how i actually ship with AI
not a prompting guide. a loop, a few rules, and what i refuse to delegate.
I don't have a clever prompt library. I have a loop.
the loop
- State the change in one sentence. Out loud, or in the prompt. If I can't, I'm not ready to ship — I'm exploring. Different mode.
- Hand the boilerplate over. Scaffolding, fixture data, type-shuffling, codemods, plumbing between two known shapes. None of this is where I add value.
- Read the diff like a reviewer who's seen the codebase. Because I have. Most of the bugs I catch from agents are the ones that look right.
- Run the real thing. Real database. Real load. Real auth. Mocked tests caught nothing the last three times I trusted them in production.
the pipeline
pencil → backend → frontend → tests → deploy. one machine. me.
I design the screens in pencil. write the go and python services behind protobuf. build the web app and admin in next.js. ship the mobile in react native. write the e2e in playwright. push to k8s through a cloudflared tunnel.
ten years ago this was a designer, two backend engineers, a frontend engineer, a mobile engineer, a QA, and a release ticket.
the surface area didn't shrink. the team did.
what i don't delegate
- The first decision on a new system. Schema, contracts, failure modes — I think these through myself, on paper. The agent helps me write them down, not figure them out.
- Anything that's load-bearing in an incident. If it pages someone at 2am, a human wrote it.
- Code review on someone else's PR. The review is the point.
what changed
Features that used to take a week land in a day. The cadence is different, the discipline isn't. I'm still the one who knows what "good enough to ship" means for this team, this quarter, this customer.
The typing got cheaper. The judgment didn't.