Here is the part nobody expects. I never wrote a single image prompt. I brought taste, a screenshot from Pinterest, and plain English. Claude, wired to Higgsfield (an image generation tool it can drive directly), did the style extraction, the generation, the recolour, and the wiring into the site.
Here are the three prompts I actually typed, and the five-step setup around them.
The setup, in five steps
- Find a look you love on Pinterest and screenshot it.
- Connect Higgsfield to Claude. This is a one-time MCP setup.
- Open Claude inside your project so it can see your code and your brand colours.
- Paste the screenshot.
- Drop in the magic prompt below.
Prompt 1: Pinterest reference to on-brand hero
Paste the Pinterest screenshot along with it.
Here's a reference I love from Pinterest [paste the screenshot]. Study the style. It's a dense neo-Renaissance line engraving, a single ink colour on a plain background, with the headline sitting in an empty clearing in the middle. I want our homepage hero to feel like this, but in our brand, not theirs. Here's how: 1. Pull our exact brand colours straight from the site code (the red and the cream or paper). 2. Using Higgsfield, generate the engraving as black ink on a plain white background. That's the trick: image models can't hit an exact brand hex, but they're brilliant at black and white, so we recolour after. Make it dense and symmetrical, framing a clean empty centre for our headline. Use motifs that mean something for us: wisdom, craft, and a bit of the cosmos, an AI brain for your business. 3. Then recolour it: map the black to our brand red, the white to our cream. 4. Wire it into the hero behind our existing headline, add a soft scrim so the text stays readable, and show me two or three options.
The magic in it:
- Point at the reference and name the style.
- Tell it to use your colours from the code.
- Black on white first, then recolour.
- Leave the centre empty, with no text baked into the image.
- Ask it to wire it in and show you options.
Prompt 2: one hero to a whole visual system
Love it. Now look at the reference again. It also has little engraved objects scattered through the lower sections. Those suit us far better than the stock photos we're using. Using Higgsfield, generate a matching set of spot illustrations in the exact same hand as the hero: one specimen sheet of about nine isolated objects on white (owl, orrery, satellite, gears, planet, ledger and quill, scales, stag, laurel). Slice each one out, knock the background to transparent, recolour to our red, and use them to replace every photo on the site, one small motif per section. I want the whole page to feel carved by the same hand, not an engraving hero plus stock photos.
The magic in it:
- Ask for a specimen sheet of isolated objects to get a consistent set in one shot.
- Reuse the same recolour trick.
- Replace the photos, and make the whole page one visual language.
Prompt 3: art-direct the polish in plain English
Now let's refine it like a designer would. Walk the live site with me and fix what's off. I'll point, you adjust and show me a screenshot, one change at a time: - the hero artwork has white gaps at the sides, scale it up so it bleeds edge to edge - the trust cards feel cramped, give them more breathing room and smaller icons - try the top nav as a solid red bar with white text Keep what I like, revert what I don't.
The magic in it: you don't describe how, you describe what's wrong. "Cramped", "white gaps", "try it red", and Claude translates your taste into code.
The one-line lesson
You're not the designer and you're not the prompt engineer. You're the art director. Pinterest gives you the taste, Claude and the image tool are the hands. You just keep saying "more like this".