curl -fsSL https://claude.ai/install.sh | bash into your terminal. On Windows PowerShell it is irm https://claude.ai/install.ps1 | iex. Homebrew users can run brew install --cask claude-code instead.claude --version. It should print a version number followed by "(Claude Code)". If the terminal is new territory for you, Anthropic's quickstart holds your hand through it.claude to start a session and follow the browser prompt to sign in with your Pro or Max account. Your credentials are stored, so this is a one-time thing.cd into it, then run claude. Claude Code's whole advantage is working on your actual files, so never start it in a scratch folder you do not care about.Intermediate · about 4 to 6 hours · free hosting. Not a template you filled in, a site you actually own, with real motion and a layout that does not look like everyone else's. The proof this is doable: Yanina Trekhleb, whose own bio describes her as a marketing automation and CRM specialist, built Flour Hour (a sourdough timing app with 22 recipes) entirely by describing what she wanted to Claude Code, in about three hours of conversation.
What it costs to put it online: nothing. GitHub Pages hosts a public repo for free, and Netlify has a free starter tier where you can drag a folder in. Claude Code walks you through whichever you pick and can push the code for you.
I want to build and deploy a portfolio site this week that looks genuinely expensive, not like a template. I'm not a developer, so explain any choice in plain language and do the work yourself.
The site is for: [describe it like you're briefing a designer, e.g. "my freelance brand strategy work: a hero, three case studies, a short about section, and a contact form"]. The feeling I want is [e.g. warm, editorial, lots of whitespace, one accent colour].
Before you start, ask me the 5 questions you most need answered. Then build it as plain HTML and CSS so I can edit the text myself, explaining each file in one line as you create it, and add one tasteful scroll animation rather than five.
When it works locally, deploy it free on GitHub Pages or Netlify, give me the live URL, and write me a short note at the bottom of the project explaining exactly which file and line to edit when I want to change the wording.Intermediate · about 6 to 8 hours · free API tiers. The difference between this and asking a chatbot for a trip plan is that this one is wired to live data, so it knows the actual travel time between two places and the actual forecast for the day you are going. That is what makes it a build rather than a prompt.
The free pieces you wire together:
.env file and add that file to .gitignore before you push anything anywhere. A key committed to a public repo is a key someone else is now using.Build me a trip itinerary planner that runs on real data, not guesses. I'm not a developer, so set up everything yourself and tell me the exact moment you need something from me.
How it should work: I enter a city and a date range, and it produces a day-by-day plan where each day has 3 to 5 stops, ordered to minimise backtracking, with the real travel time shown between consecutive stops and the forecast for that day beside the date.
Use Open-Meteo for weather (no API key needed, and add the CC BY 4.0 attribution it requires), Nominatim for looking up places (respect the OpenStreetMap usage policy: max 1 request per second, and set a User-Agent that identifies this app), and openrouteservice for travel times. Put every key in a .env file and add it to .gitignore before anything else.
Handle the failure cases properly: if a place can't be found, say so on screen instead of crashing, and if the forecast is more than 16 days out, show "forecast not available yet" rather than inventing one. Then run it once with [your city] and [your dates] so I can see it working.Intermediate · about 8 to 10 hours · free tiers throughout. This is the one with the strongest career payoff, because "I built the tool my team now uses" is the best line you can bring to an interview or a performance review. And it is not hypothetical inside AI companies either: Anthropic's own legal team, with no engineers, used Claude Code to prototype internal tools for routing questions to the right lawyer (how Anthropic teams use Claude Code).
Pick something your team currently tracks in a messy spreadsheet: content approvals, client onboarding steps, stock levels, contract renewal dates. The tool stores the records, shows their status, and emails somebody when one goes red.
The free pieces: SQLite for the database (it is a single file, already on your machine, zero setup, zero cost), or Supabase's free plan if the team needs shared access (500 MB database, and free projects pause after a week of inactivity). For the alerts, Resend's free plan sends 3,000 emails a month, capped at 100 a day, which is far more than an internal alert ever needs.
Build me an internal tracking tool my team would genuinely use. I'm not a developer, so make every technical decision yourself and explain it to me in one line as you go.
What it tracks: [describe the thing, e.g. "client onboarding, where each client has a name, owner, current stage, and a due date for the current stage"]. It needs a simple web page listing every record with a colour-coded status, a way to add and update a record without touching the database directly, and a filter by owner.
Use SQLite so there's nothing to set up, and structure the code so I could move it to a hosted database later without rewriting the app.
Then add the alerting: once a day, check for anything overdue or due within 48 hours and email a single digest to me using Resend, listing what's late and who owns it. Set it up so I paste in the API key once, and show me how to schedule that daily check on my machine. Send me a test digest with fake data before we go live.Intermediate · about 2 to 3 hours · free. A skill is a saved set of instructions Claude loads on demand, and a personal one is just a folder with a markdown file inside it at ~/.claude/skills/<name>/SKILL.md. The part nobody tells beginners: you do not write that file yourself. You describe the task, and Claude writes the skill. After that it is one command, forever.
Good candidates are the jobs you do the same way every week and still somehow do slightly differently each time: turning a messy client email into a structured brief, writing the weekly update from your notes, reformatting a report, QA-ing a piece of work against your own checklist. Full reference: use skills in Claude Code.
I want to turn a task I repeat constantly into a reusable Claude skill, and I want you to write the skill rather than me.
The task is: [describe it in real detail, e.g. "turning a messy client email into a structured brief with scope, deliverables, deadline, budget signals, and open questions I need to ask back"].
First, interview me about how I actually do it today, including the steps I always forget and the ways it goes wrong. Ask in one batch, no more than 8 questions.
Then write it as a personal skill at ~/.claude/skills//SKILL.md, with a description line that says both what it does and exactly when it should trigger, clear step-by-step instructions, and one worked example. Keep it under 500 lines.
Finally, run it on this real example so we can improve it together, and tell me what you'd change about the skill after seeing how it performed: [paste a real example here]. Intermediate · about 2 hours · free. The marketer payoff, and it is not a theory. Anthropic's own growth marketing team built an agentic workflow that reads CSV files of hundreds of ads, identifies the underperformers, and generates new variations inside strict character limits. Using two specialised sub-agents, it produces hundreds of new ads in minutes instead of hours (how Anthropic teams use Claude Code).
You need two things to start: a CSV export of your current ads (headline, description, and performance if you have it) and your brand guidelines in any format Claude can read. Sub-agents are what keep the generating and the checking separate, which is why the output stays on-brand instead of drifting after variant 40.
Here's a CSV of my current ads (headline, description, and performance where I have it) and my brand guidelines: [attach or point to the files].
Step one, analyse what my best performers actually have in common structurally: sentence shape, where the benefit lands, whether they lead with a question, length, and any pattern in the weakest ones. Show me that analysis before you generate anything.
Step two, set up two sub-agents. One generates variants, one reviews them against my brand guidelines and the character limits in the CSV. Nothing reaches me without passing the reviewer.
Step three, generate 100 new variants, each tagged with the angle it uses (e.g. outcome, objection, curiosity, social proof) so I can test angles rather than random lines. Write them to a new CSV in the same column format as my original.
Finally, give me the 10 you'd test first with a one-line reason each, and flag anything that made a claim my brand guidelines don't let me make.