Claude Code for Noobs
What I've learned using it every day as a designer
I've been using Claude Code every day since it launched and some have asked me how... I wrote about my whole terminal journey last week. But here's a more concise, to the point version... just the stuff that matters when u're starting out.
Install
Open Terminal on Mac:
npm install -g @anthropic-ai/claude-code
No npm? Get Node.js first. Then type claude. There's also a desktop app if terminals aren't your thing. But if you're an adventurous noob like me, I recommend using Ghostty.
Plan First, Build Second
When starting, hit shift+tab twice for plan mode. Every time. Go back and forth until the plan makes sense, then let Claude build it. A good plan means Claude gets it right on the first try instead of going in circles.
When something breaks mid-build, don't keep pushing. Go back to plan mode and re-plan... I have learned this the hard (and expensive) way.
Set Up Your CLAUDE.md
This is the file that matters most. CLAUDE.md sits in your project and tells Claude your rules, preferences, and patterns. Claude reads it every session.
The move: every time Claude does something wrong, end with "Update your CLAUDE.md so you don't make that mistake again." Claude writes good rules for itself. I do this constantly... every time it adds a custom Tailwind class that deviates from the system, or goes and recreates a component that could've reused, I make it add a rule. Over time, the mistakes stop.
If this is the first time Claude sees a project, then run /init, this will inspect the file and create a CLAUDE.md for u. And, again... you don't have to write this, let Claude write/update its own file.
Be Specific
Claude does better with specific prompts. "Full-height hero, center text vertically, 32px between cards, Inter from Google Fonts" beats "make it look better" every time.
Designers already know how to give feedback. Use that. Describe what u want the way u'd describe it to another designer bluntly.
Also: hit fn twice on Mac for voice dictation. This has been super helpful for me, as a new dad, I often have to be carrying baby in my arms while I tell Claude what to do. And, supposedly, we speak faster than we type, and we naturally give more context when talking out loud... but, I'll be honest, speaking sometimes gives me a mini mind shortcut lol... it's weirdly a different mental model.
Make Claude Review Itself
Don't just accept what Claude gives you. Push back.
What I do is that I open another tab with another session with Claude, and I ask it to review the work in the specific worktree the other agent just finished some work. It would give you insights on what could be improved, mostly in order of priority. Then I pass those back to the original session, and tell it someone reviewed their code lol. So, yeah, I have per PR or project two tabs, one building and one reviewing.
Pro tip: Use browser agent from Vercel... this gives Claude the ability to open the browser, test flows, test different aspect ratios for responsive design, take screenshots and review, click buttons, EVERYTHING... making Claude test its own implementation.
(disclaimer, I work at Vercel... but I really recommend this since it's just so damn useful... although it can take a ton of memory from your computer)
Use Subagents
Add "use subagents" to your prompt when u want Claude to work harder. It spins up smaller agents that handle pieces in parallel. Keeps your main conversation focused instead of one long messy thread... it just does a better job.
Run Multiple Sessions
Run 3 to 5 Claude sessions at once using git worktrees. Each session gets its own copy of the code so they don't step on each other. One builds, another tests, a third researches. You can ask Claude to do this for you.
Once u try parallel sessions, one at a time feels slow. But, be careful... idk why, but Claude does get tripped out and suddenly goes back to main, or gets trapped in another worktree... I honestly don't know why and I wish Anthropic did something about this that were more deterministic.
Skills
If u do something more than once, turn it into a skill. Skills are custom commands u define in .claude/commands/ that Claude can run. I have skills for optimizing images, generating social cards, compressing videos, helping me with images for my blog posts using Sharp...
At v0, Claude kept inventing new colors or styles that... look just right that you don't notice that it actually deviated from the system... so, made a skill instructing it on our tokens, text styles, components... and now it rarely makes those mistakes anymore.
U can also grab skills others have shared at skills.sh.
Connect Your Tools
Through MCP, you can connect Claude to your other tools. For designers, the big one is Figma MCP... Claude can pull your designs directly instead of u having to describe everything. There's also MCP for Slack, Notion, Google Drive, even Blender.
Less describing, more building.
Just Say "Fix"
When something breaks, paste the error and say "fix." That's it. Don't overthink it... that's your PROMPT ENGINEERING:
fix
For design bugs: screenshot what looks wrong, make notes if you use a tool like CleanShot, paste it in, say "this doesn't look right." Claude can see screenshots.
Oh, to paste images on the terminal, it's not cmd+v, but ctrl+v.
Start Fresh When It Gets Weird
If Claude is going in circles or the conversation gets long and messy, don't keep adding to it. Type /compact to summarize and free up context, or just quit and start a new session. A fresh Claude with a good CLAUDE.md picks up where the old one left off.
Don't waste tokens on a confused agent.
Deploy
vercel.com, sign up with GitHub, import repo, click Deploy. Every push to git auto-deploys after that.
Or install the Vercel plugin and say "deploy this" inside Claude Code.
Some tips here came from Boris Cherny's thread on how the Claude Code team uses their own tool. The Claude Code docs are worth reading too.
Tips I missed? Send me a dm!
oh, and try two things I'm involved on: v0 (day job) Efecto (side project)
and one last thing... if u know what u want, and you're sure Claude can handle it without hand holding, then before you boot Claude, run: claude --dangerously-skip-permissions ... like it says, it will stop dangerously asking for permissions.