When creating a new project, I code from scratch to get the architecture/folder/file structure coherently setup.

However, when contributing to existing repos, I use Claude Code for: 1) understanding repo structure, flow of function calls, 2) making small fixes, 3) using plan mode to iterate on how to implement a larger feature.

On the topic of small fixes, I use git worktrees to use Claude Clode parallelly. Very useful to tackle 3 or 4 small fixes at once.

The diffs created by Claude Code can get a bit disorganized in my head, so I use GitHub PRs to manually check diffs.

Regarding your question, you are not necessarily wasting time doing the exercies. Being able to "code from scratch" accelerates your ability to use LLMs in a controlled, but powerful manner, without degenerating into vibe code slop.