Shipping and Launch

Ship with confidence: reversible, observable, incremental. Every launch has a rollback path and success metrics defined before traffic moves.

Hard Rules

• No launch without rollback — feature flag, previous artifact, or DB migration down path.
• Monitoring before traffic — alerts on errors, latency, and business KPIs.
• Staged rollout — internal → beta → % canary → full unless risk is trivial.
• Checklist is blocking — unchecked security/perf/a11y items are explicit deferrals with owner.
• Communicate — stakeholders know window, owner, and rollback trigger.

---

Workflow

Step 1 — Define success and risk

Document: what ships, who is affected, success metrics, rollback triggers (error rate, latency, support tickets).

Step 2 — Pre-launch checklist

Minimum gates:

• Tests green (unit, integration, e2e on critical paths)
• No secrets in repo; dependencies audited for critical CVEs
• Performance budgets met on changed surfaces
• Accessibility spot-check on changed UI
• Runbook: deploy steps, owner, comms channel

Step 3 — Staged rollout

1. Deploy to staging with prod-like data volume if possible
2. Internal dogfood or feature flag to staff
3. Canary (% traffic or invite-only beta)
4. Full rollout with monitoring watch window

Step 4 — Monitor and decide

Watch 24–72h (risk-dependent): error rate, p95 latency, Core Web Vitals, conversion/support signals.

Step 5 — Rollback or complete

If triggers hit → rollback first, postmortem second. If stable → remove flags, archive runbook learnings.

---

When NOT to use

• Pure local dev with no users
• Changes already live behind a flag with an existing runbook (update runbook only)

---

Gotchas

• "Green staging" with 10x less data than prod hides query issues.
• Launch without feature flags forces redeploy to rollback.
• Announcing before monitoring is wired means flying blind.
• Skipping comms strands on-call.

---

Common Rationalizations

Excuse	Reality
"It's a small change"	Small deploys still need rollback and error visibility.
"We'll add alerts after launch"	You discover outages from users instead of dashboards.
"Full send is faster"	One incident costs more than a canary day.
"Rollback is unlikely"	Plan for it anyway — Hyrum's Law applies in prod.
"Checklist is bureaucracy"	It's how you remember security and a11y under time pressure.

---

Output Format

## Launch plan — [feature]

Risk: [low/med/high] | Rollback: [how]
Stages: [list] | Metrics: [list]
Checklist: [pass/deferred items]
Owner / window: [who / when]

---

Examples

<examples> <example> <input>"Ship new billing flow Friday."</input> <output>Feature flag at 5% → 25% → 100%; dashboards on payment errors; rollback = disable flag; checklist signed.</output> </example> </examples>

---

Verification

• Rollback path documented and tested (or dry-run)
• Monitoring/alerts live before user traffic
• Staged rollout plan matches risk level
• Pre-launch checklist complete or deferrals explicit
• Success metrics and rollback triggers defined

---

Red Flags

• Launch without rollback path or previous artifact
• Traffic shifted before error and KPI monitors exist
• Full rollout for non-trivial change without canary
• Staging-only validation at fraction of prod data volume

Prune Log

Last pruned: 2026-07-04

• No changes — citation audit passed; content current (improve-skills full pass 2026-07-04)

Impact Report

Feature: [name] | Risk: [level]
Stages: N | Rollback: [ready/deferred]
Checklist: [N pass / M deferred]