Claude Code Foundations for Designers
Important concepts for designers new to Claude Code
If you’re here from my previous blog post, welcome back. You now know how to use Terminal or Finder to make folders and navigate between them. Let’s make use of that knowledge! In this post, I’m going to share some higher order concepts for dealing with Claude Code.
1 project = 1 folder
As you work on various projects, the humble folder (aka directory) acts as contextual blinders for Claude. During chat sessions, Claude only sees the contents of the folder it was started in, but not up & out (unless you ask it to). I recommend structuring work into folders. To draw an equivalence to Figma, 1 project = 1 Figma file.
My personal preference is to have:
the Home directory ~
↳ a projects folder
↳ a folder for each project
This isn’t new. It’s how many of us were organizing files on our computers even before the advent of coding agents.
You can do this in Finder, but here’s how to do it in Terminal and get Claude working in a new project folder:
- Create the folder structure for your project
- Navigate to your project
- Only then do you start Claude
Here’s how that would look in Terminal:
# 1. from your home directory, create the "projects" folder ⤵
alexis@CORPK4M7R92X5N ~ % mkdir projects
# 2. cd into the "projects" folder ⤵
alexis@CORPK4M7R92X5N ~ % cd projects
# 3. create the "my-website" folder ⤵
alexis@CORPK4M7R92X5N projects % mkdir my-website
# 4. cd into the "my-website" folder ⤵
alexis@CORPK4M7R92X5N projects % cd my-website
# 5. launch Claude Code ⤵
alexis@CORPK4M7R92X5N projects % claude
Now you’ve launched Claude. An orange robot greets you and everything you type shows up between two horizontal lines. Remember this context change.
▐▛███▜▌ Claude Code v2.1.269
▝▜█████▛▘ Opus 5. (1M context) with medium effort · Claude Max
▘▘ ▝▝ /Users/alexis/projects/my-website · /rc
-----------------
>
-----------------
Slash commands
Claude Code’s graphical user interface (GUI) is in fact a Terminal User Interface (TUI) and by that virtue most of its capabilities are mostly invisible and therefore can be very hard to discover & learn; doubly so if you’re used to using desktop apps with a mouse to click on icons, actions, and contextual menus.
Operations you may be used to performing via a button or a menu are now neatly tucked away behind a /.
Very similar to Terminal commands, “slash commands” only work in Claude Code.
No one tells you about them unless you’re a fan of reading documentation
or the output of the /help command.
I mention slash commands here since we’ll go over a few of them in ther post.
Slash commands are invoked by typing a / character, after which Claude will launch into an auto-complete mode where it enumerates
the available slash commands. You can navigate these with the ⬆ & ⬇ keyboard arrows and hit Tab or Enter
to accept the command. You can also just type the full name of the slash command and hit Enter.
Get to know the context window
You may have heard of the “context window” agents have. This is their short term memory. It’s the full amount of things agents can hold in memory at any given moment. Most Claude models have a context window of ~1 million tokens.
A token is a basic unit of text used by models as they work. They are more than a letter, but less than a word.
As a conversation with Claude progress, the context window fills up.
The context window isn’t visible by default. You can ask Claude to display how full it is by running the /context command.
Anecdotal evidence says that the quality of Claude’s output diminishes once context has reached ~50%.
When the context window is full, Claude will automatically “compact” the entirety of your conversation in an effort
to extend the context window and keep the conversation going.
Unfortunately this process is opaque and some important details may get lost in the compaction.
You can manually reset the context window with the /clear command. This is equivalent to exiting Claude and launching into a new conversation.
It can be useful if the quality of output has decreased, or you want to move on to a different project task.
Be warned! Claude has no long-term memory.1 Every time you launch Claude, the context window resets. Any previous context from the converastion you just had or instructions you gave Claude about how it should answer, or how it you like working will need to be repeated, e.g.:
use our company design systemrespond briefly,modifiy the design we just worked on,don't use em-dashes,simplify technical explanations
Don’t fret! Folks have gotten around this constraint in many different ways.
Everything is a file
Unix-based operating systems has a principle whereby Everything is a file.
While this concept goes deeper than we need for our purposes, it’s a useful mnemonic device when working with agents. This is because Claude’s only way to remember things is files. It makes up for the lack of memory with its ability to infer context from its environment: (you guessed it!) files & folders.
When Claude is launched in a folder and given an instruction, it starts by getting it bearings by looking at what files are available to it in the pwd.
It does this by executing commands inside that folder; commands similar to those we learned about in the previous article.
This is how it forms a picture of what’s going on in any given folder and “knows” (e.g.):
- images are stored in the
pictures/folder - design decisions are stored in the
decision-log.txtfile - website source code is stored in
src/and CSS insrc/styles/ - this project uses React and D3.js for data visualisation
Claude excels at running these commands and giving itself the context it needs, but this comes with a time & token cost.
To save time and tokens on Claude doing this every time, Anthropic recommend using a CLAUDE.md or AGENTS.md file to save any
important findings that need to persist across conversations.
CLAUDE.md matters most
Speaking of files, CLAUDE.md2 is a file you can write that gets read by Claude at every new session. It’s a good idea to
provide very high-level guidance to Claude in that file. Things like:
- what this project is (a design system, a blog, a mobile game)
- where open tasks are tracked (Jira, Github Issues, Linear)
- anything else about how you want to work or that Claude should contextualize
For a project that’s already in flight, the fastest way to get a half-decent CLAUDE.md is to run the /init command in claude:
▐▛███▜▌ Claude Code v2.1.269
▝▜█████▛▘ Opus 5. (1M context) with medium effort · Claude Max
▘▘ ▝▝ /Users/alexis/projects/my-website · /rc
-----------------
/init
-----------------
I recommend you open this file and read it, for accuracy, and to get familiar with what’s inside. If anything needs updating, change it. It’s just a text file!
AGENTS.md works too
As of September 18 2026, Claude added support AGENTS.md: a different filename than CLAUDE.md to do the same thing.
You may prefer AGENTS.md over CLAUDE.md to ensure other non-Claude coding agents are also able to ingest the
same context as you’re providing Claude.
What’s next
By now you have an understanding of where & how to operate Claude Code. In the next post, we’ll dive into Model choice, Model Context Protocol (MCP) servers and skills. These are quick ways to help Claude Code gain superpowers.
Footnotes
1. Auto memory and some plugins exist which allow Claude to save important details about your chats and carry them over to each session. Learn more about Claude’s long-term memory
2. md is the file extension for Markdown: a lightweight markup that makes
it easier to format plain text into a structured document using symbols (e.g. # for headings and * for bold)
Loading comments…