gray-jan tupas

BUILD 09 · Case study

Running a client website build with an AI coding agent

In progress: staging, not yet live

A full website for a Singapore courier company. This page is about how I ran the build: scoping from meetings, holding back features that weren't ready, checkpoints, and a decision log.

Client
Singapore courier company
My role
Design and build, solo
Coding agent
Claude Code
Output
Multi-page static site

Problem

The requirements for this site came out of client meetings, not a written spec. An AI coding agent will build whatever it is told, including features nobody has decided on yet. I needed a way to turn each meeting into a small, checked piece of work.

My role

I designed and built the full site solo, using Claude Code as my coding agent. I wrote every prompt, reviewed every build, and decided what was ready to go in.

How it works

Every change to the site went through the same five steps.

  1. Client meetingThe client says what they want and reacts to the current version.
  2. TranscriptThe meeting transcript is the source for the next piece of work.
  3. Scoped Claude Code promptOne prompt per piece of work, with checkpoints written into it.
  4. BuildClaude Code builds up to the next checkpoint, then stops.
  5. ReviewI check the result before the next step starts.

Two files carried the project between sessions. A CLAUDE.md file held the build rules, and a decision log kept the history.

This is what that process has produced so far, on staging.

The 'Choose your delivery service' section: five service cards for Same-Day, Express, Night & Weekend, Van and Direct, above a table that matches what the customer needs to the right service.
Service cards and the service-picker table
The 'From request to doorstep' section: four numbered steps. Tell us what and when, we confirm price and pickup, a rider collects, delivered and confirmed.
The 4-step "how it works"

Key decisions

  • Scope from the meeting. Each prompt was written from a meeting transcript, so the work matched what the client actually asked for.
  • Hold back what isn't ready. Features still in development stay listed as in development. They are not presented as finished.
  • Checkpoints in every prompt. The agent stops at set points so I can review before it continues.
  • Rules and history live in files. CLAUDE.md holds the build rules. The decision log records what was decided and keeps the history.

What's done vs in progress

Built so far

  • Home, quote & booking, corporate accounts, 5 service pages, contact, privacy/terms/cookies.
  • Service-picker table matching needs to services.
  • Client logo wall.
  • Real Google reviews.
  • 4-step "how it works".
  • Story timeline.
  • FAQ.
  • WhatsApp chat link.
  • Sticky mobile Call / Schedule a Pickup bar.
  • Skippable intro.

In development

  • Booking calendar with a server-enforced weekly cap (Monday–Sunday, Singapore time).
  • Tracking link in order emails.
  • Reminder email.
  • "Track a delivery" link.

Limitations

  • The site is on staging. It is not live yet.
  • The booking calendar, tracking link, reminder email and "Track a delivery" link are still in development.
  • There are no results to report yet, because the site has not launched.
  • The client is not named here, so the screenshots leave out the site's header, logo and client logo wall.

Links