A developer instructs Claude in the system prompt to "always return valid JSON with the following fields." At a volume of 10,000 requests per day, the pipeline occasionally receives malformed JSON (roughly 1% of responses). What is the most production-appropriate fix?

Question

Accepted Answer

B. Implement a three-layer pipeline: `tool_use` schema for structural enforcement, programmatic semantic validation, and a retry loop that sends validation errors back to Claude for correction. Prompt-only JSON enforcement is an anti-pattern. Even the best prompts are probabilistic; at 10,000 requests/day, a 1% failure rate means 100 errors per day. The correct approach is the three-layer pipeline: (1) `tool_use` JSON schema for guaranteed structural correctness, (2) programmatic semantic validation for business rule compliance, (3) retry loop that sends invalid output back with error details for correction. Option A strengthens the weakest solution. Option D silently discards failures.

A developer instructs Claude in the system prompt to "always return valid JSON with the following fields." At a volume of 10,000 requests per day, the pipeline occasionally receives malformed JSON (roughly 1% of responses). What is the most production-appropriate fix?

Related Questions

Discussion