Claude Code first-week starter playbook for UK indie hackers: the four commands you actually use, your first feature, and the mistakes that burn a week

Claude Code went from "developer terminal toy" to the default agentic coding environment for solo founders in about eighteen months. If you're a UK indie hacker opening it for the first time this weekend, the official quickstart will get you installed and running in ten minutes — and then leave you staring at a prompt without a plan. This is the post that picks up where the docs stop. Day one to day seven, what to install, what to type, what to actually build, and the four mistakes that turn week one into week three.

What Claude Code is in one paragraph

Claude Code is Anthropic's terminal-native coding agent. You install a small CLI, point it at a project folder, and chat with it in plain English. It reads your code, writes patches, runs tests, opens files, runs commands, and asks before doing anything destructive. There's no graphical editor, no sidebar, no extension marketplace. The whole product is a conversation with an agent that happens to have file-system and shell tools. If you've used Cursor, the difference is that Cursor is an editor with AI bolted on; Claude Code is an agent with a terminal underneath. You'll feel the difference within an hour.

Day one: install and the four commands you actually use

Install takes five minutes if you've already got Node 18+ on your machine.

npm install -g @anthropic-ai/claude-code

Then in any project folder:

claude

It'll prompt you to log in. You can either log in with your Anthropic API key (pay-per-token, sane for occasional use) or with a Claude subscription (Pro at GBP 16.50/month, Max at GBP 81/month — the second one gets you a sensible weekly Claude Code allowance). Most UK solo founders should start on Pro, watch usage for a fortnight, and only upgrade if they hit the cap mid-flow.

Once you're in, you have a dozen slash commands available. Four of them are the ones you'll actually use this week.

/init — run this on first contact with any new repo. Claude Code reads the codebase, drafts a CLAUDE.md at the root, and writes itself a project-specific brief. Don't skip this. The CLAUDE.md is the highest-leverage thing in the entire workflow — it's the system prompt for every future conversation. Read what it generated, fix anything wrong, add anything missing.

/clear — wipe the conversation, keep the project context. You'll use this every time you switch tasks. Claude Code's context window is generous (up to a million tokens on Max) but you'll get sharper, faster responses if you start a fresh conversation each time you change topics.

/compact — summarise the current conversation into a tight context block, then keep going. Use this when you're deep in a task and the conversation is getting unwieldy but you don't want to lose the thread.

/agents — spawn a subagent for a specific task. The subagent runs in its own context, does the job, returns a handoff. This is the killer feature: long-running research, full-test-suite runs, repository-wide refactors all happen out of band so your main conversation stays clean.

That's it for week one. The other commands (/model, /permissions, /settings, /hooks, /mcp) are useful, but you can ignore them until you've shipped your first feature.

Day two: write your first CLAUDE.md by hand

The auto-generated CLAUDE.md from /init is a starting point. Replace it. Your week-one CLAUDE.md should answer six questions:

1. What is this project? One paragraph, plain English.
2. What's the stack? Languages, frameworks, hosting, database, payment. Be specific. "Next.js 15, TypeScript, Tailwind, Supabase, Vercel, Stripe" beats "modern web stack."
3. What's already built? Short bullet list. Helps Claude Code orient when you ask "is this thing already wired up?"
4. What's the next milestone? What you're working toward this week.
5. House rules. Things Claude Code should and shouldn't do. EN-UK spelling. No emoji in commits. Always run npm run lint after edits. Test before committing. Ask before deploying.
6. Sensitive files. What to never read or commit. .env, secrets, anything with key in the filename.

A solid first-week CLAUDE.md is roughly 150-300 lines. If yours is longer, you're over-specifying. If it's shorter, you're leaving leverage on the table.

Day three: build your first real feature, not a todo app

Skip the todo app. The fastest way to learn Claude Code is to build a feature you'll actually keep. Pick something with a database write, an authenticated read, and an external API call. Three suggestions for UK indie hackers, all shippable in one focused session:

• Postcode-to-region lookup. UK postcode in, region/county out. Uses postcodes.io (free), one Supabase table for usage caching, one authenticated endpoint, one rate-limit. Teaches you the full request lifecycle.
• Stripe webhook receiver with VAT logging. Stripe sends a checkout.session.completed, you write a row to Supabase with the customer email, GBP amount, VAT band, and timestamp. Useful immediately. Teaches you signed-webhook handling, which 90% of UK SaaS dashboards get wrong on the first attempt.
• Companies House lookup. Company number in, registered name + address + status out. Companies House API is free with an account. Useful for any UK B2B SaaS doing customer onboarding.

Tell Claude Code what you want, in detail. "Build me a UK postcode lookup endpoint at /api/postcode/[postcode] that calls postcodes.io, caches the result in a Supabase table called postcode_cache with a 30-day TTL, returns JSON with region, country, and lat/lng, and returns 429 if a single IP has hit it more than 60 times in an hour." That paragraph gets you a working endpoint in one round-trip. "Add a postcode lookup" gets you four rounds of clarification and probably a nodemailer integration nobody asked for.

