Building an agency with AI, in public
Red Hills Lab serves gyms, markets, talk series and other businesses with a physical space. Every project here is a real brief, what AI did, what it didn't, and the tools I used. Take the playbook, or hire me to run it for you.
NewBuilt with AI: the LX3 case study. How I directed my first end-to-end website build. Free 14-page PDF.→Problem. LX3 makes to-order clothing in New York: tees, hoodies, sweatpants, knitwear, flannel shirts and leather outerwear. They had the raw material (product descriptions, a linesheet, photography, flat lays, illustrations, a logo and a launch film) and no store to put it in.
Goal. A simple store that puts the clothes first. Polished enough to shop, but with the brand's sketches and scrapbook feel showing through, because that's where the brand is right now.
Where AI earned its keep. Matching scattered descriptions and photos into a usable collection structure. Turning flat lays into one sketch style (after a false start). Matching ten styles to the linesheet and flagging missing prices and conflicting details on its own. Turning chat feedback into working changes, including the video overlay and the mobile galleries. Diagnosing image sizing that was leaving big blank spaces.
Results. Live since October 2026. Tracking from here: mobile conversion, product-page visits, add-to-bag rate, checkout completion, and whether deliveries land inside the two-to-three-week promise. The full write-up is in the case study tab.
Shopify, a custom theme ("LX3 — NY minimal draft") built on the Horizon setup with custom Liquid sections and templates. Codex for implementation and copy, OpenAI image generation for the sketches, Python scripts for asset and catalogue sorting, FFmpeg to inspect the launch film. No storefront apps.
The prompt that produced the most, verbatim. Not a clever one. It worked because the constraints were already established in the thread.
# after the homepage + animations were signed off Ok fire.. Build out the rest of the store with the pages we will need. Keep it simple and straight forward. Use the animations we created and keep the theme that way but when they get to the actual product it shows lifestyle images etc.
The code worth showing: the mobile product gallery is swipeable using the browser's own scrolling, no carousel library.
/* product gallery, mobile */ .lx3-product-gallery { display: flex; overflow-x: auto; scroll-snap-type: x mandatory; gap: 10px; overscroll-behavior-x: contain; }
The first tee and flannel sketches didn't match each other. It took my feedback to lock a consistent style before generating the rest.
Fixed image dimensions left large blank spaces on some layouts. AI diagnosed it once pointed at it, but didn't catch it.
The linesheet's price columns weren't labelled. I had to say which figure was retail. One product's embroidery differed between the linesheet and the photography and had to be settled by a human before launch.
A working local preview doesn't prove the Shopify version looks the same. Both needed checking, every time.
Mobile was checked at small viewport sizes, not on a physical phone or a slow connection. The 63-piece limit is messaging, not an enforced sales limit.
The launch film and the photography were supplied, not generated. Creative direction, asset approval, pricing and the operating model were all human input.
What I'd tell a client. AI can build and revise a substantial store, but it still needs your product knowledge, your creative decisions and accurate business information. A convincing preview is a milestone. Checkout, stock limits and delivery promises need verifying separately, and that's the part no model did for me.
Built with AI: the full case study. How I directed the LX3 build through its references, prompts and revisions, from the first draft to the launched store. Ends with a short prompt playbook and a working checklist for your own project. 14 pages, PDF. Read more about it →
Problem. The old site described services generically. It didn't say who the agency is for (businesses with a physical space and an experiential offer) or show the work.
Goal. A site that qualifies leads before the first call, built in days rather than a two-month agency project, and cheap enough to iterate every month.
What I'd tell a client: AI got the site to 70% in an afternoon. The last 30% (voice, what to leave out, what the offer actually is) was still the job.
The one technical piece worth showing: the prompt that produced usable copy asked for constraints, not adjectives.
# positioning brief → page copy Audience: owners of gyms, markets, event series (UK) Tone: plain, confident, no agency jargon Rule: every service line names a real deliverable and a timeframe Output: H1, 2-line intro, 6 service blocks (title + 25 words each)
Generated copy defaulted to hype ("transform", "elevate"). Needed a banned-words list.
Webflow layout still had to be built by hand. AI was useful for custom code embeds and CMS field planning, not for the visual design.
Positioning decisions can't be outsourced. The model will confidently write a site for the wrong audience if you let it.
Short notes on things I tried. Kept, dropped, or mixed. Newest first.
LX3 went from a folder of assets to a live store. The build was fast; the slow part was verifying checkout, prices and policies by hand. Next time the launch checklist goes in the brief on day one.
Ran out of Codex credits mid-project and switched. First pass of this page came from a single brief. Judging on how far the second and third rounds get.
Used on LX3 so you browse by sketch and land on the real garment. First two sketches didn't match each other; once one style was locked and fed back in, the rest came out consistent.
Codex builds a local preview fast, but a page that works locally didn't prove the Liquid version looked the same. Check both, every time.
Generated HTML doesn't map to Webflow's designer cleanly. Faster to build layout by hand and use AI for copy, CMS planning and embeds.
Placeholder entry. The trick was giving it one great past proposal as the example and a list of phrases never to use.
Every project on this site follows the same loop. It's the same loop I'd set up for you.
The brief is where the value is. The draft is cheap. The edit is where taste lives. The log is what makes the next project faster.
Two ways in, depending on whether you want it done or want to learn to do it.
For agencies and businesses with a physical space who want AI in their marketing and operations without a six-month transformation programme.
Everything on this site is written so you can copy it: the briefs, the prompts, the failures, and what I'd do differently. If something here is unclear or you want the detail behind it, ask.
Tell me what you're working on and I'll reply by email. Usually within a couple of days.
I'm Damoy Robertson, Agency Lead at Red Hills Lab, a full-service marketing agency for businesses with a physical space and an experiential offer: boxing gyms, markets, panel talk series. Brand, web, content, social, strategy.
I think the businesses that get AI into their day-to-day workflow early will pull ahead of those who wait, and that most of the advice out there is written by people who don't run a business. This site is my working notes, in public.