Choose the format around the next step
Plain text is convenient when a person will read and edit the result. JSON is useful when you need named fields for comparison or another workflow step. The decision is about how you will use the response.
The instructions and the response format are separate choices. You can write a plain-language prompt asking for JSON, or provide a JSON-formatted input and ask for a readable explanation.
Compare the same task
Use these fictional notes: “Mira will draft the landing-page headline. The launch date is not confirmed. We still need an owner for the image selection.” The task is to extract actions while keeping unknown information visible.
Plain-text instruction
Summarize these notes as action bullets. For each action, identify the
owner and due date when supplied. Write "not assigned" or "not confirmed"
for missing information. Do not infer an owner or date.JSON instruction
Extract the actions from these notes. Return only a JSON object with an
actions array. Each action must contain task, owner and due_date.
Use null for an owner or date that is not supplied.Illustrative structured response
{
"actions": [
{"task": "Draft the landing-page headline", "owner": "Mira", "due_date": null},
{"task": "Select images", "owner": null, "due_date": null}
]
}This is a written example. It illustrates a response shape without claiming comparative model performance.
When plain text fits
- A person is reviewing a strategy, explanation or first draft.
- The result needs commentary and flexible section lengths.
- You will manually move a few actions into another tool.
When JSON fits
- You need the same fields across many results.
- A program will read the response.
- You want to validate field names, values and missing information.
Check both formats for accuracy. For JSON, also parse and validate the object before passing it downstream. Provider structured-output features can help software enforce a shape, while the factual content still needs review.
Explore standard prompts and JSON prompts, or read what a JSON prompt is for a complete starter example.