Show Claude the Expected Output and Let It Write the Code Backward

Describing what code should do is harder than showing what it should produce. When you have a clear picture of the output, show Claude the example and let it reverse-engineer the implementation.

"Write an API endpoint that returns this JSON:
{
  'user': { 'id': 1, 'name': 'Jane', 'plan': 'pro' },
  'usage': { 'api_calls': 1247, 'limit': 5000, 'reset_date': '2024-04-01' },
  'features': ['analytics', 'webhooks', 'custom-domains']
}"

Claude reads the structure, figures out where the data comes from in your codebase, and writes the endpoint — controller, route, resource/transformer, and any queries needed to assemble that response.

This works for any kind of output:

# CLI tool output
"Write a command that produces output like this:
┌──────────┬────────┬─────────┐
│ Service  │ Status │ Latency │
├──────────┼────────┼─────────┤
│ Database │ ✓ up   │ 2ms     │
│ Redis    │ ✓ up   │ 1ms     │
│ S3       │ ✗ down │ timeout │
└──────────┴────────┴─────────┘"

# Email template
"Write the code that generates an email that looks like this:
[paste the HTML or text you want]"

# Log format
"Write a logging middleware that produces lines like:
[2024-03-15 14:30:22] GET /api/users 200 45ms user_id=123"

# Data transformation
"Write a function that takes this input:
[paste input]
And produces this output:
[paste output]"

For complex transformations, give multiple input/output pairs:

"Write a function where:
- Input: 'john.doe@example.com' → Output: 'JD'
- Input: 'alice.wonderland@test.io' → Output: 'AW'
- Input: 'bob@company.com' → Output: 'B'"

Claude infers the pattern from your examples and writes the function — including edge cases you might not have thought of.

When you know what the output should look like, showing is faster than telling — Claude works backward from examples better than forward from descriptions.

via Claude Code