production-readiness

---
name: production-readiness
description: Use as the final pre-launch gate — verify reliability, observability, scaling, and disaster recovery are real, after security/QA/performance/accessibility audits pass. Blocks launch on critical gaps.
---

# Production Readiness Review

## When to use
- A release candidate has passed the security, QA, performance, and accessibility audits and is being considered for launch.
- Promoting a service from staging to production for the first time, or after a major architectural change.
- Periodically re-certifying a live system after significant growth or incidents.

Applies to any system that runs after release: hosted apps/APIs, background services, scheduled jobs, agents, mobile/desktop apps with backends, and embedded fleets. The four pillars below are universal even when the mechanisms differ.

## Workflow
1. **Confirm the audits are green.** Production readiness comes *after* security (`.claude/skills/security/SKILL.md`), QA (`.claude/skills/testing/SKILL.md`), performance (`.claude/skills/performance/SKILL.md`), and accessibility. If any are open, stop and route back. This skill verifies operability, not feature correctness.
2. **Reliability.** Verify graceful degradation: timeouts, retries with backoff, circuit breakers, idempotency on writes, and sane handling of dependency failure. Confirm health/readiness checks exist and are wired to the platform.
3. **Observability.** Confirm structured logs (no secrets/PII), metrics for the golden signals (latency, traffic, errors, saturation), distributed tracing on critical paths, and actionable alerts that page on user-impacting conditions — not noise. Verify dashboards exist for the top flows.
4. **Scaling & limits.** Confirm the system meets performance budgets at expected peak load (verified, not assumed), has autoscaling or a capacity plan, enforces rate limits/quotas, and applies back-pressure instead of collapsing. Identify the next bottleneck.
5. **Disaster recovery.** Verify backups exist, are automated, and — critically — have been test-restored. Define and check RPO/RTO. Confirm rollback works, runbooks exist for top failure modes, and there is an on-call/escalation path.
6. **Data & compliance.** Confirm data retention/deletion, encryption at rest/in transit, and any regulatory requirements (privacy, residency, audit logs) are satisfied for the project's domain.
7. **Operational dry run.** Walk a deploy + rollback in staging. Trigger a synthetic failure and confirm alerts fire and runbooks work. Confirm secrets rotation and on-call are real, not theoretical.
8. **Go/no-go.** Score the readiness checklist. Any unmet *critical* criterion is a no-go. Document accepted risks with owners and a remediation date; get explicit human sign-off before launch.

## Standards
- **Do** verify backups by actually restoring one; an untested backup is not a backup.
- **Do** require real, test-fired alerts on user-impacting conditions with an owner and a runbook for each.
- **Do** prove the system meets performance budgets at peak load before launch, with measurements.
- **Do** make every write idempotent or safely retryable, and every external call timeout-bounded.
- **Do** keep a tested rollback and a documented incident/escalation path.
- **Do** strip secrets and PII from logs and traces; confirm during the review.
- **Do** define RPO/RTO and confirm the recovery mechanism meets them.
- **Do-not** launch on assumptions ("it should scale", "backups probably run") — verify each.
- **Do-not** ship noisy or non-actionable alerts; alert fatigue means real pages get ignored.
- **Do-not** override a failed *critical* readiness criterion without explicit, recorded human approval.
- **Do-not** perform destructive recovery drills (restore-over-prod, failover) without approval and a safety net.

## Common mistakes to avoid
- "It works in staging" with no load test, then collapsing at real traffic.
- Backups that run but were never restored, discovered useless during an actual outage.
- Logging that captures secrets/PII, or alerts so noisy the team mutes them.
- No rollback rehearsal; the first real rollback happens during the first real incident.
- Missing idempotency, so a retried payment or webhook double-charges or duplicates data.
- No on-call/escalation defined; an alert fires at 3am to nobody.
- Treating readiness as paperwork instead of dry-running deploy, failure, alert, and recovery.

## Output format
A go/no-go report scored against the four pillars (reliability, observability, scaling, DR) plus data/compliance, each item pass/fail with evidence (load-test numbers, restore proof, fired-alert screenshot/log). Include an accepted-risk register with owners and dates, and a recorded human sign-off line. Use `.claude/checklists/production.md` as the source list and link the report to the release.

## Related checklists
- `.claude/checklists/production.md`
- `.claude/checklists/security.md`
- `.claude/checklists/performance.md`

## Related agents
- `.claude/agents/quality/production-readiness-auditor.md`
- `.claude/agents/engineering/devops-engineer.md`
- `.claude/agents/core/orchestrator.md`

More agent context in muxammadmamajonov/dot-claude

74 other files this repository gives its agents, the first 60 shown.