Module 2 • Lesson 5

Output Formatting: JSON, Tables, Markdown & More

Control how the AI formats its response — structured data, tables, lists, and machine-readable formats.

Output Formatting: JSON, Tables, Markdown & More

Controlling HOW the AI formats its response is just as important as controlling WHAT it says. For production systems, structured output is non-negotiable — downstream code needs to parse AI responses reliably.

Key Concepts

  • Structured Output: Machine-readable formats like JSON, XML, CSV

    • Controlling HOW the AI formats its response is just as important as controlling WHAT it says. For production systems, structured output is non-negotiable — downstream code needs to parse AI responses reliably.

    Key Concepts

  • Structured Output: Machine-readable formats like JSON, XML, CSV
  • Semi-Structured Output: Human-readable but organized formats like Markdown, tables, lists
  • Format Enforcement: Techniques to guarantee the model outputs valid structured data

    • Common Output Formats

    FormatBest ForParsing Difficulty
    JSONAPI responses, data pipelinesEasy (native parsing)
    MarkdownDocumentation, reportsMedium
    CSVSpreadsheet data, bulk analysisEasy
    XMLLegacy systems, SOAP APIsMedium
    YAMLConfiguration filesEasy
    Plain textHuman consumptionN/A
    HTMLWeb contentMedium

    Techniques for Reliable JSON Output

  • Schema Definition: Provide the exact JSON schema you expect
  • Assistant Prefix: Start the response with { to force JSON mode
  • Few-Shot Examples: Show 1-2 examples of the desired JSON structure
  • Validation Instructions: "Ensure the output is valid JSON. Do not include any text before or after the JSON object."

    • Example: Forcing JSON Output

    "Analyze the following customer review and output a JSON object with this exact schema:

    `json

    {

    "sentiment": "positive" | "negative" | "neutral",

    "confidence": 0.0 to 1.0,

    "key_topics": ["string"],

    "summary": "string (max 50 words)"

    }

    `

    Output ONLY the JSON object. No explanations, no markdown code fences."

    Table Generation

    "Create a comparison table with these exact columns:

    Use checkmarks (✅) and crosses (❌) for boolean features. Include at least 8 rows."

    Markdown Reports

    "Generate a technical report in this exact Markdown structure:

    Executive Summary (2-3 sentences)

    Key Findings (numbered list, 5 items)

    Data Analysis (include a table)

    Recommendations (bullet points)

    Next Steps (action items with owners)"

    Format Enforcement Strategies

  • API-Level: OpenAI's response_format: { type: "json_object" } forces JSON
  • Grammar-Based: Tools like llama.cpp support BNF grammars for exact format control
  • Validation Loop: Generate → parse → if invalid, regenerate with error feedback
  • Few-Shot + Schema: Most reliable prompt-level technique

    • Common Mistakes

  • Not specifying the EXACT format (leading to inconsistent outputs across runs)
  • Asking for JSON but not specifying the schema (model invents its own keys)
  • Mixing format instructions with content instructions (separate them clearly)
  • Not validating output format programmatically before using it downstream
  • Forgetting that complex formats (nested JSON) need more explicit instructions

    • Read to the end of the lesson to mark it as read.