Production Claude Application Capstone

By the end of this lesson you will know the 12-item production-readiness checklist every Claude app should pass before ship: secrets management, eval-gated deploys, cost caps, structured logging + trace_id, retry-with-backoff, safety layer (injection + PII + refusal), feature flags for model versions + prompts, alerts on 429 + cost + refusal-rate deltas, runbooks for common incidents, docs. Anchors every remaining CLD-AI-112 hands-on lab and the final Orion capstone.

By the end of this lesson you will know the observability discipline every production Claude app needs — a stable log-record schema (trace_id, request_id, agent_role, tokens_in, tokens_out, latency_ms, cost_usd, cache_hit, refusal), trace_id propagation across every downstream call, per-request metric emission for dashboards, and the guiding rule 'what you can't measure you can't ship'. Anchors the L4 hands-on where you wrap the Anthropic client with the emitter.

By the end of this lesson you will know the four-layer safety stack every production Claude app needs — ingress prompt-injection detection (before Claude sees the input), PII redaction before logging (so trace records don't leak SSNs or emails), refusal patterns for out-of-scope requests, and a post-hoc content-safety check on Claude's output — plus where each layer catches which class of attack. Sets up L6 hands-on where you wire all four layers into the Orion app.

By the end of this lesson you will know how to deploy a Claude app safely — hash-based feature flags that route a fraction of traffic to a canary variant (5% → 25% → 100%), per-variant metric aggregation (latency, cost, refusal rate) so you can compare before promoting, instant rollback via a flag flip when the canary breaches thresholds, and the promote/rollback decision rule as a machine-parseable scorecard. Sets up L8 hands-on where you ship a real canary.

By the end of this lesson you will know the five most common Claude-app incidents and how to respond: cost spikes (a runaway loop or a bad prompt burning tokens), refusal-rate spikes (a safety-layer overshoot after a prompt change), 429 storms (rate limit exceeded, backpressure your ingress), prompt-injection outbreaks (adversarial traffic surge, tighten ingress filter), upstream Anthropic outages (fail closed vs degrade). Each incident gets a checklist runbook. Sets up the L10 capstone.