I build it, and I answer the emails.

Not coverage for its own sake. I test the paths where failure is expensive: payment webhooks, sync reconciliation, auth boundaries, anything that touches a paid API. On Context Mesh that means a benchmark regression gate in CI, so a change that makes retrieval worse fails before it merges.

Partway through building Slote, my Flutter notes app, I replaced the WebView-based editor with a native one, a large mid-project rewrite. It went out behind a tagged rollback point. The decision was right, but I did not know that with certainty in advance, and building the escape hatch cost less than being wrong without one.

AtTalentFit went live on purpose before the build was finished, so every change after that landed on a site the company was already using. I plan for that from the start: changes small enough to check, data that moves forward with the product instead of being rebuilt, and alerts that reach me before a complaint does.

A hosting bill that creeps up every month is not something to discover after launch. One client's website had a target monthly cost before it had a homepage, and the choice of video service followed from that number rather than the other way round. The same goes for limits like function timeouts and API quotas: they shape the design in week one, not the rework in week three.

The Context Mesh benchmark ships with a disclaimer I wrote against my own results: it checks whether the expected answers turn up, not how well a real model then uses them or how much irrelevant material comes along, so the success rate flatters the tool. The cost figure is the one to trust, and the page says so.

I am usually the only developer and almost never the only person. On one client site the visual direction sits with the designer on the project; my job is to build to their calls and say early when a layout will fight the browser or cost a second of load time. And some of it is not code at all: that same site's domain took a run of emails with more than one hosting provider before it answered again.

Mainstream tools, documented decisions, and code in your repository. On client work every account transfers into your name at completion. The measure of a good handover is that a different developer can pick the project up without calling me, and that you are never paying me for access to your own software.

Slote has a maintained launch-readiness document tracking every open issue by severity. Resolved rows stay in the table for two releases, so there is a history of what went wrong and why. A clean list that has forgotten its own past is not the same as a healthy project.