How to Build $10,000 Websites With Claude Code (Full Course, 2 Hour Masterclass)

Agencies charge $10,000 for websites Claude Code can build in an afternoon. That number confuses people, because they assume the money is for the animation. It is not. You probably think a beautiful interactive site is what makes you money, but what a client actually pays for is conversions. People buying their product.

This is the written companion to that masterclass. It covers the anatomy rather than the keystrokes: what each piece of a premium site is doing, why it sits where it sits, what you have to decide at each step, and what breaks. Everything here you can run by hand tonight.

Start with the test that decides whether you get paid. A site that looks good and a site that gets buyers to fill in the contact form are two completely different things, and a client only pays for the second one. Hold every decision below against that test.

A premium site says one thing. One. That is what cuts confusion and speaks to the person who actually wants to buy. Most builds fail here, because the builder starts by describing a look instead of a claim.

The move that fixes it is to stop writing a brief and make Claude interview you instead. Ask for exactly four questions, one at a time, and answer them plainly. What is the product or brand in one line. Who is it for, and what should they feel. What is the one thing they should remember. And do you have an AI render tool connected, or should the build use stock imagery. Four answers, no essays.

In the masterclass the product is an Audi R8, and the answers are short: premium luxury car buyers and thrill seekers, a feeling of excitement and speed, and one thing to remember, that this is their dream car and it could be in their driveway this week. The first output is deliberately ugly. A bare page, no palette, no motion. That is the point. You are proving the message and the structure work before you spend anything on the look.

Scroll style is how the page moves. The header motion is the first thing your customer sees, and it decides how they feel about the product before they read a word. It is not decoration, it is the first impression.

The instruction that matters: look at the product and the thesis, pick one scroll style, and defend the choice in a single line. Three that work for physical products are a looping video header, a scroll to play sequence where the video advances as you scroll, and an exploded view that separates the product into parts.

Forcing a one line defence is doing real work. Left alone, Claude stacks three effects on top of each other and the page feels busy and cheap. One motion idea, executed properly, is what reads as expensive. The same pass also lays the scaffolding for the rest of the page: hero, thesis, feature blocks, gallery, and a single call to action.

One trick to skip. The popular hero right now is a 3D exploded animation built from a render, then animated as a start frame to end frame transition. It looks great in a demo and it does not sell anything. It is the clearest example of the looks-good trap, so spend that effort on the copy instead.

This is the one idea on this page that transfers to everything else you make. If you take one thing away, take this.

When you ask for colors by describing the brand, you get the average of every site in that category. Say sports car and you get red, black, and a gradient. Say premium and clean and you get the pale grey template everyone recognises as AI output. Adjectives have no light in them.

So do not describe the brand. Describe a scene, in four moves. Ask for one sentence only. Who is using this product. Where are they, and in what light. Then take the darker light from that scene and build the palette out of it.

For the R8, the sentence that came back was a buyer taking his new car out alone at dusk on an empty canyon road just after sunset. Every color on that finished site is in that sentence already. A different sentence for a different product, a smith pulling steel from the fire after dark, produces a completely different site, and it is obviously right the second you scroll it.

It works because a scene has real light in it, real time of day, real surfaces, and light has specific colors. Adjectives do not. You have turned a taste decision into a procedure you can repeat, which means it works for thumbnails, decks, and ads too.

Copy is every headline and every proof statement, and premium copy has rules. Expensive sites are quiet. Headlines stay short. Captions run three to five words with maybe a couple of accented words. Claims get very specific, because two thirds less acid beats smooth and delicious every single time. Nothing restates the headline, and there are no em dashes anywhere.

You can see the standard in the R8 lines. Production of the R8 ended in 2024, so every R8 that will ever exist already does, and one of them can be in your driveway by Friday. Then 8,700 RPM. Then drive it to the office on Monday, then take the long way home through the canyon and arrive grinning. Short, specific, no adjectives doing the work.

The bigger idea is where those sentences come from. Do not write in your own words. Send a research agent out to Reddit threads, reviews, and forums in the niche and have it collect the exact sentences real buyers write. What hurts, what they want, why they hesitate. Then write the page in those words.

The HVAC build shows what that finds. The number one word in five star HVAC reviews is honest. The industry has a trust problem so severe that customers are shocked when a tech does not try to upsell them. No amount of taste gets you to that. It produced the line, your AC breaks on the hottest day of the year, we pick up the phone, we show up, we fix it honestly.

