Shipping an AI feature in a single weekend
The constraints, scaffolding, and observability I lean on to take an idea from notebook to production by Monday morning.
Side projects only ship if the scaffolding is boring. I keep the stack opinionated: Next.js App Router, tRPC for the thin API, drizzle for migrations, and a single feature flag guarding the new path to prod.
The weekend checklist
- Friday: write a one-pager, choose latency/error budgets, decide what gets logged
- Saturday: build happy path, seed realistic fixtures, wire eval harness
- Sunday: remove sharp edges, add feature flag + rollout script, prepare rollback
Every prompt is treated like code: checked in, versioned, and logged as structured JSON so I can diff behavior across model upgrades. Shadow deployments let me compare responses in production without risking the primary path.
Instrumentation that catches surprises
- Trace every request with request id, prompt hash, and model version
- Emit structured errors for rate limits, content filters, and tool failures
- Keep a rolling sample of raw prompts/responses for fast root cause analysis
Practice
If you only have one weekend, bias toward traceability. You cannot tune what you cannot see.
The release ritual ends with a kill switch, a 15-minute runbook, and a follow-up task list for Monday. Speed is fine; reversible speed is better.
Includes prompts, runbooks, and rollout steps referenced here.