Claude Code is Anthropic's agentic coding tool. It runs in your terminal, reads your project files, edits code across multiple files, runs commands such as tests and builds, and works through tasks you describe in plain English. To build a client app with it, set up a project folder, write a CLAUDE.md file that defines your stack and rules, and give Claude Code small, scoped tasks that you review before they merge.

This guide is for agencies and senior developers who want a repeatable process for client work. It covers the terminal workflow, how to use CLAUDE.md as project memory, a walkthrough of a typical build, and a decision framework for when Claude Code is the right tool compared with Cursor or Lovable.

What Claude Code Does Differently

Most AI coding assistants suggest code inside the file you have open. Claude Code works at the project level. It searches the repository, reads the files it needs, plans a change, applies edits across the codebase, and runs your commands to check the result. You stay in control through permission prompts and by reviewing every diff.

That matters for agencies because client work rarely starts from zero. Most engagements begin with an existing repository, earlier technical decisions, and a team that will read the code long after you have left.

Set Up the Terminal Workflow

A reliable setup has four parts: a git repository, a documented stack, a local environment that runs, and a single test command.

Step 1: Initialize the repository

Create the project folder and initialize git before your first prompt. Every change Claude Code makes should be reviewable as a diff, and a clean starting commit makes rollback simple.

mkdir client-booking-app && cd client-booking-app
git init
git commit --allow-empty -m 'Initial commit'

Step 2: Start a session in the project root

Launch Claude Code with the claude command from the folder that contains your code. The session treats the current directory as its working context, so starting from your home folder gives the agent the wrong scope.

Step 3: Give scoped tasks

Ask for one outcome at a time. A task such as 'Add a booking form with name, email, and preferred date fields, validate the inputs on the server, and save to the bookings table' is a good size. 'Build the whole app' is not. Scoped tasks produce smaller diffs, fewer hidden assumptions, and faster review.

Step 4: Review, test, and commit

After each task, read the diff, run the test suite yourself, and commit with a message that describes the change. If the output is wrong, correct it in the same session or revert the commit. Do not accept multi-file changes you have not read.

CLAUDE.md: Project Memory for Client Work

CLAUDE.md is a markdown file in the project root. Claude Code reads it at the start of a session, which makes it persistent project memory. Without it, you repeat the same instructions in every prompt, and the agent falls back on its own defaults.

A useful CLAUDE.md is short and specific. Include:

• The stack and versions, such as framework, database, and hosting target
• Commands for install, dev server, tests, lint, and build
• Folder structure, including where components, routes, and tests belong
• Coding conventions the team already follows
• Hard rules, such as never committing secrets or editing generated files

Example CLAUDE.md

# Client Booking App
Stack: Next.js (App Router), TypeScript, PostgreSQL via Prisma
Commands: npm install | npm run dev | npm test | npm run lint
Structure: UI in /app, data access in /lib/db, tests beside source files
Rules:
- TypeScript strict mode. No any types.
- Validate all form input on the server.
- Never edit files in /generated or commit .env files.
- Every new feature ships with at least one test.

Update CLAUDE.md when a decision changes. Treat it as documentation the agent reads at the start of every session, because that is exactly what it is.

Walkthrough: Building a Client Booking App

This is how a typical scoped build runs. The sequence matters more than the specific features.

1. Scaffold the project. Ask Claude Code to create the framework skeleton defined in CLAUDE.md, then run the dev server to confirm it starts.
2. Model the data. Define the schema for clients, appointments, and availability. Review the migration before applying it.
3. Build the core flow. Implement the booking form, server-side validation, and confirmation view as separate tasks.
4. Add authentication. Name the auth provider in CLAUDE.md first, then request the integration. Verify role checks with tests rather than by eye.
5. Wire integrations. Connect email confirmations and calendar sync one at a time, each with a sandbox key.
6. Harden before launch. Run lint, tests, and a dependency audit. Review environment variable handling and the error messages users see.

Each step ends in a commit. By the end, the client has a readable repository, a test suite, and a CLAUDE.md that the next developer can pick up without a long handover.

Claude Code vs Cursor vs Lovable

These tools overlap, but they serve different buyers. The table compares them on the criteria that usually drive agency decisions.

Tool	Interface	Best for	Who should use it
Claude Code	Terminal agent (also available in some IDEs)	Existing codebases, multi-file changes, backend logic, scripted and repeatable tasks	Engineers comfortable with git, tests, and the command line
Cursor	AI-native code editor with chat and agent modes	Hands-on development where the developer edits code directly and wants inline AI help	Developers who want a full IDE with AI built in
Lovable	Browser-based chat that generates and hosts a full-stack app	Fast UI-first prototypes, MVPs, and stakeholder demos	Founders, product teams, and agencies that need a working demo quickly

When an Agency Should Use Claude Code

• The client already has a repository that needs new features, fixes, or migration work
• The project involves complex backend logic, data migrations, or refactors across many files
• The client requires code review, automated testing, and version control standards
• You need a documented, repeatable process that another engineer can audit

When Lovable or Cursor Is the Better Fit

• Lovable when the goal is a clickable prototype in days, a validation build, or a UI-heavy MVP that the client's team will take over later
• Cursor when your engineers want to stay in an editor, write most of the code themselves, and use AI for suggestions and targeted edits
• Claude Code when the work is code-first and a bad architectural decision would be expensive to undo

One practical pattern is to prototype in Lovable to test the flow with stakeholders, then move the project into a repository where engineers use Claude Code for hardening, tests, and integrations. To see how we scope AI-assisted engagements, visit our AI development page. For the Lovable side of the workflow, see our Lovable development services.

Review and Security Practices for Client Delivery

• Require human review of every diff before merge, especially for authentication, payments, and data access
• Keep secrets in environment variables and never paste production keys into a prompt
• Review each permission prompt before approving it, and limit what the session can run
• Write tests for business rules first, then have the agent implement against them
• Record architectural decisions in CLAUDE.md so the client inherits context, not just code

Frequently Asked Questions

What is Claude Code?

Claude Code is Anthropic's agentic coding tool. It runs in your terminal, reads your project files, makes multi-file edits, runs commands such as tests and builds, and carries out tasks you describe in plain English.

Do I need to be a developer to use Claude Code?

In practice, yes. Claude Code assumes familiarity with the command line, git, and a local development environment. Non-engineers can build early prototypes in a browser-based tool such as Lovable and bring in engineers for production work.

What is a CLAUDE.md file?

CLAUDE.md is a markdown file in your project root that Claude Code reads at the start of a session. It stores project-specific instructions such as the stack, commands, coding conventions, and rules the agent must follow, so you do not repeat them in every prompt.

When should an agency use Claude Code instead of Lovable?

Use Claude Code when the client already has a codebase, needs complex backend logic, or requires code review, testing, and version control standards. Use Lovable when speed to a clickable prototype matters more than code-level control.

Can Claude Code build a production client app on its own?

Not without engineering oversight. Claude Code can produce working code quickly, but production readiness depends on the review, testing, security configuration, and deployment practices your team applies. Treat its output like a pull request from a capable developer.

Next Steps

If you are new to building with Claude more broadly, start with our guide to building a web app with Claude AI. If you are scoping a client build and want an engineering team to run this workflow, talk to Revex about AI development.