Get Claude Code running once, and the hard part is behind you.
Now we turn it into your AI — an AI that knows your voice, your work, and the people you work with, and builds you tools on demand.
Let's make sure everything is in place, then get to the fun part.
This guide assumes Claude Code is already on your Mac and signed in. If it's already set up, you're good; skip to the check below. If you opened this cold and typing claude does nothing, two things need to exist first: the app itself, and a one-time sign-in (it runs on a paid Claude plan like Pro or Max). The official Mac steps are at code.claude.com/docs/en/setup: install it, run claude, then follow the sign-in prompt once.
Some of this will feel strange for a day or two. You'll type something and it won't do what you expected. That's normal, and it isn't you. Everyone hits it. Give it a few real tries on work that actually matters, and the strangeness quietly turns into "oh, this is just how I work now." Stuck for more than a few minutes? Back up and ask Claude itself what went wrong. It's a conversation, not a wall.
It's your terminal — a single window where you talk to your AI. Open it like any app.
claude
You should see Claude Code start up, ready for a message. That's it. You're in.
Hi! In one sentence, what can you help me do?
It answers. You're talking to the most capable writing tool you've ever used.
If claude says "command not found," quit your terminal and reopen it once. If it asks you to sign in, follow the prompt — that's normal the first time. If you're truly stuck, ask Claude itself what went wrong; it's good at explaining.
Install the starter kit that came in this folder — double-click install.command (the README shows you how). It drops the six commands and two skills below onto your computer so everything in this guide actually works when you type it.
Every brief used to mean: a browser tab to research, a doc to draft in, email to find the thread, a PDF of the last one, paste into a chatbot, back to the doc. Six apps. Energy burned on logistics instead of the work.
Now it's one place. You describe what you need in plain English, and your AI reads your files, drafts the document, researches whoever it's for, sharpens the language, and saves it — without you leaving the window. The setup was an afternoon. The payoff is every workday after.
You never need to know how. You describe the outcome you want, in your own words. Your AI figures out the steps. That's the whole game.
Open your terminal, type claude, then paste any of these. Pick the tab that fits your day. Hit Copy, paste, press Return.
The installer put brief-sample.txt and draft-sample.txt in your Documents/ResearchKit/samples folder — both made up — so the very first prompt below works right now, no setup. Use your own files anytime.
Read the two files in my Documents/ResearchKit/samples folder — brief-sample.txt and draft-sample.txt. The brief is the situation; the sample shows the flat tone to avoid. Draft a clean first version for the brief — clear and warm, a strong opening, no filler. Save it as draft.txt in my Documents/ResearchKit folder.
Read draft.txt. The opening feels flat. Rewrite it: lead with the reader's problem, show I understand their world, then say what I bring. Under 150 words.
I have a call with Acme Co in an hour. Search the web and give me a one-page brief: what they do, what's putting them under pressure right now, the smartest thing I could lead with, and three good questions to ask.
Write a firm-but-polite email disputing a medical bill. They charged me for a visit my insurance should have covered. I want it to sound reasonable but make clear I expect this fixed.
Help me think through whether to switch our family's phone plan. Ask me what I'm paying now and what we actually use, then lay out the trade-offs simply and give me a recommendation.
I'm going to paste a long contract. Read it and tell me in plain English: what I'm agreeing to, anything that could bite me, and what I should ask about before signing.
Plan a week of simple weeknight dinners for a family with young kids — nothing fancy, real ingredients, one sheet-pan night. Then make me a grocery list grouped by aisle.
My 6-year-old asked me a hard question about something scary in the news. Help me answer it honestly but gently, in words a 6-year-old understands.
Write a warm, short thank-you note to my kid's teacher for going out of her way this term. Specific and genuine, not gushy.
End any request with "save it as [name].txt in my Documents folder" and your AI writes the file for you. No copy-paste, no losing the draft. Say "now make that shorter" or "give me three subject lines" and it keeps going from where you left off.
A slash command is a shortcut you trigger by typing / and a name. Claude Code has built-in ones, and the starter kit gives you six of your own, tuned for real work.
| Type this | What it does |
|---|---|
/help | Lists everything you can do |
/clear | Starts a fresh conversation (forget the last thing) |
/init | Sets up an AI memory for the project you're in |
/memory | Opens your AI's memory to view or edit |
/agents | Create and manage your AI's specialist helpers |
/mcp | Connect new tools (web, files, and more) |
These came in the kit. Type the command, then a few words, and press Return.
| Command | What it does |
|---|---|
/draft-kickoff | Asks a few quick questions, then builds a first-draft outline in your voice |
/sharpen | Diagnoses why a draft is flat, then rewrites it tighter |
/research | A one-page brief on any person, company, or topic before a call |
/pressure-test | Pressure-tests a draft — where it loses, before anyone else sees it |
/council | A small board of experts debates a hard call and gives you a recommendation |
/coach | A board of ten legendary coaches debates a hard people-or-team call, then gives you one recommendation |
/draft-kickoff the Q3 launch brief
It takes over from there. The words after the command are yours to change.
