Claude Code's Projects beta reviewed: Claude as coordinator, less visibility

The reviewer normally runs several Claude Code CLI sessions side by side on Ghostty via herdr, watching each agent's state in a list and deciding himself what to delegate and how to connect the results. Projects shifts exactly that job — the splitting of work and the pulling together of results — to Claude, which is why he describes the difference as being about who is in command.

A second change is where the work happens. Threads run in Anthropic's cloud, so work continues after the laptop is closed and can be checked from a phone, but almost nothing from the local setup comes along. Each thread gets a fresh cloud environment, so globally installed skills and plugins do not work; the reviewer ended up committing only what was needed to the repository, which he found tidied things up. His Codex-based external review flow did not carry over to cloud threads.

What he found least clear was how far the coordinator goes. When a thread was waiting for input, nothing happened, and permissions still had to be answered inside the thread. Threads work on separate branches so they do not collide, but nothing prevents them from touching the same files, and conflicts are resolved at the PR stage. He tried a style of development where design decisions are settled one at a time while verifying, and found a layer he could not see through uncomfortable for that; for work with fewer decisions and more volume, his impression might have been different.