Day four: learn to interrupt cleanly

Claude Code will sometimes go off on a tangent. It'll start refactoring something you didn't ask it to. It'll add a dependency you don't want. It'll write a test suite when you wanted a quick prototype. The trick is the escape key. Press it and Claude Code stops mid-action, no questions, no orphaned half-applied diffs.

Combine that with the plan mode toggle (Shift-Tab cycles modes — chat, accept-edits, plan). Plan mode tells Claude Code "show me the plan first, don't touch the files." Use it for anything you're not sure about. Cheap, fast, prevents 80% of the rework you'd otherwise do.

The mental model: Claude Code is a junior developer with infinite stamina and zero ego. Interrupt it the moment the plan smells wrong. Re-prompt with one extra constraint. It'll get sharper every iteration.

Day five: hook it into your real workflow

Two integrations are worth wiring up in week one.

Git. Claude Code commits cleanly if you ask it to. The default commit message style is fine; if you want Conventional Commits or anything bespoke, drop a line in CLAUDE.md. Set a hook in .claude/settings.json to require lint-passing before commit and you'll never push broken code again. Don't auto-push on commit — that's a recipe for accidental main-branch chaos.

Tests. Tell Claude Code your test runner once (npm test, pytest, whatever) and from then on it'll run tests after meaningful changes. The first time it loops on a flaky test until it passes is the first time you'll genuinely trust the agentic loop.

If you want to push further, MCP servers extend Claude Code with structured tools: a Supabase MCP lets it query your database directly; a GitHub MCP lets it open PRs; a Vercel MCP lets it tail deployment logs. Don't install all of them on day five. Pick the one that matches the bottleneck in your workflow and add it. Add the next one when you hit the next bottleneck.

Day six: the four mistakes that burn a week

These are the patterns that send most first-week solo founders backwards.

One. Letting it run unsupervised on a fresh codebase. Day one is the day Claude Code knows least about your code. Day seven is when its instincts are sharpest. Don't hand it autonomy on day one. Watch every edit until you've seen its taste line up with yours roughly twice.

Two. Ignoring /clear. Letting one long conversation sprawl across three unrelated tasks degrades quality fast. Claude Code starts mixing up what file the current task is even editing. The fix is /clear between every meaningful task switch.

Three. Vague prompts. "Add auth" is a six-round-trip request. "Add Supabase auth with magic-link email login, a /login page, a /dashboard page that requires a session, and a Vercel middleware that redirects unauthenticated requests on /dashboard/* to /login" is a one-round-trip request. The agent doesn't read minds. Front-load the constraints.

Four. Mistaking speed for shipping. Claude Code can write 1,200 lines of code in eight minutes. That doesn't mean you should ship 1,200 lines of code in eight minutes. Read the diff. If the diff is too big to read, the task was too big — break it up.

Day seven: review what you've shipped and write next week's plan

End the week with a 20-minute review. Open your project, scroll through the commits, and ask Claude Code to summarise. "Summarise this week's commits, list any patterns you notice, flag any code that you'd want to refactor before it grows further, and suggest the three next features that would unlock the most user value." You'll get a one-page brief that doubles as your week-two plan.

Add anything notable to CLAUDE.md. House rules that worked. Patterns that didn't. Tools you started using. Then /clear and you're set up for week two with a sharper agent than you started with.

What it costs in week one

Realistic UK indie hacker budget for a first week of Claude Code use:

• Anthropic Pro subscription: GBP 16.50/month
• Domain (if you need one): GBP 8/year via 123-Reg or Cloudflare
• Supabase free tier: GBP 0
• Vercel free tier: GBP 0
• Total: under GBP 17 for the whole week, all-in.

If you're hitting your Pro weekly cap mid-flow, the upgrade-to-Max question is: are you genuinely shipping more than four hours a day? If yes, the GBP 81/month Max is a no-brainer. If no, stay on Pro and use the cap-hit moments as natural breakpoints.

Closing

Week one of Claude Code is mostly a calibration exercise. You're learning how the agent handles your taste, your tone, your stack. By Friday you should have shipped one real feature you'd put in front of a paying customer, written a CLAUDE.md that matches how you actually work, and started getting comfortable interrupting the agent the moment it heads sideways. From week two, the same workflow scales — the four commands stay the same, the prompts get sharper, the features get bigger.

If you want a concrete weekend-shaped follow-on, our companion post on shipping a UK SaaS weekend with Claude Code + Vercel + Supabase walks through Friday-evening to Sunday-night. And if you're still picking between Claude Code and Cursor, the UK indie hacker buyer's guide gives the GBP-priced answer for three founder profiles.

Free this week's data-backed UK builder report. Every Thursday IdeaStack publishes one deeply researched UK opportunity — keyword data, SERP gaps, GBP pricing, and a builder prompt. Read the latest free report.