You decide how it gets built — not the other way around

The pitch is always the same: describe an idea, let the tool handle the rest. The cost is visibility. You don't know what you'll get back until the AI has already produced it, and if it built the wrong thing — or the right thing the wrong way — you're stuck in a round of "no, like this," because you never got a say in how it was implemented. Guardrail tools help at the edges, but they still don't give you the control or visibility to steer the work while it's happening.

Software teams have always shipped the same way: incrementally. Break a large feature into user stories. Spec the hard ones. Implement in small, reviewable steps. That's the software development lifecycle, and it's incremental for a reason — smaller pieces are easier to diagnose, test, release, and control. nightshift enforces those exact principles when you work with AI agents, on open Claude Code primitives.

Each phase ends at a hard gate that hands control back to you: refine, spec, plan, implement, review. You approve each one before the next begins, so you catch a wrong turn before it's implemented instead of after. Catching it at the gate also means no tokens spent building code you'd have thrown away.

Software built in one shot gets some things right and some things wrong, and you can't tell which until you've read the whole thing. nightshift doesn't work that way. Work breaks down the way an agile team breaks it down — epic → stories → spec → implementation — so every unit is small enough to review on its own.

With the plugin installed, Claude Code follows these principles by default. Ask it to build something and it works the way you would if you were writing it yourself: refine, spec, implement in steps, review before merge. You stay the developer. It does the typing.