Skip to main content

Output Formats: JSON, XML, Markdown

You ask AI to analyze customer feedback and it responds with three paragraphs of insights. Helpful — but you needed a table to paste into your quarterly report, or a list of scores to drop into a spreadsheet.

The problem is not what AI says. It is how AI says it. And you can control that.

Tell AI How to Format​

AI does not guess your preferred output shape unless you spell it out. The fix takes one extra sentence in your prompt: specify the format upfront.

Three formats cover almost every business use case.

JSON: Structured Data​

JSON organises information in labelled pairs. Think of it as a filled-in form.

When you need names, ratings, dates, or any data destined for a spreadsheet, ask for JSON:

"Return a JSON object with: name, email, satisfaction score (1-10), main complaint."

The AI responds with:

{
  "name": "Sarah Chen",
  "email": "sarah@example.com",
  "satisfaction_score": 6,
  "main_complaint": "Long wait times on phone support"
}
Use JSON

You plan to move the output into a spreadsheet, database, or another tool.

Both OpenAI and Anthropic support structured JSON modes that guarantee parseable output. For everyday use, simply asking for JSON works perfectly fine.

XML: Structured Prompts​

XML uses angle-bracket tags like labels on filing boxes. <context>, <instructions>, <data>, <task>.

Here is the surprising part: Claude responds better when you structure your input with XML tags. It separates your instructions from the information they apply to — reducing confusion.

The 30 Percent Trick

Put your data at the top and your question at the end of a prompt. Tests show this alone improves response quality by up to 30 percent, especially with long documents.

Example structure:

<context>
Q3 customer survey results: 2,400 responses. Average satisfaction: 6.4/10.
Top complaint: delivery times. Top praise: product quality.
</context>

<task>
Draft an email to the logistics team summarising key findings and recommending three actions.
</task>

By boxing each section, you prevent the AI from mixing instructions with context — a common source of confused outputs.

Markdown: Human Readable​

Markdown is the format people actually read: tables, bullet lists, headers.

Instead of asking for raw data, ask for a table:

"Return a table with columns: Feature | Status | Priority"

The AI gives you a clean table ready for a slide deck, memo, or email.

FeatureStatusPriority
Online bookingLiveHigh
Loyalty programPlannedMedium
Chat supportIn developmentHigh

Markdown excels when the reader is a human. JSON excels when the reader is a system.

Try This Now​

Pick one of these exercises. It takes two minutes.

  1. JSON practice: Paste five customer reviews into ChatGPT and add: "For each review, return JSON with: reviewer_name, sentiment (positive/negative/neutral), key_issue, suggested_action."
  2. XML practice: Take any long email or memo, wrap it in <context> tags, then add <task>Summarise in three bullets</task> at the bottom.
  3. Markdown practice: Ask any AI to "Compare ChatGPT, Claude, and Gemini in a table with columns: Best For, Strength, Weakness, Price."
Common Mistake

Never assume the AI will choose the right format. Always state the output shape you want as the last line of your prompt — right before you hit enter.

Good Read​