50 mins lesson duration•10 mins read

Lesson 18: Structured Outputs & Reusable Templates

Force consistent, machine-readable outputs with JSON, tables, and XML tags — and stop rewriting prompts from scratch.

Why Structured Outputs Matter

Free-form answers are useless when your pipeline needs to parse them. The fix: tell the model exactly what shape the output must take — and show it an example.

1. Explicit Format Instructions

Analyze this feedback and return JSON with EXACTLY this structure:
{
  "sentiment": "positive | negative | neutral",
  "key_themes": ["..."],
  "urgency": "high | medium | low"
}

Feedback: "Your product broke after 2 days and support hasn't replied!"

The schema doubles as instructions: keys, allowed values, and types are all specified.

2. The JSON Output Template

Return ONLY valid JSON, no markdown fences:
{"summary": "...", "findings": [...], "confidence": 0.0-1.0}

Then validate with json.loads() in your code — never assume the model got it right (use JSON mode / structured outputs when the API offers it).

3. Markdown Tables

For comparisons and structured data:

Compare React and Vue in a table:
| Aspect | React | Vue |
|---|---|---|
| Learning curve | | |
| Performance | | |
| Ecosystem | | |

4. XML Tags for Long, Multi-Part Outputs

Write a blog post about AI in healthcare.

<outline>
[outline here]
</outline>

<introduction>
[1 paragraph]
</introduction>

<conclusion>
[2 sentences]
</conclusion>

Tags let you (and the model) keep track of many sections — and let your code extract each part with a simple regex.

5. Delimiters Keep Parts Separate

Delimiter	Use
```	Code blocks, long data
---	Section separation
\|\|\|	Separating examples
XML tags	Nested/multi-part structure

Golden rule: never put user-provided text directly next to instructions — wrap it in delimiters (also a prompt-injection defense).

Reusable Prompt Templates

A template = a prompt with placeholder variables:

You are a [ROLE] specializing in [DOMAIN].

TASK: Create a [CONTENT_TYPE] about [TOPIC].

CONTEXT:
- Audience: [AUDIENCE]
- Tone: [TONE]
- Goal: [GOAL]

OUTPUT FORMAT: [FORMAT]

EXAMPLE: [SHOW_ONE]

Store templates in code with variables, and fill them per request:

template.format(role="SEO writer", topic="local SEO", audience="plumbers")

Template Library Ideas

Content brief generator
Data-analysis report (summary → findings → recommendations)
Code review (correctness, style, security)
Meeting notes (decisions, action items, owners)
Customer reply drafts (tone-preserving)

Key Takeaways

Specify the exact output shape (JSON/table/XML) and show an example.
Validate structured output in code — JSON mode when available.
Delimiters separate data from instructions and resist injection.
Templates make your best prompts reusable; version them like code.

Next up: Prompting for agents — system prompts, tool schemas, and multi-step workflows.

Interactive Lesson Code Snippet

# A reusable prompt template + JSON output validation in code
template = """You are a {role} specializing in {domain}.

Analyze the following input and return ONLY valid JSON matching:
{{
  "summary": "string",
  "top_findings": ["string"],
  "recommendation": "string"
}}

INPUT: {user_input}"""

prompt = template.format(
    role="customer success analyst",
    domain="SaaS churn",
    user_input='"We lost 40 customers this month, mostly from the starter plan."',
)

print(prompt)

# After the model replies, ALWAYS validate before trusting:
import json
model_reply = '{"summary": "Starter-plan churn is high.", "top_findings": ["40 customers lost"], "recommendation": "Investigate onboarding for starter plan."}'
try:
    parsed = json.loads(model_reply)
    print(f"\nValid JSON! Keys: {list(parsed.keys())}")
except json.JSONDecodeError:
    print("\nInvalid JSON - retry or repair the output.")

Language: python

Lesson Code (Python)

# A reusable prompt template + JSON output validation in code
template = """You are a {role} specializing in {domain}.

Analyze the following input and return ONLY valid JSON matching:
{{
  "summary": "string",
  "top_findings": ["string"],
  "recommendation": "string"
}}

INPUT: {user_input}"""

prompt = template.format(
    role="customer success analyst",
    domain="SaaS churn",
    user_input='"We lost 40 customers this month, mostly from the starter plan."',
)

print(prompt)

# After the model replies, ALWAYS validate before trusting:
import json
model_reply = '{"summary": "Starter-plan churn is high.", "top_findings": ["40 customers lost"], "recommendation": "Investigate onboarding for starter plan."}'
try:
    parsed = json.loads(model_reply)
    print(f"\nValid JSON! Keys: {list(parsed.keys())}")
except json.JSONDecodeError:
    print("\nInvalid JSON - retry or repair the output.")

Console Output

You are a customer success analyst specializing in SaaS churn.

Analyze the following input and return ONLY valid JSON matching:
{
  "summary": "string",
  "top_findings": ["string"],
  "recommendation": "string"
}

INPUT: "We lost 40 customers this month, mostly from the starter plan."

Valid JSON! Keys: ['summary', 'top_findings', 'recommendation']

Code Visualization Tips

Draw the JSON schema as a form with fields and allowed values the model must 'fill in'.
Diagram the template as a fill-in-the-blanks card with variables as empty slots.
Color-code a structured-output prompt: instructions vs. example schema vs. real data.

Professional Tips & Tricks

Add 'no markdown fences, no explanations, JSON only' — models love wrapping JSON in prose.
Use the API's structured output / JSON mode when available; fall back to prompts elsewhere.
Wrap any user-supplied data in <user_input>…</user_input> tags to resist prompt injection.

Python Code Judge & Practice Arena

LeetCode Style

Run real Python 3.12 WebAssembly code directly in your browser against automated test suites.

Solved:0 / 1

0 / 20 XP

Challenges:

Problem 1 of 1

Design a JSON Prompt

Medium+20 XP

Write a prompt that extracts {product, price, currency} from a shopping message and returns clean JSON. Include the schema and the 'JSON only' rule.

main.pyPython 3.12 (WASM)

# Design a JSON Prompt
# Write your solution below

def solve():
    pass

Press Run Code to test or Submit to verify test cases

Up next · Continue learning

Prompting for Agentic Workflows

System prompts for agents, tool-use instructions, prompt chaining, and evaluation — production prompting for AI tools.

11 mins read60 mins

Start next lesson

Previous: Zero-Shot, Few-Shot & Chain-of-Thought Next: Prompting for Agentic Workflows

Made withbyAmol Shukla·amolshukla.online