Digital publishing · AI-accelerated delivery
ICF Website Rebuild
A legacy CMS site rebuilt on a modern headless stack, with AI agents trained on the team's own conventions doing the repetitive work.
50–60%less development effort
The challenge
A legacy CMS caused slow page loads, publishing friction and weak mobile engagement.
The rebuild had to ship a large surface of pages, templates and content blocks with a four-developer team, without letting speed erode consistency, accessibility or review quality.
Our Founder's role
- Led delivery and architecture alongside the client's technical director, and designed how AI agents were used across the team's development lifecycle.
How it works
- 01
Headless re-architecture
Rebuilt on Next.js with a headless CMS organised as pages, templates and blocks, fetched through GraphQL, styled with Tailwind CSS and deployed on Vercel.
- 02
Automated scaffolding from the CMS schema
A meta file describes the content structure, and an automated task pulls the CMS schema and generates typed templates. Developers build components directly against those blocks, and nobody hand-maps the CMS again.
- 03
AI agents across the lifecycle
Cursor plan mode produced first drafts of component design, with final architecture decisions kept by humans. Agents generated components, page templates, integration code and unit tests, and ran a first-pass self-review before every PR. Human review stayed mandatory.
- 04
Preview-first QA
Every PR deployed automatically to a Vercel preview, where developers and QA tested together before merging.
Engineering decisions
Encode the team's standards as agent rules
- Next.js conventions, Tailwind standards, accessibility practices and component patterns were captured as reusable Cursor skills and rules. The agent writes code the way the team writes it, so output from four developers stays consistent.
- A feedback loop closes the gap: every new kind of review comment from the technical director became a rule, so the same comment never came back.
Measure the gain honestly
- The 50–60% is scoped to development effort, measured against story-point history for similar work. Lead time also depends on requirements, designs and product feedback, which AI doesn't speed up.
Outcomes
- Same scope delivered in roughly half the development time, often ahead of estimate.
- Shorter review cycles, with style, structure and accessibility basics no longer debated in review.
- Fewer post-merge bug-fix PRs, because basic bugs were caught and fixed in the same PR on preview.
- Sub-second page loads and zero layout shift across editorial articles.
Technology
- Next.js
- Frontend
- Sanity CMS · GraphQL
- Headless content and queries
- Tailwind CSS
- Styling
- Vercel
- Hosting and per-PR previews
- Cursor agents
- Design drafts, code, tests, self-review
Contact
Have something you want to build?
Tell us what you're building, or what you'd like to automate. Let's figure out the next step.