A few years ago, AI entered our development process in a modest way: autocomplete. The technique was simple. Give a function a clear, descriptive name, write a short summary of what it should do, and see what the tooling suggested. The results were often surprisingly useful, and it gave us an early glimpse of what AI-assisted development might become.
Exploring the landscape
As the technology matured, so did our ambitions. We piloted a range of AI coding tools, including Amazon Q, GitHub Copilot, Kiro and Cursor, and more recently ChatGPT Codex. Each had its place, with distinct strengths and limitations.
To make sense of this fast-moving space, we formed a small, dedicated group of AI implementers who met regularly to share experiences, compare results and identify improvements. From this work, we rolled out GitHub Copilot to part of the team. It proved a valuable productivity aid and got things right most of the time, though it occasionally fell short on more complex tasks.
A turning point with Claude Code
Over the past six months, we have invested significantly more time in Claude Code, and it has marked a genuine turning point in our experience of AI-assisted development.
The key difference is context. Claude Code can see our entire codebase across multiple repositories, which has transformed how effectively we can use AI at every stage of the cycle. It has become a powerful tool for exploring ideas for new functionality, building understanding of existing code, and ultimately planning and delivering new features through agentic AI workflows.
Our development AI approach
Today, our use of AI is structured around a clear set of principles and practices.
AI tools for every developer. All developers have access to Claude Code, and our UK development team also has access to GitHub Copilot.
Standardised workflows. We use Claude Skills to define and standardise common development processes, such as writing user stories in the Retail247 style. This keeps output consistent regardless of who is doing the work.
A structured agentic process. AI-assisted development follows a defined Understand → Plan → Break Down → Deliver workflow, providing a consistent framework for moving from investigation and requirements through to implementation.
Shared context. We maintain a shared CLAUDE.md file that gives Claude common project context, conventions and guidance, helping it quickly understand how our systems and development processes work.
Human decision gates. Developers remain firmly in control. Manual decision gates ensure AI never takes significant action without developer review and approval.
IDE-based quick wins. For smaller changes, code completion and immediate in-editor productivity gains, we continue to make use of GitHub Copilot.
Looking ahead
Our journey from simple autocomplete to structured, agentic development has been one of steady experimentation and learning. By combining capable tools with clear processes and strong human oversight, we have found an approach that makes our developers more effective while keeping quality and accountability where they belong: with the team.