Using your audience's own language back to them is the whole conversion mechanism. They read the page, see themselves in it, and conclude that whoever wrote this understands them, so the offer must be the right fit. Run the same research on your own offer before you build for yourself. One pass produced a full competitive landscape plus thirteen quotable customer sentences.

Less is more, and there is a hard budget: the whole site runs on five assets. Two videos and three images. Same lighting, same color palette, a mix of close ups and wide shots.

The restraint is what makes it read as expensive. A wall of twelve AI images with mismatched lighting reads as a template no matter how good any single image is. Five consistent assets read as a shoot.

Have Claude write the render prompts and then fire the render tool itself in the same pass, so the prompts inherit the palette and the scene sentence you already locked. If the client already has footage, the strongest move is to turn their own MP4 into a scroll driven image sequence. The site stays theirs, and it costs you nothing to render.

The last pass wires everything up and verifies it. Every section present and in order. Every image and video actually placed. The palette applied everywhere, including the states nobody checks. One call to action, and everything funnelling to it.

Then have the build test itself. It can take screenshots of the page at phone width, tap every button, and submit the form. That is how you catch the dead link and the button that goes nowhere before a client does.

Two things you will hit on this pass. Contrast bugs are common and the fastest fix is a screenshot plus a plain sentence: this is not easy to read, please make it better. And the same prompt run twice will not give you the same site. There is a real element of chance in these tools. Commit to git before any big rebuild pass, because the version you preferred can quietly disappear.

The research step does something more valuable than harvesting sentences. It shows you who the competition is all shouting at, which tells you who nobody is talking to.

The plumbing build found it immediately. Every plumbing site screams identically at the panic Googler with a burst pipe. Nobody designs for the $8,000 repipe decision. That is a considered, expensive, planned purchase, and the person making it is being served by nothing at all.

So the site went after that buyer instead. Copper Line Plumbing, whole home repipes and ground up plumbing for houses worth doing. The page opens with a line aimed straight at an old house: your house was built in 1926, so were its pipes. Then it names the real fear without shouting. Small leaks never stay small, and every year you wait the water gets closer to the floors, the framing, the things that cannot be replaced.

The reassurance is just as specific. One walkthrough, one number, a fixed bid line by line that stays fixed, three days of work, a lifetime workmanship warranty. It closes with the actual promise the buyer wants: you will never think about your pipes again, book a walkthrough.

• Almost every local trade has two buyers: the emergency one everyone competes for, and the big ticket planned one nobody addresses.
• The emergency buyer picks on speed and price, which is a race you cannot win for a client. The planned buyer picks on trust and clarity, which is exactly what a good page supplies.
• Prospect the same way. Search a city plus the trade on Google Maps and look for strong ratings with a modest review count, because those businesses could use more customers.
• That filter is how the HVAC target was picked: 4.9 stars and 694 reviews, good enough to be real, small enough to want growth.

Most of the difference between a $500 looking site and a $10,000 one is a handful of choices made on purpose instead of by default. Every one of these has a tell, so you can pick without guessing.

• Plan mode or auto edit. Auto edit goes straight to changing files. Plan mode states the entire game plan first. Use plan mode for anything touching more than one file: stronger result, far less back and forth. The tell that you got this wrong is spending your session undoing work.
• Which model. Sonnet 4.6 for almost everything, the sweet spot on speed and cost. Haiku for quick, easy responses. Opus 4.6 when a job is genuinely complicated, like reading a messy file and writing a build spec from it. Fable 5 for generating and self testing a whole site in one run. The tell to move up is a long multi step build losing the thread halfway through.
• Default effort or the highest effort mode. Default while you build piece by piece and want to inspect each one. Highest effort when you want a single prompt to spawn a team of sub agents to build plus another team to check the work. One run produced three complete draft sites, scored them against the design rules, and picked a winner.
• Terminal, an agent IDE, or the desktop app. The terminal is fine and fastest for the piece by piece build. An agent IDE earns its place on complicated jobs, because you can watch every file being written, run parallel agents in tabs, and comment inline on the plan before approving it. The desktop app is the pick when you need a connector, like reading confirmations out of Gmail.
• One off build or a packaged skill. Build it by hand the first time so you understand every piece. The second time you build for the same industry, package it. After that a company URL plus a note like make it dark mode returns a finished site in minutes, which is the difference between selling hours and selling a system.
• MVP or perfect. Ship the working version and start taking money before you make everything perfect. The polish list you write before launch is almost never the list customers give you.

