Claude Code patterns: spec-first prompts, git worktrees, and e2e harness

A community write-up shares practical ways to make Claude Code more reliable: drive features from an "expectations" spec into requirements/design/tasks, isolate A/B implementations with git worktrees, and keep context lean by pruning skills and using /context. Users also report short autonomous sessions can implement most of a feature, and Playwright user-journey tests work well as a regression harness.