Anthropic API workflows with clear tensor boundaries

Choose the documented message boundary

Begin with the task you need to complete. Summarizing supplied passages or proposing a category can fit a message interface. Extracting a model’s hidden activations requires an implementation that explicitly exposes those values. A number array written inside a prompt remains message content; its appearance does not create a native tensor contract.

Keep the task object separate from the provider request. Your application might accept a document identifier and category set, then build a message containing only the relevant evidence. Preserve model selection and request metadata in the adapter. This lets the rest of the application work with stable task types while the provider-specific representation remains in one place. Review the boundary whenever you add a new input modality or output requirement.

Give the adapter a small, testable job

The provider adapter should translate validated application input into a request and translate the observed response into a controlled result. Keep numerical preprocessing and document retrieval in clearly named components.

• Validate input size and required fields before making a request.
• Attach stable identifiers to supplied evidence so the response can reference it.
• Handle supported content blocks explicitly instead of assuming one plain string.
• Record usage and stop metadata from the response rather than asking the model to invent them.
• Use a deadline and bounded retry policy for transport failures.

Store sanitized fixtures for representative responses. They let you exercise parsing and failure handling independently of a live generation call, while a separate integration check verifies the actual provider interface.

Verify structure and the proposed decision

Define the answer contract before writing the prompt. For classification, decide which labels are allowed, whether the model may return unknown and what evidence is required. A parser can establish that an object is valid JSON; task validation must establish that its fields make sense for the supplied record.

Reject references to evidence that was never provided. Flag unsupported categories and incomplete responses for a controlled retry or review, according to the failure type. Keep a reviewed set of normal, ambiguous and adversarial examples to compare changes. If retrieval omitted the decisive passage, changing the output format will not fix the answer. Diagnose the stage that failed, then update the corresponding component and rerun the relevant examples before promoting the change.