stays with you. Use AI to clear the plumbing so you have time for the work that matters.

1. The dirty data cleaner

You have messy addresses or coordinate formats, and you are about to spend hours fixing them by hand. I have done this with addresses and strange polyline strings, and it is miserable.

This is the cleanest example of AI for GIS. Take a sample of the messy data, hand it to an LLM, and ask for a standardization script. The model can write the regex pattern for you. If the file is small enough, you can paste the whole thing in and have it do the work directly. Agentic tools like Claude Code can go further, searching documentation and building the full workflow.

Common cases this handles:

• Latitude and longitude values cut off at a certain digit
• Coordinates that include westings, northings, and stray letters
• Values that import as a mix of strings and numbers
• Address abbreviations like “ST” that should become “Saint” or “Street”

Here is the prompt structure I use:

Here is a sample of messy address data:
[paste 10-20 rows]

Write me a Python script that standardizes these. Handle the
ST/Saint/Street ambiguity, normalize coordinate formats, and
flag any rows it cannot confidently fix. Show me the regex you use.

Do not spend a day on this. Let the AI handle it.

2. The documentation black hole

You want to use a new library, but the docs are dense and you do not want to lose a day reading them. Maybe you are trying to render a 3D hexagon layer in pydeck and Streamlit and you just need a working starting point.

This is another strong case for ChatGPT for GIS, with one caution. The model is trained at a fixed point in time. It may not have seen the toolkit you are using. So feed it the source. Paste in the link to the documentation, or pull the page down as plain text first using one of the web-to-text tools online, then paste that.

Then assign it a role:

Act as my senior developer. Here is the documentation for [library]:
[paste docs text]

I want to build a 3D hexagon layer in pydeck rendered in Streamlit.
Read the docs, describe the approach, then give me boilerplate code
I can use as a starting block.

Telling the model how to act is the part most people skip. “Act as my senior developer” focuses it on the outcome you actually want. Use the result to skip the setup and get to real work faster.

3. The error output you cannot read

A GDAL error lands and you have no idea what it wants. Maybe it is telling you to install something. Maybe it is an import error buried in a long log. You could Google the code, but you might grab the wrong fragment.

Take the whole error log and paste it in. Then ask the model to narrow the cause:

Here is the full error log:
[paste log]

Is this an environment issue, a data issue, a programming issue,
or a network issue? Tell me which, then tell me exactly how to
resolve it. If I need to install something, give me the command.

A few tips that make this work in any AI GIS workflow. Start a fresh chat for each error. If you go too deep in one thread, the model starts pulling you in different directions. Ask it the same question a few different ways. And if you have a clue about where the error lives, bring in the relevant documentation from step two and have it answer as a senior developer.

4. The projection mismatch

Projections are the scariest one for me. You build an entire workflow that should run, and then an upstream coordinate system mismatch breaks it. Your raster is in one CRS, your GeoDataFrame is in another, and you never validated that they match. I forget this whether I am in QGIS, a GeoPandas frame, or a Sedona DataFrame.

Before you even start, have the AI write a safety-checked script. Do not just ask it to look up an EPSG code. Ask it to validate and fix:

Write me a safety-checked Python script that:
1. Checks whether the CRS of my vector and raster match.
2. If they do not match, reprojects the vector to match the raster.
3. Prints what it did at each step.

Stack: GeoPandas for the vector, rasterio for the raster.

This is uncomplicated code, but it makes the whole workflow more robust and easier to rerun. It is one of my favorite ways to use AI for GIS because it removes a class of silent failures.

5. The file conversion loop

You have data in a dozen formats and you are still opening QGIS to do “save as” one file at a time. Install GDAL and let the AI write the command instead.

For a single file, ask for the one-off conversion. For a batch, describe the loop:

I have 50 shapefiles in /data/raw. Write me a bash script using
ogr2ogr that converts every one to GeoParquet and writes them
to /data/out. I do not know bash, so add comments explaining
each line. Also tell me how to install GDAL if I do not have it.

This is the power of the command line, unlocked by AI. Even if you do not know bash, the model does. Copy, paste, run. It is one of the easiest ways to automate GIS with AI and start saving time immediately.

The point of all this

The goal of AI is not to replace you. It is to replace the work like this, the heavy manual grind, so you can focus on architecture and analytics. A handful of these prompts will solve a large share of the friction in a normal week. Knowing what to build and how to assemble it still belongs to you.

FAQ

What is the best AI for GIS tasks?

There is no single best tool. For cleaning data and decoding errors, any strong general LLM like ChatGPT or Claude works well. For workflows that need to search documentation and build code across files, agentic tools like Claude Code are a better fit. Match the tool to the task.

Can ChatGPT for GIS actually clean my coordinate and address data?

Yes. Cleaning addresses and coordinate formats is a text problem, which is what language models are built for. Give it a sample of the messy data and ask for a standardization script, or paste the full file if it is small enough.

How do I get reliable code from an AI when it does not know my library?

Feed it the documentation. Models are trained at a fixed point in time, so paste the relevant docs as text, then prompt it to act as a senior developer and build boilerplate from those docs rather than from memory.

Will an AI GIS workflow replace GIS analysts?

No. AI replaces the manual work, not the person. It cannot design system architecture or decide what to build. Those decisions, plus the analysis on top, stay with you.

If you want to learn how to put an AI GIS workflow into real practice, with a community and portfolio projects to build on, take a look at the Spatial Lab: https://forrest.nyc/go/lab/