Behind the scenes
Behind the scenes. How Dewgo builds a site in 2 weeks
Two-week web design works because we did the boring infrastructure work once, on our own time, instead of charging clients for it on every project. Here is exactly what runs behind the scenes from intake to launch, and where the human judgment lives.
7 min read · published May 21, 2026
The intake form
Five steps. Industry, tier, ten base questions, five industry-specific questions, contact info. Direct file uploads on the logo and photos questions. The whole thing runs as a multi-step React form on /start/web-design.
Submission goes through a Zod schema, then into a Supabase RPC that writes to dewgo.intake_submissions. Two transactional emails fire via Resend: one to info@dewgo.org with the full intake, one to the prospect confirming we got it.
The seven starter templates
We maintain seven Next.js 15 starter repos under our GitHub org. One per industry: local services, restaurants, health, professional services, real estate, creators, portfolio. Each is a working standalone Next.js app with industry-specific sections (services grid, menu cards, practitioner bio, project grid, etc.) and a dewgo.config.json file that holds all the placeholders.
The templates were designed once. Every project starts from one of them. That is most of the time savings.
The AI first-draft pipeline
Triggered from the admin console once we accept the intake. A Vercel serverless route runs nine steps:
- 1. Load the intake row from Supabase.
- 2. Pick the matching template based on the industry field.
- 3. Generate dewgo.config.json values with Claude Haiku 4.5 using a strict per-industry JSON schema.
- 4. Fork the template repo into the org with a unique name.
- 5. Wait for GitHub to finish the fork.
- 6. PUT the generated config file into the new repo.
- 7. Create a Vercel project pointed at the new repo.
- 8. Trigger the first deployment, capture the preview URL.
- 9. Create a Notion runbook page with intake ID, repo URL, preview URL.
What is automated
Repository creation. Initial copy generation. Vercel project setup. The first deploy. Sitemap. Open Graph image generation. Per-page meta tags. Image cropping suggestions. The Notion runbook.
These are the high-volume, low-judgment tasks. A robot that gets them right 95% of the time saves us about three days per project.
What is human-touched
Headline rewrites. Brand voice. Photography selection and direction. Section ordering. Custom layout decisions. Hover states and interactive details. The choice of when to break the design system on purpose.
These are the decisions that make a site feel intentional. We hold them for human time. The pipeline buys us the hours to spend there.
The review and revision loop
You get a Loom walkthrough plus the preview URL on day 10. We use the Loom because written feedback on a live URL plus a visual walkthrough is the fastest way to align. You leave comments inline or send a written list. We have one revision pass scheduled for days 11 and 12.
The launch and handoff
Domain pointing. SSL verification. Performance check. Lighthouse 95 plus on every page. Final QA. Launch. Then we send the runbook: GitHub repo, Vercel admin, env var docs, the answers to the predictable questions about the CMS, the contact form, and how to update copy.
Why this is honest about AI
AI does not write your headline. AI does not pick the photo of you. AI does not decide what to remove. AI does the boring parts so the human time goes to the parts that matter. We tell clients exactly which step is which. The pricing is based on the time it actually takes us, not the time it would take without the pipeline.