The first rung of Boris Cherny's 4-Step AI Maturity Framework, above baseline chat-only use. One engineer works with one agent, mostly supervised: a single session at a time, with almost every change reviewed before it merges — having an assistant actually do a task for you, rather than just answering questions. The bottleneck this step removes is copy-paste friction and the low trust that forces manual review of everything.
Self-diagnostic: you're not genuinely at this step if you still copy-paste Claude's output elsewhere, or copy-paste outside data into Claude — a "yes" to either signals unresolved copy-paste friction.
Practical setup: open Claude Desktop → Code → select or create a project folder → run /init to optimize the project. Pair with a file browser (e.g. Obsidian) to inspect what Claude writes to disk without needing to code, and connect external systems (e.g. Notion) via the '+' icon so the agent reads/writes data directly instead of the user relaying it by hand.