If you just ran the installer and the commands don't show up, quit your terminal and reopen it once (or type /reload-skills). Claude Code only needs to notice the new files once.
Out of the box, your AI is a brilliant stranger. It becomes that same AI after it's read everything about how you work — your voice, your work, the people you deal with, your recurring tasks. It never needs to be briefed twice.
That's all a "personal AI" really is: an AI that remembers you. Yours already has a starter memory from the kit. Here's how to make it truly yours.
Open your terminal, type claude, and paste this. Answer honestly — there are no wrong answers, and this is what makes it yours.
Interview me so you can remember who I am and how I work, then save it to your memory. Ask me one question at a time about: what I do, the kinds of writing I do most, how I want my work to sound, who I make things for, what tasks I repeat, and the parts of my week that frustrate me. When we're done, update your memory file so you remember all of this next time.
Anytime, just ask your AI in plain English:
What do you remember about me, and where is your memory file saved?
That file is your AI's brain (usually ~/.claude/CLAUDE.md — the installer told you where it landed). You never edit it by hand unless you want to — just talk to it.
Whenever something's worth remembering, just say so:
Update your memory: my biggest project right now is the website relaunch, and I like drafts under two pages.
It edits its own memory. Tomorrow, it already knows.
Every time you correct it — "warmer," "shorter," "we don't say that" — it can remember the lesson. A stranger you re-explain things to becomes a colleague who just gets it. That compounding is the whole point.
A skill is a playbook you teach your AI once, so it follows it forever. Write down how you do a recurring task, and from then on your AI does it your way — every time, without re-explaining.
Two came in your kit and are already working:
| Skill | What it does |
|---|---|
voice | Your AI writes anything audience-facing in your voice automatically |
review | A structured checklist that scores a draft before it goes out |
Try the review skill on anything you've written:
/review draft.txt
It's literally a plain text file in a folder. To install one someone sends you: drop its folder into ~/.claude/skills/ and restart Claude Code — that's the whole thing (your kit's installer already did exactly that for these two). To make your own, you don't even touch a folder — use the builder right below.
Think of a task you do over and over. Describe the pain below and get a ready-made skill you can paste straight in. You never write the file from scratch.
Every skill you build is institutional knowledge that never walks out the door. A new hire would take months to learn your voice. Your AI learns it once and applies it on the thousandth draft exactly as well as the first.
An agent is a specialist helper your AI sends off to do a job on its own, then report back. You stay in the conversation; it goes and works in the background.
The simplest version needs no setup at all — just ask for the legwork:
Research Acme Co across several sources — their website, recent news, their leadership, and what their competitors are doing. Pull it all together into one brief and flag anything that changes how I'd pitch them.
Behind the scenes, your AI can split that into separate searches running at once, then hand you one clean answer. You asked once; it did five things.
Want a dedicated helper — say, a "prospect researcher" that always works the same way? Type:
/agents
That opens a simple menu to create one: give it a name, describe its job, and your AI will hand the right tasks to it from then on.
You're the one with the goal. Your AI is the chief of staff. Agents are the team it delegates to. You never manage the team. You just say what you need.
Some calls deserve more than one opinion. The /council command assembles a small board of expert voices, has them debate from different angles, surfaces where they disagree, then gives you a clear recommendation.
/council Should we take on the new project, or pass and protect the team's time?
You'll get four distinct points of view — the numbers person, the delivery realist, the competitive strategist, the other side's own view — the real fault line between them, and a single honest call with a first step.
| Use it for | What you get |
|---|---|
| Pursue or pass on a pursuit | An honest go/no-go instead of wishful thinking |
| Reviewing a finished draft | Four critics finding what one read would miss |
| How to position against a rival | The angle that actually separates you |
| A hard personal decision | Same method, different experts — works for anything |
A council to decide whether to take it on. A skill to write it in your voice. A pressure-test pass before it ships. That sequence is the unfair advantage — and it all lives in one window.
This is the part that makes people sit up. Claude can already use tools on its own — read your files, search the web, look at a photo, write a little code, make a chart, save a document. You never say how. You say the outcome; it picks the tool. Try one. Pick your lane.
I have last-quarter-pipeline.csv in my Documents folder. Read it and tell me which deals are stuck, what the stuck ones have in common, and the single thing I should do this week to unstick the most revenue.
Search the web for Acme Co's three biggest competitors and what each one announced in the last six months. Give me one page, with sources.
Read the HOA bylaws PDF in my Downloads folder and tell me, in plain English, whether I'm allowed to put a shed in my backyard — and quote the exact rule that says so.
Here's a photo of my kid's soccer schedule (soccer.jpg in my Downloads folder). Read it and turn every game into a list I can paste into my calendar — date, time, and location for each.
Here are our new sign-ups for the last six months: 3, 5, 4, 8, 11, 14. Make a simple bar chart and save it somewhere I can open it — an image, or a little web page, whichever is easier. Then tell me the trend in one sentence.
