How Dataguessr Content Is Created

Everything you can play on Dataguessr is based on simple datasets. Every new topic that gets added follows the same semi-automated process. AI does most of the legwork, but all topics start with human input and are reviewed at multiple steps in the process. Below, I outline how exactly this human-machine-collaboration works.

This page is inspired by PlotLines’ «How it’s made» page.

Human Input

1. Getting started

The AI agent always waits for human input to kick into action. Once activated, the agent follows the instructions defined in a Cursor Rule.

Start from a dataset

I can input a link to a specific chart or a dataset, e.g. from Our World in Data. The agent follows the link and moves onto source check.

Start from a topic idea

I can input a rough topic idea, like «I want a new topic on internet access». The agent then starts looking for matching datasets.

AI

2. Source check

The agent checks the data source for a number of predefined criteria. Is the source reliable? Are we allowed to use the data? Is the data recent enough? Does it cover a large number of countries? If it checks all boxes, the agent proceeds.

AI

3. Turn data into a topic

The agent drafts all metadata that turn a raw dataset into a playable Dataguessr topic. It creates a title, question copy for different contexts and assigns the topic to a category. It also documents the data source, year, footnotes if needed and adds a link where users can learn more about the topic. All of this gets written into a json file that the game draws from.

AI

4. Data Processing

The agent formats the data according to the data structure the game expects and makes sure all country names match the spelling we use in the game. The final csv gets saved.

Human Review

5. Review

A human (hi, it’s me 👋) reviews all metadata, usually there is one or two things I’ll rewrite to make the questions crisper. I also do spot checks on the data set to make sure values and countries are correct.

AI

6. Package

The agent starts a python script that generates an Open Graph image for social media based on the metadata. Then, the agent runs a few last checks: is the JSON with the newly added topic valid, is the topic added to the list of playable topics, does the Open Graph image exist?

Human Review

7. Playthrough

I play the new topic to see if everything works and to spot irregularities from the perspective of a player.

Human Signoff

8. Publish

If all looks good, the topic is ready to be published on dataguessr.com. From then on, it is available in the Explore section, is added to the rotation of the Daily challenge, and is being used in the Feed.