What problem does it solve? Wiring an LLM API into a real application introduces failure modes that normal API integration does not: nondeterministic output, per-token billing, long streaming responses, rate limits, and untrusted generated content. This Skill provides the production playbook for handling all of them before the invoice or the outage teaches you. ## Core Features & Use Cases - Client Layer Hardening: Generation-sized timeouts, jittered retries on 429/5xx, client-side rate limiting, pinned model versions, and server-side key management. - Cost Control: Per-call token metering tagged by feature and tenant, prompt caching on stable prefixes, response caching, per-task max_tokens, and enforced budget caps. - Safety & Resilience: Schema validation with corrective retry, sandboxing of LLM output, circuit breakers, cross-provider fallback plans, and monitoring dashboards for latency, errors, and cost anomalies. - Use Case: You are shipping a support-chat feature on Claude. Use this Skill to set up streaming with per-chunk timeouts, cache the system prompt, meter cost per tenant, and define a fallback model for provider outages. ## Quick Start Use the llm-integration skill to review my LLM API wrapper and produce a production ship checklist covering streaming, cost metering, and fallback behavior.
