3. Setup¶
← Working with context · Contents · Next: Project layout →
Install¶
Follow the official installation guide
for your operating system, then open a terminal in your project and run claude.
The native installer is the recommended route. If you already use npm, the guide also
covers that installation method.
First run walks you through login. Run claude from inside a project directory — it takes
the current directory as its working root and reads any CLAUDE.md it finds there.
Plans and seats¶
Ask Micah before buying a subscription: the lab may already have seats. Check the scientists programme for current eligibility and terms. Prices and usage limits change, so they are not reproduced here.
Python¶
Lab members use macOS, Windows and Linux, including the Cybertron server. There is no single lab-wide Python path. Use the environment documented by the project you are working on; for a new practice project, the setup below works across these platforms.
Use uv to manage the project's Python environment: a separate set of packages for
this analysis. Its lockfile records the dependency versions. If setup is unfamiliar,
ask the agent:
Help me set up a new practice project. Check my operating system and shell,
then check whether Git, Python and uv are installed. Explain anything
missing and help me install what I need using instructions for this system.
Keep the setup local to this project where possible.
Use pandas and matplotlib for a small analysis, and pytest for tests.
Explain the commands and save the setup in a Git commit.
With uv installed, these commands initialise a new practice directory:
uv init
uv add pandas matplotlib
uv add --dev pytest
uv run python --version
uv run pytest --version
These commands check the environment; they do not run an analysis. See the uv project guide for details.
For an existing project, follow its environment instructions. Some neuroimaging tools
need conda, containers or separately installed software. Record the chosen setup in
CLAUDE.md and avoid adding a second package-management workflow without a reason.
On a shared server, follow its existing environment and installation arrangements. Cybertron and GenomeDK are different systems; the SLURM instructions later in this book are specifically for GenomeDK.
Terminal basics you will actually use¶
| Key | Does |
|---|---|
Shift+Tab |
Cycle modes — get to Plan Mode before any nontrivial task |
Esc |
Interrupt the current response or action |
Esc Esc |
Rewind to an earlier point in the conversation |
/clear |
Wipe context and start fresh. Do this between unrelated tasks |
/model |
Switch model |
! prefix |
Run a shell command yourself, output goes into the conversation |
@ prefix |
Reference a file so it gets read into context |
Sanity check before you start real work¶
- Save the starting state in Git and use a branch for substantial changes. The agent can help initialise a practice repository; see page 4.
- You are using fictional practice data. Read page 13 before granting access to research files.
- You have a short
CLAUDE.mdwith the environment, scope and data-access rules.
Continue with the internal lab tutorial or the fictional warm-up. You can add more project structure later.