Architecture scenarioHard~10 minPrompting & Structured Output / 11 of 20

Interview question

How would you evolve a structured-output schema safely?

Treats the model-facing schema and downstream consumer contract as a versioned API with compatibility and rollout concerns.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

I treat the schema as a versioned API contract. Adding an optional field may be compatible for tolerant consumers, while renaming a field, changing an enum, or making a field required is usually breaking. I update consumers and evaluation cases deliberately rather than assuming the model can bridge incompatible versions.

Full answer below

Practice focus

Whether ordinary API evolution discipline is applied to AI-generated payloads.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
Changing the schema changes both constrained generation and every consumer that parses or acts on the result.
I treat the schema as a versioned API contract. Adding an optional field may be compatible for tolerant consumers, while renaming a field, changing an enum, or making a field required is usually breaking. I update consumers and evaluation cases deliberately rather than assuming the model can bridge incompatible versions.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
For a breaking change I introduce a new schema version, run both paths on representative inputs, and migrate consumers before removing the old one. I keep prompt instructions and examples aligned with the selected version. I also account for provider schema compilation or caching so a rollout does not create unexpected cold latency.
Finish
Advanced answer
Optional deeper angle for production judgment and nuance.
I prefer additive optional fields when they preserve meaning and reserve breaking changes for a new contract version. I record that schema version with stored outputs and events, especially when asynchronous consumers may process them later. Rollback restores a compatible model-prompt-schema-consumer set, not only old prompt text. Contract tests cover both syntactic payloads and semantic cases such as new enum values, partial results, and old readers.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Changing enums without checking downstream consumers.
Assuming an optional field is harmless to strict deserializers.
Updating examples but not the formal schema.
Storing outputs without their schema version.

Red flagsCandidateRed flagsSignals that an interviewer may treat as risky.Open details

Details

A rollback can reintroduce payloads current consumers cannot read.
Prompt and schema versions can drift independently.
Asynchronous events have no contract version.

Follow-up prompts

Quick interviewer prompts inside this drill.

Follow-up questionsCandidateFollow-up questionsLikely next questions after the main answer.Open details

Details

Treat these as the next interviewer turn. Answer one out loud, then save the gap in your practice notes.

Which schema changes are breaking?
How would you support old asynchronous consumers?
What must roll back together?

Continue the section

Next in Prompting & Structured Output: Output vs tool calling

Step 12 of 20

When should you use structured output versus tool calling?

Practice route

Read core answer Review checklist Follow-up prompts Track progress

Finished this drill?

Practice freely. Sign in only when you want to save completion, confidence, notes, or bookmarks.

New to Aporeon? Create account

Drill details

Human editorial direction

Aporeon is shaped by Aleksandar Tomovski, a software developer with experience on both sides of technical interviews. Content is reviewed for accuracy, natural spoken delivery, useful depth, and honest trade-offs.

About the creator Editorial standards

Was this drill useful?

Help improve the interview library.

Found this answer useful?

Say thanks with a standalone, one-time $5 contribution. No account required.

No subscription or perks.Choose another amount

Opening drill

Architecture scenarioHard~10 minPrompting & Structured Output / 11 of 20

Interview question

How would you evolve a structured-output schema safely?

Treats the model-facing schema and downstream consumer contract as a versioned API with compatibility and rollout concerns.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

Full answer below

Practice focus

Whether ordinary API evolution discipline is applied to AI-generated payloads.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
Changing the schema changes both constrained generation and every consumer that parses or acts on the result.
I treat the schema as a versioned API contract. Adding an optional field may be compatible for tolerant consumers, while renaming a field, changing an enum, or making a field required is usually breaking. I update consumers and evaluation cases deliberately rather than assuming the model can bridge incompatible versions.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
For a breaking change I introduce a new schema version, run both paths on representative inputs, and migrate consumers before removing the old one. I keep prompt instructions and examples aligned with the selected version. I also account for provider schema compilation or caching so a rollout does not create unexpected cold latency.
Finish
Advanced answer
Optional deeper angle for production judgment and nuance.
I prefer additive optional fields when they preserve meaning and reserve breaking changes for a new contract version. I record that schema version with stored outputs and events, especially when asynchronous consumers may process them later. Rollback restores a compatible model-prompt-schema-consumer set, not only old prompt text. Contract tests cover both syntactic payloads and semantic cases such as new enum values, partial results, and old readers.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Changing enums without checking downstream consumers.
Assuming an optional field is harmless to strict deserializers.
Updating examples but not the formal schema.
Storing outputs without their schema version.

Red flagsCandidateRed flagsSignals that an interviewer may treat as risky.Open details

Details

A rollback can reintroduce payloads current consumers cannot read.
Prompt and schema versions can drift independently.
Asynchronous events have no contract version.

Follow-up prompts

Quick interviewer prompts inside this drill.

Follow-up questionsCandidateFollow-up questionsLikely next questions after the main answer.Open details

Details

Treat these as the next interviewer turn. Answer one out loud, then save the gap in your practice notes.

Which schema changes are breaking?
How would you support old asynchronous consumers?
What must roll back together?

Continue the section

Next in Prompting & Structured Output: Output vs tool calling

Step 12 of 20

When should you use structured output versus tool calling?

How would you evolve a structured-output schema safely?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue the section

How would you evolve a structured-output schema safely?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue the section