Orienting before acting
When you start Claude Code in a project directory and give it a task, its first move — before writing anything — is usually to build a mental model of the codebase relevant to that task. It might list the directory structure, search for files matching a keyword, open a config file to see what package manager and test runner the project uses, or read a few files that seem central to the request. This isn't wasted motion; it's the equivalent of a new engineer spending the first ten minutes on a ticket reading the surrounding code before touching anything, except compressed into seconds and done every single session rather than only once.
This matters because the agent has no persistent memory of your codebase between separate sessions by default — each new session starts cold and re-derives context from what it reads. That's a deliberate tradeoff: it means the agent's understanding always reflects the current state of the code on disk rather than a stale mental snapshot, at the cost of spending some effort on orientation every time you start fresh.