Output Control: JSON, Citations, and Caching

Output Control: JSON, Citations, and Caching

Many production issues are output-side concerns: stable structure, traceable citations, predictable cost, and high-throughput batch processing.

Recommended notebooks

1) Stable structured output

Prompting patterns for consistent JSON output

Structured JSON + tool use

Use tools as a schema contract

2) Citations and traceability

Traceable citations across document types

3) Cost & latency: prompt caching

Cache-friendly prompt construction patterns

Speculative prompt caching

Speculative caching comparisons

4) Throughput: batch processing

Message Batches API

High-throughput batch workflows

5) Safety: moderation filtering

Moderation filter

Build a moderation filter with Claude

6) Long outputs (optional)

Sampling past max tokens

Strategies for very long outputs

7) Prompt engineering workflow (optional)

Prompt templates + testing loop

Rules of thumb

Treat format as an API contract: validate JSON and retry; don’t rely on prompts alone.
Caching is a measured tradeoff: quantify hit-rate and savings before shipping.
Batch needs observability: queueing, failures, retries, and result collection.