The agent-skills workflow I actually ship with

AI coding is only as good as the process around it. Left unstructured, it produces code that is confident, plausible, and occasionally wrong. I lean on a small set of agent-skills that turn that into a repeatable loop.

Spec, then plan, then build

I start with a spec, break it into ordered tasks with acceptance criteria, then build one slice at a time. The point is to never hand the agent a vague “build the thing.” A task with a clear done condition is a task the agent can finish and I can verify. This site was built exactly that way: a plan first, then small incremental slices.

Tests first, especially with an agent

Test-driven development matters more with an agent, not less. A failing test written before the code is a specification the agent cannot fake passing. Red, then green, then move on. It is the cheapest way to keep a fast-moving agent honest.

Review before it merges

Before anything lands I run a review pass across the usual axes: correctness, readability, architecture, security, and performance. Independent eyes catch what the author, human or agent, is too close to see. A simplify pass on top trims the cleverness that tends to creep into generated code.

A performance and ship pass

Before it goes live there is a performance audit against Core Web Vitals and a short pre-launch checklist. This portfolio got exactly that: an independent performance review flagged a few real improvements, and they were applied before it shipped.

The skills are just discipline, made default

None of these skills are magic. They encode the discipline I would want on any serious project: write it down, prove it works, get a second look, check it before launch. The win is that the discipline becomes the default path instead of the thing I skip when I am in a hurry. With an agent moving quickly, those checkpoints are what keep speed from turning into debt.