Anthropic Production Checklist

Overview

Complete checklist for deploying Claude API integrations to production with reliability, observability, and cost controls.

Pre-Launch Checklist

Authentication & Keys

• Production API key from dedicated Workspace
• Key stored in secret manager (not env files on servers)
• Key rotation procedure documented and tested
• Separate keys for each environment (dev/staging/prod)

Error Handling

• All 5 error types handled: authentication_error, invalid_request_error, rate_limit_error, api_error, overloaded_error
• SDK maxRetries set (recommended: 3-5 for production)
• Custom error logging with request-id captured
• Circuit breaker for sustained API failures

Rate Limits & Cost

• Usage tier verified at console.anthropic.com
• Application-level rate limiting implemented
• Cost alerts configured (monthly spend caps)
• Model selection optimized (Haiku for simple tasks, Sonnet for complex)
• max_tokens set to realistic values (not inflated)
• Prompt caching enabled for repeated system prompts

Reliability

• Timeout configured (timeout parameter, recommended 60-120s)
• Graceful degradation when API is unavailable
• Health check endpoint tests API connectivity

async def health_check():
    try:
        # Use token counting as a cheap health probe (no generation cost)
        count = client.messages.count_tokens(
            model="claude-haiku-4-20250514",
            messages=[{"role": "user", "content": "ping"}]
        )
        return {"status": "healthy", "tokens": count.input_tokens}
    except Exception as e:
        return {"status": "degraded", "error": str(e)}

Observability

• Request/response logging (redact content, keep metadata)
• Latency tracking (p50, p95, p99)
• Token usage tracking (input + output per request)
• Cost tracking per feature/customer
• Error rate alerting (429s, 5xx, timeouts)

import logging
import time

logger = logging.getLogger("anthropic")

def tracked_create(**kwargs):
    start = time.monotonic()
    try:
        response = client.messages.create(**kwargs)
        duration = time.monotonic() - start
        logger.info(
            "claude_request",
            extra={
                "request_id": response._request_id,
                "model": response.model,
                "input_tokens": response.usage.input_tokens,
                "output_tokens": response.usage.output_tokens,
                "duration_ms": int(duration * 1000),
                "stop_reason": response.stop_reason,
            }
        )
        return response
    except Exception as e:
        duration = time.monotonic() - start
        logger.error("claude_error", extra={"error": str(e), "duration_ms": int(duration * 1000)})
        raise

Content Safety

• System prompts reviewed for injection resistance
• User input validated and length-limited
• Output scanned for sensitive data leakage
• Content moderation for user-facing responses

Infrastructure

• Deployment uses canary/rolling strategy
• Rollback procedure documented and tested
• Runbook created (see anth-incident-runbook)
• On-call escalation path defined

Alerting Thresholds

Metric	Warning	Critical
Error rate (5xx)	> 1%	> 5%
p99 latency	> 10s	> 30s
429 rate	> 5/min	> 20/min
Daily cost	> 80% budget	> 100% budget
Auth failures (401/403)	> 0	> 0 (immediate)

Prerequisites

• Have an approved release/artifact digest, production workspace, secret-manager reference, owner/on-call, change record, canary plan, and tested rollback command.
• Define the model/version, data classification, allowed destinations, retention, budget, rate-limit, latency, error, and content-safety thresholds for this release.
• Prepare synthetic fixtures and a staging environment that matches production policy; never validate readiness with live customer content or by printing credentials.

Instructions

1. Confirm every checklist item with an evidence link or redacted receipt: authentication, workspace isolation, model/version, error handling, limits, cost, observability, content safety, and rollback.
2. Run staging contract, health, synthetic redaction, timeout, rate-limit, permission, and output-safety tests. Verify logs/metrics contain metadata only and that deletion/retention behavior is proven.
3. Deploy the approved artifact to a small internal canary. Monitor p95/p99 latency, 4xx/5xx/429, token/cost aggregates, rate-limit headroom, and policy probes; halt on any critical threshold.
4. Require owner and on-call approval before staged production promotion. Preserve the prior revision and ensure the rollback path is executable without exposing secrets or content.
5. After rollout, issue a redacted receipt, revoke temporary test access, and retain only the evidence required by the documented policy.

Output

Produce a go-live receipt containing artifact/config digests, workspace/model classes, checklist evidence, synthetic test results, canary and threshold outcomes, approvals, rollout state, retention cleanup, and rollback reference. Exclude API keys, prompts, responses, customer identifiers, and raw exception text.

Error Handling

Gate failure	Required response
Authentication, workspace, or permission check fails	Do not deploy; verify secret binding and scope, then rotate/revoke only through the approved process.
429/5xx, timeout, latency, or budget threshold fails	Halt promotion, apply bounded degradation/circuit breaking, and roll back to the prior revision.
Redaction, content-safety, or retention check fails	Stop traffic, quarantine affected artifacts, correct the boundary, and rerun staging evidence.
Missing approval or unverifiable evidence	Mark release not ready; do not bypass the gate.

Examples

For artifact=sha256:fixture in staging, run synthetic fixture-request-001, assert sensitive_content_logged=0; contacts_exported=0; rollback_test=pass, then canary 1% internal traffic. A failed 429 gate records go_live=halted; rollback=prior-revision and sends no further production traffic.

Resources

• API Status
• Pricing
• Rate Limits

Next Steps

For version upgrades, see anth-upgrade-migration.