82 labs
← Field notes

Five levels of using Claude

How much of a team’s work Claude does at each level, and the setup each level needs.

We use five levels to describe how much of a team’s work Claude does. Level 1 is a chat window. Level 5 is work that runs on a schedule with nobody at the keyboard. This post lists the setup that each level needs.

  • Level 5 Systems that run on their own
  • Level 4 Claude Code
  • Level 3 Scheduled work
  • Level 2 Projects and files
  • Level 1 Chat
Each level builds on the one below it.

Level 1: Chat

A person asks a question and gets an answer: a draft email, a short script, a summary of a document. Most people start here. Results are better when Claude gets the material itself (the pasted document, the screenshot) and when follow-up requests stay in the same conversation, so it keeps the context.

Level 2: Projects and files

A project holds the documents and the instructions for one area of work, so the background is explained once. Memory carries what Claude learned about the work from one session to the next. Connectors let it read the tools a team works in, such as chat and shared files. At this level Claude also returns files: the spreadsheet, the slide deck, the Word document.

Level 3: Scheduled work

Cowork runs on the desktop and works in a person’s own folders. A task that is done every week becomes a saved workflow on a schedule, and it can be started from a phone. A weekly report is one example: Claude builds it on Monday morning and a person reads it before it goes out.

Level 4: Claude Code

This level is for technical leads and people who are comfortable in a terminal. A CLAUDE.md file holds the standing instructions for a project. Plan mode has Claude write out its plan before it changes anything. Subagents take separate parts of a job at the same time, and git worktrees give each one its own branch. The Model Context Protocol (MCP) connects Claude to outside tools, and custom commands store the prompts a team repeats. Two habits start here: Claude checks its own work, and someone watches token use so the cost stays predictable.

Level 5: Systems that run on their own

Routines run recurring work in the cloud with nobody at the keyboard. Hooks run a team’s own scripts at set points in Claude’s work, for example after every file edit. Headless mode and the Agent SDK make it possible to build software on top of Claude. People still set the limits, review the output, and correct the system when it gets something wrong.

Skipping levels

Teams that jump from level 1 to level 4 after a demo video tend to stall. Level 4 builds on the habits from levels 2 and 3: shared context in projects, and a workflow that someone ran by hand until it was reliable.

For a team at level 1, the next setup is a project for each area of work and the connectors. One scheduled task comes after that.