Session 5.5: Structured Output: Getting Predictable Formats

Course → Module 5: Prompt Engineering

Session 5 of 10

From Free-Form to Predictable

When AI generates free-form text, every output is different. The structure varies. The section order changes. The formatting is inconsistent. This is fine for one-off writing. It is unacceptable for production, where your next pipeline stage expects a specific format and your publishing system requires a specific structure.

Structured output means specifying the exact format of the AI's response. Instead of "write a product review," you specify "return a JSON object with fields: title (string), summary (100 words max), pros (array of strings), cons (array of strings), rating (integer 1-5), and recommendation (string)." The AI produces data you can parse, not prose you have to interpret.

Structured output turns AI from a creative tool into a production component. When the output format is predictable, you can build automated pipelines around it. Formatting scripts, quality checks, and publishing systems all require knowing what the output looks like before they see it.

Three Levels of Structure

Structure exists on a spectrum. You choose the level based on what your pipeline needs.

Level	Format	Example Use	Parseable by Code
Loose	Prose with requested headers	Blog posts, articles	Partially (header extraction)
Template	Markdown with fixed sections	Reviews, comparisons, reports	Yes (section parsing)
Strict	JSON, XML, or YAML	Product data, metadata, structured content	Yes (native parsing)

Template-Level Structure

For most content production, template-level structure is the sweet spot. You define a markdown template with required sections and the AI fills it in. The template ensures consistency while allowing the AI freedom within each section.

A content template for a product review might look like this:

## Overview
[2-3 sentences summarizing the product]

## What It Does Well
[3-4 specific strengths with examples]

## Where It Falls Short
[2-3 specific weaknesses with examples]

## Who Should Use It
[1-2 paragraphs defining the ideal user]

## Verdict
[1 paragraph, clear recommendation]

Include this template in your prompt with the instruction: "Fill in each section of the template below. Do not add, remove, or reorder sections." The AI produces output that matches your template exactly, making downstream processing predictable.

graph LR A["Template
(fixed structure)"] --> B["AI fills sections"] B --> C["Consistent output"] C --> D["Automated formatting"] D --> E["Publishing"] F["No template
(free-form)"] --> G["AI decides structure"] G --> H["Inconsistent output"] H --> I["Manual reformatting"] I --> E style A fill:#222221,stroke:#6b8f71,color:#ede9e3 style C fill:#222221,stroke:#6b8f71,color:#ede9e3 style F fill:#222221,stroke:#c47a5a,color:#ede9e3 style H fill:#222221,stroke:#c47a5a,color:#ede9e3

Strict JSON Output

For data-heavy content or pipeline components that feed into code, JSON is the right choice. Modern AI APIs support structured output natively. Claude, GPT, and Gemini can all return valid JSON when instructed properly.

The key is providing a schema. Do not just say "return JSON." Specify the exact fields, types, and constraints. Many APIs now support response schemas that enforce the structure at the API level, guaranteeing valid output.

Testing Structural Consistency

A structured prompt is only useful if it produces consistent structure. Test by running the same prompt five times and comparing the outputs. Check: do all five have the same sections? Are the sections in the same order? Do all JSON outputs parse without errors?

If consistency is below 4 out of 5, the prompt needs refinement. Common fixes: more explicit section markers, stricter formatting instructions, or adding "Do not deviate from this structure" as a constraint.

Consistency Score	Diagnosis	Action
5/5	Prompt is production-ready	Deploy to pipeline
4/5	Minor drift in one run	Add explicit constraints, re-test
3/5 or below	Structure is not enforced	Rewrite prompt with stricter formatting rules

Assignment

Define a content template for your most common content type. Express it as a markdown structure with section headers and placeholder descriptions. Create a prompt that instructs the AI to fill in the template exactly. Test it five times. How consistent is the structure? Document any deviations and refine the prompt until you achieve 5/5 structural consistency.

Structured Output: Getting Predictable Formats