How I Work with AI
I treat AI as a multiplier, not autopilot — using spec-driven development and agentic engineering under versioned rules, not vibe coding.
Narrative
I've moved from ad-hoc "vibe coding" to a governed, repeatable workflow built around spec-driven development and agentic engineering so AI output stays consistent, reviewable, and honest. Ad-hoc prompting still has its place for small, low-risk changes; anything substantial goes through the governed workflow below.
Practices
Governance files
Versioned rules (truthfulness, scope, conventions) that every AI agent follows.
Spec-driven development
Define the goal, constraints, and acceptance criteria before implementation so AI can pressure-test the plan.
Agentic engineering
Break work into bounded, reviewable tasks that AI agents can execute under governance.
Per-task scaffolding
Implementation plan, before/after, QA verification, changelog, and root-cause analysis for each change.
Session handoffs
A written snapshot so any session (or machine) can pick up without relying on the model's memory.
Reusable skills
Repeatable commands for common tasks (run locally, pre-deploy checks, release).
Proof
This site itself is built and maintained this way — its governance files and skills live in the repo.