How would you repair invalid structured model output?

Uses constrained output where possible, deterministic parsing and validation, bounded targeted repair, and safe fallback.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

I prefer provider-supported schema-constrained output. Otherwise I parse and validate the response, classify the failure, and retry only when a bounded repair is likely to help. A repair prompt includes the validation error and the original required contract without asking the model to change unrelated values. After the retry budget, I use a safe fallback or request review.

Full answer below

Practice focus

Whether repair is controlled and observable rather than an endless reprompt loop that hides contract defects.

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
The first choice is a stronger output contract. Repair exists for recoverable failures, not to make an unsuitable task appear reliable.
I prefer provider-supported schema-constrained output. Otherwise I parse and validate the response, classify the failure, and retry only when a bounded repair is likely to help. A repair prompt includes the validation error and the original required contract without asking the model to change unrelated values. After the retry budget, I use a safe fallback or request review.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
I do not fix JSON with broad regex substitutions that can silently alter content. Deterministic repairs are limited to cases with unambiguous semantics, such as removing a known wrapper before parsing. Missing required evidence, an invented identifier, or a policy refusal is not a formatting repair. I record first-pass and repaired success separately.

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

Retrying indefinitely until something parses.
Using regex to guess the meaning of malformed JSON.
Treating semantic errors as format errors.
Counting repaired responses as clean first-pass success.

How would you repair invalid structured model output?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue practice

Try a variant

Tech stacks

Skill areas

Human editorial direction

Was this drill useful?

Found this answer useful?