Troubleshooting is most of the job. Anyone telling you these builds just work is skipping the part where you sit and figure out what went wrong. These are the failures worth knowing in advance.

• The generic AI look. Card layouts, small emoji, gradient headers. The cause is level one prompting, meaning you described what you wanted in plain language and stopped. The fix is to have a chat model write the detailed spec first, then apply a design skill on top.
• Non determinism. Run the same prompt twice and get two different sites, and sometimes the second one is worse. Commit before every big pass so you can go back to the version you liked.
• Looks good, worth nothing. Before you send anything to a client, ask what on this page makes someone fill in the form.
• The API key precedence bug. A chat feature that will not respond and reports that it cannot connect to the server. The cause is often two keys: one set correctly inside the project and a blank one at the system level that overrides it. Work out which key wins before you hunt in the code.
• Missing environment variables after deploy. The site works locally and dies live. Every key has to be added in the host dashboard too: the model API key, the scraper key, and every payment link. With three pricing tiers, that is three separate links.
• An authentication token error on the first local run even though the keys are in place. Usually the dev server needs restarting so it picks up the file. Kill it and start it again before you debug anything else.
• Not reading the generated spec. When you have a model write the build spec, read it. One run quietly specified user accounts, per user connections, and dashboards for a product meant to be one input and one table. Cut that before it gets built, not after.

A website is a one time invoice. There are two ways to make the same work pay more than once, and both come out of the build you already did.

The first is the brand guidelines wedge. While researching the client you scrape their live site for their logo, colors, fonts, spacing, and images, and save all of it into a single brand guidelines file. That file is the retainer. With it you can produce their emails, ads, and landing pages on demand, in their brand, without asking them for anything. You stop being the person who built the site and become the person who has their brand on file. That is what makes the ongoing relationship obvious rather than a pitch.

The second is building a product instead of a page, because a web page sells once and a product pays every month. It does not have to be ambitious. The version in the masterclass takes a YouTube URL, pulls the comments, filters them for questions and suggestions, and returns content ideas with hooks and outlines in a table you can export. One run turned 100 comments into 28 usable ideas. The idea itself came from an automation that already existed, and the validation came free: the post sharing that automation had 732 likes and 990 comments.

The margins are the part worth copying, because they are small and real. The comment scraper costs $2 per 1,000 comments, so a 100 comment run costs about 20 cents. Charging 99 cents leaves roughly 70 cents after a 10 cent buffer for the model calls that write the ideas. Around 70% margin on a product that runs while you sleep.

One more habit worth stealing: pay for your own product with your own card to test it. Test mode is a separate project of its own, and a real 99 cent charge landing in your dashboard proves the whole path in one go.

The shortest honest sequence from finishing this guide to a real site on a real URL. One evening, one target, no shortcuts skipped.

1. 1Pick one real target. Search your city plus a trade on Google Maps and shortlist a business with strong ratings and a modest review count.
2. 2Make a folder for it, open Claude Code inside that folder, set the model, and switch to plan mode before you type a prompt.
3. 3Scrape the target's existing site for their branding and images and save all of it into a single brand guidelines file you keep.
4. 4Send a research pass to Reddit, reviews, and forums for the exact sentences their buyers write. Save the quotes to a file, and note which buyer every competitor is ignoring.
5. 5Run the four question interview and answer in one line each. Accept the ugly first page as proof the structure works.
6. 6Pick one scroll style and make it defend the choice in a single line before it builds anything.
7. 7Write the scene sentence, take the darker light from it, and build the palette from that.
8. 8Write the copy from your harvested quotes file, not from your head. Short headlines, specific claims, one call to action.
9. 9Render exactly five assets, two videos and three images, in the palette you just locked.
10. 10Run the spine pass, then have it screenshot itself at phone width, tap every button, and submit the form.
11. 11Create a private GitHub repo, push to it, then import it into your host and add every environment variable during the import rather than after.
12. 12Open the live URL on your phone, fill in the contact form yourself, and confirm the message arrives. Then make the call.

The $10,000 is not for the scroll animation. Anyone can produce a scroll animation now, which is exactly why it stopped being worth money. It is for knowing who the buyer is, saying it back to them in their own words, and building the one page that gets them to act.

Everything in this guide is a procedure, not a talent. The interview, the scene sentence, the harvested copy, the five assets, the verification pass. Run them in order on one real local business this week. Then pick up the phone and start selling some websites.