Why This Matters

Beginner output shaping (BE-18) is 'output as JSON'. Intermediate output schemas give Copilot a precise shape to fill — and you a programmatic way to validate. This unlocks automation: Copilot output → parser → next step.

Concept Explained

An output schema is a typed shape specification: field names, types, optionality, allowed values, nested structures. Specifying it precisely in the prompt produces consistent, parseable output.

How It Works

Define schema in TypeScript or JSON Schema. Include in prompt. Specify 'output as JSON matching this schema, no prose, no markdown'. Validate with a parser (zod, pydantic, json-schema). On violation, retry or fall back.

Step-by-Step Tutorial

1. Define schema

2. Include in prompt

3. Add 'no prose'

4. Validate programmatically

5. Handle violations

Real-World Example

A team built a 'bug triage' pipeline: paste stack trace → Copilot outputs JSON {severity, category, likely_cause, suggested_fix} → script files issue in tracker. Schema validation caught 5% of outputs that didn't match; those got retried with stricter prompts. Pipeline processed 100 bugs/hour with 95% valid output.

Example Prompts / Commands / Code

Output as JSON matching this TypeScript type:

type BugAnalysis = {
  severity: 'high' | 'medium' | 'low';
  category: 'runtime' | 'logic' | 'security' | 'performance' | 'other';
  likely_cause: string;          // 1-2 sentences
  suggested_fix: string;         // 1-2 sentences, no code
  affected_files: string[];      // file paths
  confidence: number;            // 0.0 to 1.0
};

Output ONLY the JSON. No markdown, no surrounding text, no explanation.

import { z } from 'zod';

const BugAnalysisSchema = z.object({
  severity: z.enum(['high', 'medium', 'low']),
  category: z.enum(['runtime', 'logic', 'security', 'performance', 'other']),
  likely_cause: z.string().min(10).max(200),
  suggested_fix: z.string().min(10).max(200),
  affected_files: z.array(z.string()).min(1),
  confidence: z.number().min(0).max(1),
});

// Usage:
try {
  const parsed = BugAnalysisSchema.parse(JSON.parse(copilotOutput));
  // ... use parsed
} catch (e) {
  // Schema violation — retry or fallback
}

Common Mistakes

• Vague schema ('output as JSON object') — produces inconsistent shapes.
• No validation — silent downstream failures.
• Allowing prose around JSON — parser breaks.
• No retry strategy for violations.

Best Practices

• Define schema precisely (TypeScript type or JSON Schema).
• End prompt with 'Output ONLY the JSON, no prose'.
• Validate with zod/pydantic/json-schema.
• Retry on violation with stricter prompt.
• Build schema-driven pipelines for automation.

Troubleshooting

Practical Exercise

Professional Challenge

Frequently Asked Questions

Further Reading

• Prompt engineering

Official References

• Prompt engineering

Related lessons: BE-18, IN-04