CI you'll actually use

Most teams build elaborate CI pipelines on day one. Within weeks, half the steps get bypassed because they're slow or flaky. The minimum that survives is shorter than people expect.

What stays

Type check. Lint. Build. Three commands, one job, a few minutes total. Run on every push.

These three catch most accidental regressions and run fast enough that nobody bypasses them.

What goes flaky

End-to-end tests. They start green and decay. Within months, half are skipped or marked "intermittent." The signal-to-noise ratio collapses. We've watched this happen in every project that started with E2E tests as a CI gate.

The compromise: run E2E tests on a schedule, not on every PR. They catch regressions without blocking sensible PRs.

What we add later

Lighthouse CI on the main branch. Catches performance regressions that don't show up in code review. Reports a delta from last run.

Bundle-size tracking. Flags when an unintentional dependency adds 100kb to the JS bundle.

What we don't add at all

Custom shell scripts that paper over framework deficiencies. If the framework needs a custom CI step to be safe to ship, the framework is the problem.