How to write Claude Code prompts for website changes

A good Claude Code prompt for a website change says five things: which pages or files to work on, exactly what should change, what must not change, how to check the result, and what to show you before editing anything. "Make the services page better" leaves Claude guessing on all five. "On /services/roof-repair/, rewrite the meta description to under 155 characters using only facts already on the page; list the old and new text and wait for my approval" leaves it guessing on none.

You don't need special syntax or magic phrases. Anthropic's own advice is to "talk to Claude like you would a helpful colleague," and to be specific: "The more precise your instructions, the fewer corrections you'll need." This guide turns that into a method for website work, with before-and-after examples, prompts for common jobs, and the things a prompt can't do for you, as of October 2026.

The five parts of a website prompt

Not every prompt needs all five spelled out. A one-line typo fix doesn't need a plan. But the bigger the change, the more of these you should state, and the more each one saves you in review.

Before and after: vague versus specific

Anthropic's best-practices page shows the pattern with coding examples: scope the task, point to sources, reference existing patterns, and describe the symptom. Here's the same idea applied to website jobs.

Anthropic notes that vague prompts have a place: "what would you improve in this file?" can surface things you wouldn't have thought to ask. Use them for exploration, then switch to specific prompts for the actual change.

Ask for a plan or a list first

The single most useful habit is to ask for a list or a plan before any file changes. Reviewing a list takes a minute. Undoing a bad site-wide edit can take a day.

• Plan mode. In the terminal, press Shift+Tab until the mode reads Plan, or start one prompt with /plan; in the desktop app, pick Plan in the mode selector. Typing the words "plan mode" in a prompt doesn't switch modes. Claude reads files and explores, then presents a plan for approval without editing your source. Our guide to Claude Code plan mode covers the approval options.
• "List first" in the prompt. End with "list what you'll change and wait for my approval." This works in any mode.
• Approve a sample. For repeated edits, ask Claude to do one page, show you the result, and wait. Then do the rest.

Anthropic's guidance is that planning helps most when you're unsure of the approach, the change touches several files, or you don't know the code. "If you could describe the diff in one sentence, skip the plan."

Give Claude a check it can run

Claude stops when the work looks done. Without a check, "looks done" is all it has, and you become the only test. Anthropic's advice: "Give Claude a check it can run: tests, a build, a screenshot to compare." On a website, useful checks include:

• The site's build command, which catches broken templates and syntax errors.
• A link checker run against the built site.
• A count: "confirm the old phone number appears zero times in the built site."
• A screenshot of the changed page at mobile and desktop widths, compared with the original.

Then ask for evidence, not reassurance: "paste the last lines of the build output" or "show me the count." Anthropic notes that reviewing evidence is faster than re-running the check yourself.

Point to files and sources instead of describing them

• Use @ to reference a file. Typing @ followed by a file name pulls that file into the conversation, so Claude reads it before answering.
• Keep business facts in one file. Prices, phone numbers, addresses and service areas belong in a single facts file that your prompts and CLAUDE.md point to. Claude can't know your new Saturday hours unless you tell it, and it can write a believable wrong one.
• Give URLs for rules. If a change depends on a platform rule or a law, give Claude the official page to read and ask it to quote the line it relied on.
• Attach a screenshot when describing a visual problem. The desktop app supports attaching images and PDFs.

Anything Claude writes about your business, its prices or a regulation is a draft until a person checks it against a primary source. Our guide to common Claude Code mistakes on websites shows what happens when that step is skipped.

Put permanent rules in settings, not just prompts

"Don't push" in a prompt is useful, but it isn't a guarantee. Anthropic's docs say a boundary stated in conversation "can be lost if context compaction removes the message," and that CLAUDE.md content is treated "as context, not enforced configuration." For rules that must hold every time, use the right layer:

Our guide to writing a CLAUDE.md for a website covers the middle layer, and our templates page has a permissions file and eight copy-ready prompts for weekly website jobs.

For bigger jobs, let Claude interview you

For a new page type, a form or anything with many decisions, Anthropic suggests starting small and having Claude ask the questions. A website version of the prompt Anthropic publishes:

I want to add a quote request form to our service pages. Interview me in detail: ask about fields, where submissions go, spam protection, tracking, confirmation messages and edge cases. Don't ask obvious questions. When we've covered everything, write the spec to SPEC.md and stop.

Then start a fresh session to build from the spec (the written description of what to build), so the build session starts clean. Anthropic says the most useful specs name the files involved, state what's out of scope, and end with a verification step that proves the feature works.

When the answer is wrong

• Stop early. Press Esc the moment Claude heads the wrong way. Context is kept, so you can redirect.
• Undo. Press Esc twice on an empty prompt, or run /rewind, to restore an earlier point in the session.
• Start over after two corrections. Anthropic: "If you've corrected Claude more than twice on the same issue in one session, the context is cluttered with failed approaches." Run /clear and write a better first prompt that includes what you learned.
• One task per session. Mixing unrelated jobs in one conversation is what Anthropic calls "the kitchen sink session." Use /clear between them.

Specific prompts also cost less. Anthropic notes that vague requests "trigger broad scanning," while specific ones let Claude work with fewer file reads. More on that in our guide to keeping Claude Code costs under control.

Prompts for common website changes

These are illustrative. Replace the file names and details with your own.

For larger, repeated jobs such as rewriting titles across hundreds of pages, see our step-by-step on title and meta rewrites with Claude Code.

What a prompt can't do

A well-written prompt makes a good result more likely. It doesn't make the result correct. Claude can still misread a template, invent a plausible fact or change a file you didn't mention. The review step stays: read the diff, check the page on staging, and have a person approve anything that goes live. Our guide to reviewing a diff when you aren't a developer shows what to look for.

Conclusion

Write website prompts the way you'd brief a careful contractor: name the pages, describe the change, set the limits, ask for a plan or a list first, and give Claude a check whose result it has to show you. Put the rules that must always hold into permission settings, start fresh when a session goes sideways, and keep a person reviewing what ships. If you'd like your team to have a shared prompt library, CLAUDE.md and review routine, our Claude Code consulting includes that setup and training. More guides are in our Claude Code Guide.

Frequently asked questions