AI coding tools for websites, such as Cursor and Claude Code, are assistants that read a project’s files, propose code changes and run commands on a developer’s instruction. At an agency they speed up repetitive build work, refactoring and debugging, while people still own architecture, design decisions, security review and the final sign-off before anything reaches a client’s live site.
What are Cursor and Claude Code?
Cursor is a code editor built on the open-source base of Visual Studio Code, made by the company Anysphere. It adds AI features directly into the editor: inline completions, a chat panel that can see the open project, and an agent mode that edits several files at once. Developers can choose between models from different providers.
Claude Code is Anthropic’s command-line tool. It runs in a terminal, reads the repository, and can edit files, run tests and use git when the developer allows it. It is also available inside editors and desktop apps. Its strength is working through multi-step tasks, such as tracing a bug across a theme and two plugins, while asking for permission before risky actions.
Both tools are only as good as the context they get. A project with a clear structure, written conventions and working tests gives far better results than a tangle of copied snippets.
Which parts of a website build do they speed up?
The gains concentrate on work that is well defined but tedious. A human still decides what should happen; the tool handles much of the typing and searching.
- Scaffolding: creating block templates, custom post types or component files that follow an existing pattern.
- Refactoring: renaming functions across files, splitting a large stylesheet into component files, removing dead code.
- Debugging: reading error logs, finding where a value goes wrong, and suggesting a fix to test.
- Accessibility passes: adding missing labels, alt attributes and focus styles, then checking contrast.
- Data scripts: one-off imports, URL redirect maps and content clean-ups for migrations.
- Documentation: writing README notes for decisions so the next developer understands them.
Where do humans still make the calls?
Several areas stay firmly with people, whatever the tool suggests. Information architecture, the visual design system and brand choices come from strategy and research, not code generation. So does any decision about what data a site collects and where it goes.
Security is the clearest example. An assistant can write a form handler that works but misses nonce checks, input sanitising or capability checks. Every change that touches user input, payments or authentication gets a line-by-line human review, no matter how clean it looks.
How Alien Road uses Cursor and Claude Code in client projects
Discovery comes first and happens without AI tools in the loop: we agree goals, page types, integrations and content sources with the client. The output is a written brief and a project conventions file that describes folder structure, naming rules and things that must never change. That file becomes context for the coding tools later.
Setup runs in a local development environment, typically Docker, never directly on a live server. Developers use Claude Code or Cursor for tasks such as generating a set of block patterns from approved designs or writing the import script that moves old blog posts into a new structure. For example, a migration might involve hundreds of posts where the tool writes the script, but a developer runs it first in dry-run mode and checks a sample by hand.
Every change goes through git. Commits are small, messages explain the reason, and nothing merges without a human reviewing the diff. Measurement covers build time per page template, defects found in testing, and Core Web Vitals after launch. Clients get a short report that lists what was automated, what was reviewed and what was tested.
How do you keep AI-written code reliable?
The main risk is code that looks right and passes a quick glance but fails on edge cases. Models also tend to add new helper functions instead of reusing existing ones, which slowly bloats a codebase.
Write the rules down
Both tools read project instruction files. Claude Code uses a CLAUDE.md file; Cursor supports rules files. Record the decisions that cost you time to learn, such as why a template file must not be deleted or which function handles paths. The tool then avoids repeating old mistakes.
Test what matters
Run linting and syntax checks after every edit, such as php -l for PHP files. Keep a short list of URLs to load after each change and check for errors. For layout work, check widths like 390, 768 and 1440 pixels before calling a task done.
Limit permissions
Allow read access and file edits freely, but require confirmation for commands that delete data, push code or touch databases. Never paste production API keys into a prompt or a file the tool can commit.
What should a client ask an agency using these tools?
- Is every AI-assisted change reviewed by a named developer before deployment?
- Where is development done, and does any client data leave that environment?
- Which parts of the code are covered by tests or manual checks?
- Who owns the code and the instruction files at the end of the project?
- How are secrets such as API keys kept out of prompts and repositories?
Clear answers to these show the agency treats the tools as instruments, not as a replacement for craft. Vague answers are a warning sign.
What mistakes do teams make when they first adopt these tools?
The most common mistake is asking for too much in one request. A prompt such as “rebuild the header” invites the tool to rewrite files it does not need to touch. Smaller, specific tasks, like moving the search button after the language switcher at widths above 1080 pixels, produce changes a reviewer can check in minutes.
The second mistake is skipping the local copy. Running an agent against a staging server feels faster, but one wrong command can overwrite uploads or settings. A disposable local environment with a fresh database export removes that risk, and it lets developers try two approaches side by side before choosing one.
Where to start this week
If you run your own site, begin with a read-only task. Install one of the tools on a local copy of the project and ask it to explain how the theme loads its styles or where a specific feature lives. Compare its answer with what you know. Then write a one-page conventions file for the project. Even if you never let an assistant edit a line, that file will save the next developer hours.
Keep reading: Web Design and Development · Web Design