Case study · Inbox Innovation, built with Claude
A freelance practice rebuilt as a product company in five days.
In November 2026 Inbox Innovation had seven years of client work and no current clients. The brief was to stop selling time: turn the practice into products that scale, with a platform clients could use themselves and a business someone could one day buy.
Douglas Baker directed the work. Claude, Anthropic’s AI model, wrote the code, tests and first drafts. This page is the record of how, and the method we now bring to client systems.
Days
5
from an empty repository to a live platform, 6 to 10 November 2026.
Automated tests
127
guarding the email system, the Health Check, the Composer and the brand tool.
Accessibility violations
0
found by axe across every page, in light and dark mode, at desktop and mobile widths.
Health Check rules
19
applied to every email the system or an AI produces.
What was built
Website
Services, pricing, sector and comparison pages, case studies and a journal, in a brand system refreshed twice on feedback.
Email design system
Three demo brands, light and dark token sets, a hand-coded component library and nine journey emails, each under 20 KB.
Health Check
Nineteen client-side checks including WCAG text contrast, with a printable report.
Composer
A governed builder with personalisation, undo, share links and export for six ESPs.
Your brand tool
One colour becomes an AA-checked brand, rendered across five emails.
Sales material
Strategy document, brand guidelines, capabilities deck, proposal template and a competitor audit.
Who did what
Douglas
- The business: what to sell, to whom, at what price
- Fourteen years of email knowledge and the portfolio
- Taste: two redesigns until it felt right
- Every decision that couldn’t be undone
- Review of every claim the site makes
Claude
- Code: pages, components, tools and tests
- First drafts of copy, decks and documents
- Accessibility and rendering checks on every change
- Research, including the competitor audit
- Fixing what the tests and checks found
What made it work
01
System first, then speed
The email components and tokens came before the tools. Once they existed, every new feature was assembled from tested parts, which is why the Composer and the brand tool took hours, not weeks.
02
Check, don’t trust
Nothing the AI wrote shipped on its say-so. Unit tests, an accessibility scan, the Health Check and continuous integration ran on every change, and caught real problems: a mobile overflow, a misleading score, a regular-expression bug.
03
Honesty is a review step
An early accessibility statement implied screen-reader testing that had not been done. It was rewritten. Every figure on this site is one we can show the working for.
04
Taste stays human
The first two visual directions were rejected. The AI is fast at producing options; deciding which one is right for a luxury buyer is still the job.
What it means for your programme
The same method works on an email programme. Build the system first, make it readable by AI, and put automated checks between every draft and your platform. Your team briefs; Claude drafts inside your components; nothing ships unscored.
The result is faster production without the drift: every AI-written email looks, renders and reads like the rest of your brand.