Claude Code best practices when your first website is the whole project
If one website is your whole project, five habits from Anthropic's guidance do most of the work: say what finished means, read a plan first, work in small steps, keep a notes file, check on a phone.
Short answer: you need about five habits, and none of them needs a terminal. Anthropic’s overview page says Claude Code runs in a desktop app and in the browser at claude.ai/code as well as in the terminal and code editors, and every habit below works in any of them.
Why most Claude Code best-practice lists are written for someone else
Anthropic’s own best-practices page says it collects patterns that work for “Anthropic’s internal teams and for engineers using Claude Code”. So it covers permission allowlists, hooks, skills, custom subagents, parallel sessions and scripts that run Claude over thousands of files. Most of the articles that rank for this search go further down the same road.
If your whole project is one website (a portfolio, a page for a class, a CV), most of that can wait. Below is the part of the same advice that matters when the site is the entire project, in plain English, alongside what actually went wrong in our recorded build runs.
Say what ‘finished’ looks like before Claude starts
The first section of Anthropic’s page is “Give Claude a way to verify its work”. The reasoning is simple: Claude stops when the work looks done, and without a check it can run, “you become the verification loop”. The page also asks for specific prompts: “Claude can infer intent, but it can’t read your mind.”
For a website, write two or three lines on what finished means before you ask for anything. “Make me a nice homepage” leaves a lot of room. “A homepage with my name at the top, three photos of my work and a contact button that opens an email. Readable on a phone with nothing spilling off the side” leaves much less. Then ask Claude to check its work against those lines and tell you what it checked. Anthropic’s page says to have Claude “show evidence rather than asserting success”.
Asking Claude to check its work makes a real difference. In our landing page run, the first version of the sign-up form never showed its messages, so “Please add your name” and “Press Send” never appeared. Claude found the cause itself, but only because the next prompt asked it to try to break the form.
Ask for a plan and read it before anything is built
Anthropic recommends four phases: explore, plan, implement, commit. Its warning: “Letting Claude jump straight to coding can produce code that solves the wrong problem.”
Plan mode is built for this. The permission modes page says Claude “reads files, runs shell commands to explore, and writes a plan, but does not edit your source”, and edits stay blocked until you approve the plan. At the time of writing (9 October 2026), that page says you switch to it with the mode dropdown next to the prompt box on claude.ai/code or the mode selector in the desktop app, or by starting a single message with /plan.
Read the plan the way you’d read a builder’s quote before they knock a wall down. If it mentions a booking system and you wanted a contact form, you’ve caught it before anything was built. In our blog run, Claude’s first reply said the home page would build its own list of posts. Its second reply said it would update the list itself each time instead, which means every new post is at least two files to upload. Both choices are reasonable. You only find out which one you’re getting by reading.
The same page says plan mode “adds overhead” and gives a rule for skipping it: “If you could describe the diff in one sentence, skip the plan.” Fixing a typo doesn’t need a plan. A new page does.
Work in small steps and keep every version
Anthropic’s page builds most of its advice on one constraint: Claude’s context window (everything in the current conversation) fills up, and “performance degrades as it fills”. When it’s nearly full, Claude “may start ‘forgetting’ earlier instructions or making more mistakes”.
In practice:
- Ask for one change per message: the header, then the photos, then the contact section.
- Start a fresh conversation for an unrelated job. The page calls mixing them “the kitchen sink session”.
- If you’ve corrected the same thing twice and it’s still wrong, start again with a better first message that includes what you learned. The page says a clean session with a better prompt “almost always outperforms” a long one full of corrections.
Keep every version as well. Claude Code saves checkpoints you can rewind to, but the page warns they only track changes made through Claude’s file editing tools and are not a replacement for git. The simplest habit with no terminal is to copy your whole site folder before any big change and put the date in the copy’s name.
After each change, ask Claude to list every file it changed. Our blog test shows why: when we uploaded the new post page without the updated index.html, the post was online but the home page didn’t list it.
Keep a notes file so Claude remembers your decisions
Anthropic’s memory page puts it plainly: “Each Claude Code session begins with a fresh context window.” CLAUDE.md is how decisions carry over. It’s a file you write in plain text, kept in your project folder, and Claude reads it at the start of every session. Typing /init asks Claude to draft a starter one for you.
