Structured Output Prompting: Getting Reliable JSON, CSV, and Tables

Structured Output Prompting: Getting Reliable JSON, CSV, and Tables

One of the most common frustrations in production prompt engineering is inconsistent output format. The model produces JSON on the first run, a prose description on the second, and a hybrid of both on the third. Structured output prompting is the set of techniques that eliminates this inconsistency and makes model outputs reliably parseable by downstream code.

Why Structure Matters in Production

When you are processing a single prompt manually, you can tolerate format variation — you just read the output and extract what you need. But in a batch pipeline processing hundreds or thousands of rows, every format deviation is a parsing error that requires manual intervention. Structured output is not a nice-to-have; it is a production requirement.

Technique 1: Explicit Schema in the Prompt

The most direct approach is to specify the exact output schema in the prompt, including field names, types, and an example.

Extract the following information from the product description and return it as JSON.

Schema:
{
  "product_name": string,
  "price_usd": number,
  "category": string,
  "in_stock": boolean,
  "key_features": string[]  // array of up to 3 features
}

Product description: {{description}}

Return only the JSON object. Do not include explanation or markdown code fences.

Extract the following information from the product description and return it as JSON.

Schema:
{
  "product_name": string,
  "price_usd": number,
  "category": string,
  "in_stock": boolean,
  "key_features": string[]  // array of up to 3 features
}

Product description: {{description}}

Return only the JSON object. Do not include explanation or markdown code fences.

The instruction "Return only the JSON object" is critical. Without it, many models wrap the JSON in a markdown code block or add a preamble like "Here is the extracted information:".

Technique 2: One-Shot Format Demonstration

Showing the model a single input-output example is often more reliable than describing the format in prose.

Extract product data as JSON.

Example:
Input: "The AeroGrip Pro Tennis Racket is available for $89.99 in the Sports category. Currently in stock. Lightweight frame, vibration dampening, and ergonomic grip."
Output: {"product_name": "AeroGrip Pro Tennis Racket", "price_usd": 89.99, "category": "Sports", "in_stock": true, "key_features": ["Lightweight frame", "Vibration dampening", "Ergonomic grip"]}

Now extract from:
{{description}}
Output:

Extract product data as JSON.

Example:
Input: "The AeroGrip Pro Tennis Racket is available for $89.99 in the Sports category. Currently in stock. Lightweight frame, vibration dampening, and ergonomic grip."
Output: {"product_name": "AeroGrip Pro Tennis Racket", "price_usd": 89.99, "category": "Sports", "in_stock": true, "key_features": ["Lightweight frame", "Vibration dampening", "Ergonomic grip"]}

Now extract from:
{{description}}
Output:

The trailing "Output:" with no content after it is a powerful signal — it tells the model to continue the pattern from the example.

Technique 3: CSV Output for Batch Processing

For tabular data, CSV is often more reliable than JSON because it has less structural overhead.

Analyse the following customer review and return a single CSV row with these columns:
sentiment,confidence_pct,primary_topic,action_required

Rules:
- sentiment: one of POSITIVE / NEGATIVE / NEUTRAL
- confidence_pct: integer 0-100
- primary_topic: one of PRODUCT / SHIPPING / SUPPORT / PRICING
- action_required: YES or NO

Do not include a header row. Do not include quotes unless the value contains a comma.

Review: {{review}}

Analyse the following customer review and return a single CSV row with these columns:
sentiment,confidence_pct,primary_topic,action_required

Rules:
- sentiment: one of POSITIVE / NEGATIVE / NEUTRAL
- confidence_pct: integer 0-100
- primary_topic: one of PRODUCT / SHIPPING / SUPPORT / PRICING
- action_required: YES or NO

Do not include a header row. Do not include quotes unless the value contains a comma.

Review: {{review}}

Technique 4: Markdown Tables

For human-readable structured output, Markdown tables work well and are easy to render.

Compare the following two products across five dimensions.
Return a Markdown table with columns: Dimension | Product A | Product B

Dimensions to compare: Price, Durability, Ease of Use, Customer Support, Warranty

Product A: {{product_a}}
Product B: {{product_b}}

Compare the following two products across five dimensions.
Return a Markdown table with columns: Dimension | Product A | Product B

Dimensions to compare: Price, Durability, Ease of Use, Customer Support, Warranty

Product A: {{product_a}}
Product B: {{product_b}}

Handling Format Failures Gracefully

Even with the best prompts, format failures occur. In production pipelines, build a validation layer that:

1. Attempts to parse the output (JSON.parse, CSV split, regex)
2. On failure, retries with a stricter prompt that includes the failed output and asks the model to fix it
3. After two failures, flags the row for human review rather than silently dropping it

Structured Output in PromptProcessor

PromptProcessor's batch export feature outputs results as CSV by default, but the structured output techniques above let you embed structured data inside each result cell. You can then post-process the CSV to extract nested JSON or split multi-column CSV rows into separate columns — making PromptProcessor a complete structured data extraction pipeline.