💡 Structured Outputs Beta — Native JSON Schema Enforcement in the Claude API

Anthropic has launched a Structured Outputs beta for the Claude API, introducing native JSON schema enforcement that guarantees Claude's response conforms to a developer-specified schema without requiring post-processing validation. The feature, accessible by passing a response_format parameter with a JSON schema definition, uses a constrained decoding approach during generation to ensure schema compliance at the token level rather than relying solely on instruction-following.

# Python — native structured output
response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "extraction_result",
            "schema": {
                "type": "object",
                "properties": {
                    "entities": {"type": "array", "items": {"type": "string"}},
                    "sentiment": {"type": "string", "enum": ["positive","negative","neutral"]},
                    "confidence": {"type": "number"}
                },
                "required": ["entities", "sentiment", "confidence"]
            }
        }
    },
    messages=[{"role": "user", "content": "Analyse: ..."}]
)

The beta is available to all API customers on Sonnet 4.5 and Opus 4.5. Anthropic notes that structured outputs mode is incompatible with extended thinking in this initial release; the combination is on the roadmap for a later update.

💡 System Prompt Templates Library — 50 Starter Configurations for Common Patterns

Anthropic has published a System Prompt Templates Library containing 50 curated starter system prompts for the most common operator deployment patterns. The library is intended to reduce the time from first API call to production-ready system prompt by providing well-tested starting points that encode Anthropic's current best practices for each category. Each template is fully editable and accompanied by inline comments explaining the rationale for specific instructions.

Template categories included

• Customer support — templates for B2C support assistants, with variants for SaaS, e-commerce, and financial services contexts including appropriate escalation instructions
• Code assistance — developer-facing coding assistants with variants optimised for code review, debugging, documentation generation, and test writing
• Document analysis — templates for contract review, research summarisation, and multi-document comparison use cases
• Data extraction — paired with the new structured outputs feature; templates that specify extraction schemas alongside the system prompt instructions
• Agentic task execution — templates with explicit minimal-footprint and confirmation-before-action instructions for tool-using agent deployments