The page suggests keeping each CLAUDE.md under 200 lines, and the best-practices page gives a test for every line: “Would removing this cause Claude to make mistakes?” It also says Claude treats the file as context, which means a strong steer it can still miss, especially when the file gets long.
For one site, a short file is plenty. Here’s one for a made-up ceramics studio:
- The site is for Hollow Clay, a small ceramics studio (made-up example).
- Friendly, plain wording. No jargon.
- The only way to get in touch is the email address on the contact page.
- The site goes live on GitHub and Cloudflare Pages: plain files, nothing to install.
That last line would have saved a detour in two of our runs. In the blog and CV website builds, Claude’s first going-live steps were for Netlify, until we said we already had GitHub and Cloudflare accounts.
Check the result on your phone, not just in Claude’s description
Anthropic lists “the trust-then-verify gap” as a common failure: a plausible-looking result that misses things. Its fix: “If you can’t verify it, don’t ship it.”
We test every build at phone size in a browser, and the first drafts show why. In our website run, the first build passed 78 of 94 checks: two pages had no main heading, there was no ‘skip to content’ link for keyboard users, and the smallest text was 11.5 to 12.5px. The follow-up prompt fixed the first two. In the CV website, small capital labels were still 12.8px even after Claude’s own check. And in the landing page run, we never tried a real phone, which is why that page says to work through Claude’s test plan before you share the link.
A five-minute phone check:
- Open the live address on your own phone.
- Scroll every page. Does anything make the page slide sideways?
- Can you read every line without zooming?
- Tap every link and button, and fill in any form.
- If something’s wrong, send Claude a screenshot and describe what you see. The best-practices page says you can paste images straight into your message.
What to ignore until you have a live site (hooks, subagents, test-driven development)
- Hooks are scripts that run automatically at set points. The page says that, unlike CLAUDE.md, “hooks are deterministic and guarantee the action happens”. That’s useful when a rule must never be skipped, and too much setup for a first site.
- Custom subagents are helpers that work in their own separate conversation, for jobs that read lots of files.
- Writing tests first, and the page’s pattern of one Claude writing tests while another writes the code, belong to bigger codebases.
- Connecting outside tools, permission allowlists and running several sessions at once can all wait.
Come back to these once the site is live and you’re doing the same job for the twelfth time.
Where beginners go wrong and how to put it right
Most of these come straight from Anthropic’s list of failure patterns:
- Everything in one conversation. Start a fresh one for each separate job.
- Correcting over and over. After two tries, start again with a clearer first message.
- A notes file that grows into an essay. Cut any line Claude would get right anyway.
- Trusting the description. Open the real page on your phone.
- Uploading half a change. Ask Claude which files changed, and upload all of them.
If you haven’t set anything up yet, start here. If you’re still deciding where the site should live, read the comparison of Wix, Squarespace, WordPress and Cloudflare.
Questions
Questions people ask.
What are the best practices for using Claude Code?
Anthropic's best-practices page covers giving Claude a way to check its work, exploring and planning before building, writing specific prompts, keeping a short CLAUDE.md file, and managing the conversation so it doesn't fill up. For one website, that becomes five habits: say what finished means, read a plan first, work in small steps, keep a notes file, and check the result on your phone.
Do I need a CLAUDE.md file to build one website?
No, but it helps. Anthropic's memory page says each session starts fresh and Claude reads CLAUDE.md at the start of every one, so the decisions you write there don't need repeating. For one site, a dozen short lines is plenty. Anthropic suggests keeping it under 200 lines.
Should I ask Claude to plan before it writes code?
For anything bigger than a small fix, yes. Anthropic recommends exploring and planning before building, and plan mode lets Claude write a plan without editing your files until you approve it. For something you could describe in one sentence, like a typo, its guidance says to skip the plan.
How do I stop Claude changing things I did not ask it to change?
Ask for one change per message, say exactly what to change and what to leave alone, read the plan before you approve it, and copy your site folder before any big change. After each change, ask Claude to list every file it changed so you know what to upload.
Which best practices can a beginner skip?
Hooks, custom subagents, writing tests first, connecting outside tools and running several sessions at once. Anthropic's own page aims these at engineers. Come back to them once your site is live and you find yourself repeating the same job.