Turn these rough notes into a clean, formatted one-page summary and save it as a document I can send. Here are the notes: [paste your bullet points]
You never told Claude how. It chose the right tool on its own — reading a file, searching the web, looking at a photo, writing a little code to draw a chart. That's the whole magic: you name the outcome, it picks the tools. The first time it wants to run something on your computer it'll ask permission — just say yes.
Out of the box your AI can already read your files, write documents, and search the web — so the research prompts above just work, no setup. Tools (the technical name is MCP) are for later: they connect Claude to other systems — a calendar, a database, a design file — when you have a specific need.
To see what's connected (nothing, at first) and manage it, type:
/mcp
When you hit an "I wish it could also reach into ___" moment, the easiest path is to let your AI set it up for you:
I keep my notes in Obsidian (it's free and stores everything as plain files on my Mac). Help me connect it so Claude can read them, one step at a time.
This is the one "power user" corner, and you do not need any of it to get enormous value. The commands, skills, and memory above are 95% of the magic. Only connect a tool when there's a real system you want Claude to reach — and let your AI walk you through it.
A list to work through at your own pace. Don't go in order — pick whatever's annoying you right now. Check it off when you've done it; your progress saves automatically. The ★ starred ones are the highest-leverage places to start.
Before you push through something tedious alone, ask one question: "Could my AI help with this?" The answer is yes far more often than you'd think. That habit — not any single command — is what changes your week.
At first, Claude asks before it does anything. That's good, it builds trust. Once you're comfortable, a few moves let you stop clicking "yes" and just fly. You decide how much it checks with you.
Shift + Tab to switchPress Shift+Tab inside Claude Code to cycle through these. The current gear shows at the bottom of the window.
| Gear | What it does | When to use it |
|---|---|---|
| Normal | Asks before each action | Start here, while you're learning. |
| Accept edits | Stops asking before each file change | When it's drafting/editing and you trust it. |
| Plan | Reads and thinks, changes nothing until you approve a plan | Big or important tasks — see the whole plan first. |
| Auto | Approves the safe stuff itself, still checks on risky things | When you want speed with a safety net. |
For anything bigger than a quick ask, press Shift+Tab to Plan and say what you want. Claude lays out exactly what it'll do before touching anything — you read it, then approve. It's "show me your plan first" instead of "just go." You can also start right in it: claude --permission-mode plan
When you're doing low-stakes work and the steady "can I?" prompts slow you down, you can turn them off for that session by starting Claude this way:
claude --dangerously-skip-permissions
This lets Claude act without asking — including running commands and changing files. Use it only when (a) you trust what you're asking and (b) you're in your own documents, not somewhere a slip would hurt. It's like switching off "are you sure?" — great for everyday writing, not for anything you couldn't undo. When in doubt, stay in Normal or Plan.
Typing the long command gets old. An alias is a short word you invent that stands for a longer command. Easiest way — let your AI set it up for you:
Set up a shortcut so that when I type "fly" in my terminal, it starts Claude with permissions skipped. Make the change for me and tell me how to use it.
After that, you just type your word and you're flying:
fly
Other handy ones to ask your AI for: kit (jump to your kit folder and start), plan (start straight in Plan mode).
None of this is required. But when you want more, these three are free and open — nothing you have to pay to use, and nothing that locks your work inside a service you'd have to pay to leave. Claude teaches you each one, one step at a time. You never learn it alone.
Obsidian is a free notes app that keeps everything as plain text files on your own computer, not in someone else's cloud. Your AI's memory and anything Claude saves for you are already plain text, so Obsidian just gives you a friendly window into them. Ask your AI to set it up:
I'd like to start using Obsidian for my notes. It's free and keeps everything as plain files. Walk me through installing it and pointing it at my Documents/ResearchKit folder, one step at a time, in plain English.
You don't need to become a programmer. But a few lines of Python can do in seconds what takes an hour by hand: rename two hundred files, pull the totals out of a stack of PDFs, tidy a messy spreadsheet. Claude writes it and explains every line as it goes.
Teach me a tiny bit of Python by solving a real problem of mine. I have a folder of PDFs and I want one number from each in a single spreadsheet. Write the code, explain each part in plain English, and walk me through running it.
Playwright is a free, open-source tool that lets Claude control a real web browser for you: fill the same form every week, log in and pull a report, grab figures off a page. You describe the task; it does the clicking.
I do the same thing on a website every week. Help me use Playwright so Claude can do it for me. Ask what the site is and what steps I take, then set it up one step at a time.
None of these lock your work inside something you have to pay to leave. You learn them the same way you learn everything here: describe what you want, and Claude teaches as it builds.
Pick one "First Win" prompt and run it on something actually on your plate.
Do the interview. Let it learn your voice and your work.
Use the generator above on a task you repeat. Install it. Use it twice.
"Could my AI help?" before anything tedious. Watch the list fill in.
The terminal isn't for coders. It's for people who'd rather stop clicking through a dozen apps and start talking directly to the most capable writing tool ever built. That's you now.