AI-DLC

Every build starts with the business outcome, in plain language — who it's for, what changes for them, and how we'll know it worked. Not a feature list. The problem, stated clearly enough to argue with.

Intent becomes a written spec — the contract. What it does, what it explicitly doesn't, and the acceptance criteria that define completion. The spec is reviewed before a line of code exists, and it stays the source of truth. When the AI drifts, the spec wins, not the code.

Data model, service boundaries, failure modes, cost shape, and security are decided before code — the calls AI has no opinion on because it sees the code, not the bill. The architecture lives as plain markdown (CLAUDE.md and topic docs) any tool or model can read, so every session starts with shared context.

AI writes against the spec — Claude, Google Antigravity, Codex, and Cursor, each used where it's strongest, not all at once on everything. Every PR is read by a senior engineer before it lands. The AI accelerates the output; the human keeps the bar. A small team ships at the velocity of a much larger one.

Observability, error handling, edge cases, security review, and cost review — before real users arrive. Structured logs and error tracking at every boundary, a runbook tied to specific alerts, least-privilege access enforced. This is the part vibe-coding skips and production punishes.

100% infrastructure as code on AWS. Fixed dev / QA / production environments, promotion gated on tests passing, rollback in one click. No console clicks, no hand-edited config in production. Every environment reproducible from a single command and a Git history.

Real usage becomes the next intent. Monitor what actually happens, learn where it strains, and feed that back to the top of the loop as the next problem worth solving. The lifecycle isn't a line from idea to launch — it's a loop that keeps turning.