AI Workflows / Field guide
The AI Workflow We Use to Build Beautiful Websites
GeorgePublished
The exact tools we use at Lucent Works to go from a blank page to a fast, beautiful website: where we find the look, how we make and animate images, how we code with Claude Code, and what we check before launch.
The full stack at a glance
People often ask which AI tools actually matter for building a website. There are hundreds, and most guides list them without saying how they fit together. This is the stack we use every day at Lucent Works, in the order we use it.
It is the foundation, not the whole story. Each stage will get its own deeper guide, with more workflows to follow.
- 1. Find the lookInspora · Recent · Godly · ReferoScreenshots, references, a style file for the AI
- 2. Make the imagesChatGPTHero art, product shots, illustrations
- 3. Animate themGoogle FlowShort loops for the hero or a section
- 4. Clean the cut-outsBria RMBG + Claude Code skillTransparent PNGs, batch-processed
- 5. Build the siteClaude Code · Opus 5.5The real code, components and animated details
- 6. Make it fastWebP · Lighthouse MetricsLight images, speed checked from several regions
- 7. Connect the businessKit · CalendlyEmail signups and booked calls
Steps 2 to 4 only apply when the design needs custom images or motion.
Step 1: Find the look before you write any code
A beautiful site starts with good references, not with a prompt. We spend the first hour in four inspiration libraries and collect what we like: full pages, single sections, buttons, type pairings.
| Library | Best for | What we grab |
|---|---|---|
| Insporainspora.design | A curated archive of recent visual design: sites, logos, brand work | Overall mood, colour and type pairings |
| Recentrecent.design | Web, interface, motion and branding posts, updated constantly | Micro-interactions and motion ideas for buttons and sections |
| Godlygodly.design | Bold, experimental websites, filterable by hero, CTA and footer | A specific section to rebuild: a hero, a footer, a pricing block |
| Referorefero.design | Real product UI screens, plus Refero Styles: a site’s look as a DESIGN.md file | Layout patterns and a style file Claude Code can follow |
Take references, not copies. Mix ideas from several sites so the result is yours.
How we collect references:
- Screenshot sections, not whole pages. A hero from one site and a footer from another are easier to brief than a full page.
- Grab code snippets where the site allows it. Some libraries show the CSS or component behind an effect, which saves Claude Code a lot of guessing.
- Use a style file when one exists. Refero Styles turns a site’s look into a DESIGN.md file. Drop it into the project and Claude Code follows the colours, type and spacing.
- Keep 5 to 10 references per project. More than that and the direction gets muddy.
Step 2: Make the images in ChatGPT
We generate hero art, product scenes and illustrations with ChatGPT’s image generation. The trick is to describe the image the way a photographer or art director would: subject, surface, light, palette, framing and where the text will sit.
What makes a website image prompt work:
- One clear subject. Busy scenes compete with the headline.
- Light and palette. Name them; they decide whether the image matches your brand.
- Empty space for text. Say where the headline goes, for example “space on the left”.
- No text, no logos. Add those in code, where they stay sharp and editable.
- The format. Wide 16:9 for heroes, square for cards.
Hero image for a boutique coffee roaster website. A single ceramic cup on a warm stone surface, soft morning side light, steam rising, shallow depth of field. Muted terracotta and cream palette. Lots of empty space on the left for a headline. Photographic, no text, no logos. Wide 16:9.
Animate this image. Keep the camera locked and the composition unchanged. Only the steam moves: slow, soft curls rising and fading. A very slight flicker in the light. Calm, 6 seconds, designed to loop seamlessly. No new objects, no text.
Generate the still in ChatGPT, then upload it to Google Flow with the second prompt. Swap the subject and palette for your brand.
Step 3: Animate the image in Google Flow
A still hero can come alive with a few seconds of subtle motion. We upload the ChatGPT image to Google Flow and use its image-to-video feature to animate it.
The animation prompt in Step 2 shows the pattern we follow:
- Lock the camera. Moving cameras look impressive but distract from the copy.
- Move one thing. Steam, water, light or fabric: one element is enough.
- Ask for a seamless loop. A hero video plays forever, so the end must meet the start.
- Keep it short. A few seconds is plenty and keeps the file small.
When the video goes on the page:
- Export MP4 or WebM and keep it as short and small as you can.
- Autoplay it muted, looping and inline, so phones play it without a tap.
- Use the original still as a WebP poster, so the image appears first and the video fades in.
- Pause it for visitors who prefer reduced motion and show the still instead.
Step 4: Remove backgrounds with a Claude Code skill
Product shots and cut-out objects need clean transparent edges. We use Bria’s RMBG background-removal model, set up from its public repository as a skill inside Claude Code. That means we can say “remove the background from every image in this folder” and get transparent PNGs back in one batch, with the original size kept.
