Structured Prompting: XML, JSON, and Frameworks

The biggest leap in prompt quality comes from structure. Not better words. Not longer prompts. Structure.

A structured prompt tells the AI exactly where each piece of information lives: what’s an instruction, what’s context, what’s an example, and what format the output should take. This eliminates the ambiguity that causes inconsistent results.

XML Tags: The Foundation

XML tags are the most versatile structuring tool. They work across all major AI models, and Claude was specifically trained with XML in its training data.

Basic Structure

<role>You are a financial analyst specializing in SaaS metrics.</role>

<task>Analyze the following quarterly data and provide insights.</task>

<data>
Q1 Revenue: $2.1M, Churn: 4.2%, NRR: 112%
Q2 Revenue: $2.4M, Churn: 3.8%, NRR: 118%
Q3 Revenue: $2.9M, Churn: 3.1%, NRR: 125%
</data>

<output_format>
1. Executive summary (2 sentences)
2. Three key metrics trends (with % changes)
3. One risk to flag
4. One opportunity to highlight
</output_format>

Compare this to the unstructured version: “You’re a financial analyst. Look at Q1: $2.1M revenue, 4.2% churn, 112% NRR. Q2: $2.4M, 3.8%, 118%. Q3: $2.9M, 3.1%, 125%. Give me insights about the trends and what we should focus on.”

Both contain the same information. The structured version produces more consistent, predictable results.

Common XML Tags

✅ Quick Check: You have a prompt with instructions, context, user input, and expected output format — all in one paragraph. What’s the risk? (Answer: The AI might misinterpret which parts are instructions vs. context vs. input. For example, if your context includes “always double-check calculations,” the AI might treat that as an instruction to it, when it was actually describing a company policy. XML tags prevent this: <context> content is read as background information, <instructions> content is treated as commands to follow.)

JSON Schema for Output Control

When you need structured output that code can parse, specify a JSON schema:

<task>Extract contact information from the following email.</task>

<output_format>
Return a JSON object with these fields:
{
  "name": "Full name of the sender",
  "email": "Email address",
  "company": "Company name (or null if not mentioned)",
  "request": "One-sentence summary of what they want",
  "urgency": "low | medium | high"
}
</output_format>

<input>
Hi, I'm Marcus Chen from DataFlow Inc. We urgently need help with our
API integration — our production system is down and we need someone
on a call within the hour. My direct line is marcus@dataflow.io.
</input>

Expected output:

{
  "name": "Marcus Chen",
  "email": "marcus@dataflow.io",
  "company": "DataFlow Inc.",
  "request": "Urgent help needed with API integration due to production system outage",
  "urgency": "high"
}

The COSTAR Framework

COSTAR provides a complete prompting framework that works for any task:

COSTAR in Practice

<context>
Our company (TechCorp) sells project management software. We just added
AI-powered task prioritization. Beta testers loved it (92% satisfaction).
We launch publicly next Tuesday.
</context>

<objective>Write the product announcement email.</objective>

<style>Professional but warm. Short sentences. Similar to how Notion
communicates product updates.</style>

<tone>Confident and excited, but not hyperbolic. No "revolutionary"
or "game-changing" — let the feature speak for itself.</tone>

<audience>Current Pro plan customers who have been using the product
for 3+ months. They're familiar with the interface.</audience>

<response>
- Subject line (under 50 characters)
- Pre-header text (under 100 characters)
- Email body (3 paragraphs, each 2-3 sentences)
- One CTA button with text
</response>

✅ Quick Check: Two prompts produce the same content but one uses COSTAR and the other is a single paragraph. In which scenario does COSTAR matter MORE? (Answer: When the prompt is reusable — used by a team, embedded in an application, or run hundreds of times. A one-off personal prompt can be informal. But a prompt that runs in production or is shared across a team needs to produce the same quality regardless of who runs it. COSTAR’s explicit components prevent drift: every team member uses the same context, style, and output spec.)

Combining Structures

The most powerful prompts combine XML tags with COSTAR principles and JSON output:

<system>
You are a customer feedback analyst. You process product reviews
and extract structured insights.
</system>

<context>We sell a productivity app. Reviews are from the App Store.</context>

<task>
Analyze the following review and extract structured data.
</task>

<output_format>
{
  "sentiment": "positive | negative | mixed",
  "topics": ["list of topics mentioned"],
  "feature_requests": ["any features the user wants"],
  "bugs_mentioned": ["any bugs reported"],
  "summary": "one-sentence summary"
}
</output_format>

<input>
{{REVIEW_TEXT}}
</input>

The {{REVIEW_TEXT}} placeholder makes this a template — you can swap in different reviews while the structure stays constant. This is how production prompts work: define the structure once, vary only the input.

Practice Exercise

1. Take a prompt you use regularly and restructure it with XML tags
2. Apply the COSTAR framework to a complex task (e.g., writing a proposal, analyzing data)
3. Create a JSON output schema for extracting information from unstructured text
4. Compare the output of your old prompt vs. the structured version — note consistency differences

Key Takeaways

• Structure beats length — organized prompts produce more consistent results than long, unstructured ones
• XML tags create clear boundaries between instructions, context, examples, and output format
• JSON schemas ensure output is parseable by code — essential for production applications
• COSTAR (Context, Objective, Style, Tone, Audience, Response) covers all dimensions of a complete prompt
• Templates with placeholders ({{VARIABLE}}) make structured prompts reusable across different inputs
• Claude was trained with XML tags; all major models respond well to structured prompts

Up Next

In the next lesson, you’ll learn reasoning techniques — chain-of-thought, tree-of-thought, and self-consistency — that make AI significantly better at complex, multi-step problems.