By now, you've heard me talk about the four-legged stool of AI growth. We started with API, which gave us control. Then we added Markdown, which gave our knowledge structure. Now we're ready for the third leg: JSON.
If the first two legs made sense to you, take a breath. JSON sounds technical, but it's not a new language you have to learn. It's simply a way of giving the AI a template to fill in, the same way you give learners a form with spaces for their name, their score, and their feedback. Think of JSON as that form written in plain text.
If you've ever worked in Excel or Google Sheets, you already know the concept. Each column is a field name. Each cell holds a value. JSON is the same thing, just expressed in text so the AI can follow it exactly.
Here's a simple example:
{
"title": "Lockout/Tagout Basics",
"objective": "Identify when and how to apply LOTO procedures",
"estimated_minutes": 12
}
In plain English, we've said: give me a course with a title, an objective, and an estimated time. The AI doesn't get to wander. It has to fill in those boxes. That's what we mean by structured response.
Now let's bring it closer to home. In most authoring tools you've probably used variables. A variable is just a container that holds text, a number, or a yes/no value. JSON works well with variables. You tell the AI, through the API, to return its answer in a JSON format. Then your course can map each JSON field into the matching variable. Suddenly the course isn't static. The AI is filling in the variables for you.
Think about a feedback screen. Normally, you'd type out every message by hand. With JSON, you can tell the AI to send back three fields: feedback, next hint, and retry (true or false). Your course reads those values and displays the right message. It's still your design. The AI is just populating the fields.
Here's what that could look like in JSON:
{
"feedback": "Correct. You selected the right PPE for aluminum MIG.",
"next_hint": "Before you adjust voltage, check your wire speed.",
"retry": false
}
And in your authoring tool, those values would drop neatly into variables you already use:
- vFeedback (text)
- vHint (text)
- vRetry (true/false)
You don't have to worry about the technical connection yet. For now, just see how the fields in JSON line up one-to-one with the variables in your tool. That's the bridge. In the next step, we'll look at how to actually pass that data back and forth.
This is where JSON shows its strength. Markdown let you shape ideas into clear sections. JSON goes a step further by locking those ideas into specific boxes. Because the boxes are predictable, you can reuse the output anywhere: in your course, in a database, or in a report. One format that stays consistent across uses.
Here's another example outside of courses. When people write prompts for generating images or video, JSON can be used to organize the details: subject, setting, camera angle, style. Laying it out this way forces you to be mindful of the parts, and it makes it easier to scale. Want twenty consistent images? Change the subject field, keep the rest the same. The structure does the heavy lifting.
For example, you might ask for an image with a simple line like: a welder standing on the factory floor with sparks flying in the background. That is freeform text. Now compare it to a structured JSON version:
{
"subject": "welder",
"setting": "factory floor",
"detail": "sparks flying in background",
"composition": "medium shot",
"camera_angle": "eye level",
"style": "realistic photo"
}
The first approach gives the AI room to interpret. The second makes sure every element has a place. That is the added clarity JSON brings.
If you think about it, much of the prompt training is doing the same thing, just not in JSON format. Frameworks like RISE, TAG, and ACE are meant to guide best practices for prompting the LLM. They provide structure. JSON expresses that same idea in a format that makes it easy for different platforms to communicate.
If you understood the first two articles, you can handle this one as well. JSON is a way to tell the AI, "fill in these boxes so I can trust what comes back." It might look like code, but it's just structure. When you reach that point, your AI work shifts from trial and error to something you can rely on.
Top comments (0)