Check the licence before client work. Bria releases the RMBG-2.0 weights under the CC BY-NC 4.0 licence, which allows non-commercial use only. Commercial use needs an agreement with Bria or its paid API. For a business website, sort this out before you ship.
Step 5: Build the site in Claude Code
Claude Code, Anthropic’s coding agent, is our main tool. It writes the actual site: pages, components, forms and the small details that make a site feel premium. We give it the references from step 1, the images from steps 2 to 4, and a short brief for each section.
It runs in a terminal or the Claude desktop app and needs a paid Claude plan. You don’t have to be a developer, but you do need to be comfortable reading what it changed and checking the result in a browser. For build costs and whether to do it yourself or hire someone, see our guide “The Complete Guide to Building a Website with AI”.
We use Claude Opus 5.5 and match the effort setting to the job. Small edits run on low effort, anything new on medium; that gives us the best balance of cost and quality. Signature details, such as an animated button or a hover effect, start on medium too, and we iterate with screenshots until the timing feels right. We raise effort only when a result keeps coming back wrong.
| Job | Setting | Why |
|---|---|---|
| Pages, sections, copy edits, layout fixes | Opus 5.5 · low | Small, clear tasks; in our work low effort is faster, cheaper and usually good enough |
| New components, forms, integrations, bugs | Opus 5.5 · medium | More moving parts; our default for anything new |
| Animated buttons, hover effects, signature details | Opus 5.5 · medium, raise if it keeps missing | Taste and timing matter; iterate with screenshots |
One rule: start low for small edits, medium for anything new, and raise effort only when a result keeps coming back wrong.
Build the hero section for the homepage. Reference: hero.png (attached) for layout and spacing; follow DESIGN.md for colours and type. Assets: /images/hero.webp as the poster, /video/hero.mp4 as a muted, looping background. Content: headline “Coffee roasted this week”, one line of supporting text, one button “Order beans”. Requirements: headline on the left over the empty space in the image; stacks cleanly at 375px wide; video pauses when the visitor prefers reduced motion. When done, open it in the browser at mobile and desktop widths and show me screenshots.
One section, a reference, real assets, the content and the mobile requirement. Then check the screenshots before moving on.
Habits that keep the code clean:
- One section per request. Small, clear tasks give better results than “build the whole site”.
- Show, don’t describe. Paste the screenshot or snippet you want matched.
- Review in the browser. Check every change at mobile and desktop widths before moving on.
Step 6: Convert every image to WebP
In our builds, oversized images are the usual reason a beautiful site feels slow. We convert every image to WebP and size it for the space it fills on the page. Claude Code can do the conversion in bulk with a free command-line tool, cwebp. WebP keeps transparency, so the cut-outs from step 4 convert without losing their clean edges.
| File | PNG | WebP | Smaller by |
|---|---|---|---|
| lucent-symbol@2x.png | 221 KB | 112 KB | 49% |
Converted with cwebp at quality 80. Savings vary with the image: photos and soft 3D renders shrink most, flat graphics less.
On our own site, a 3D logo PNG dropped from about 979 KB to 27 KB with no visible difference. A flatter graphic shrank by about half. Both are worth it.
Step 7: Check speed from several regions
Before launch we test the site with Lighthouse Metrics. It runs Google’s Lighthouse audit from different regions and keeps a history of past tests, so you can see how the site performs for visitors far from your server and whether a change made things better or worse.
What we look at:
- Performance on mobile. Many visitors arrive on a phone, so test mobile first.
- The biggest image or block on first load. If it is slow, resize or compress it.
- Layout shifts. Elements that jump while the page loads feel broken.
- Repeat tests after each fix, so you know which change helped.
Step 8: Connect the business tools
A website that looks good but captures nothing is a brochure. Two integrations turn ours into a working sales tool:
- Kit for email. Our guides signup sends each new address to Kit and tags each person by language and by the topic they care about, then sends a confirmation email. New guides and follow-up emails reach the right people.
- Calendly in the contact window. Visitors pick a time for a call inside the site, without email back and forth.
- Every image is WebP and sized for the space it fills
- Lighthouse scores checked from more than one region, mobile included
- Hero videos are MP4 or WebM, muted, with a WebP poster, and pause for reduced-motion users
- The Kit signup tags people by interest and asks them to confirm (double opt-in)
- The contact window books a Calendly call without leaving the site
- Image and background-removal licences allow commercial use
This is only the foundation
Everything above is the base layer we use for every website. Each step can go much deeper: better prompts, more advanced animation, automating the content itself. Those are the next guides in this series.