From AI-assisted to AI-first, in a year

From AI-assisted to AI-first, in a year

Over the past year I've led a radical change in how an engineering department works, from first introducing AI-assisted development to a point where the vast majority of code is generated by AI and reviewed by AI alongside human reviewers. It's checked against style and coding guidelines iterated on in the repo, so what's produced meets the team's standards rather than the model's defaults.

• We started with AI-assisted coding (GitHub Copilot and Amazon Q) for routine tasks, before moving to AI-first development.
• Most code is now written by AI and checked by both AI and people before it merges.
• Standards live in the repo, so the models and the team work to the same rules.
• Claude rates every pull request green, amber or red, so reviewers know where to look before they open a diff.
• Terraform plans get an AI summary as a GitHub check before anyone approves them.
• Chores like cost audits, dataset restores and stacked PRs are one command for any engineer or agent, through 20+ Claude Code skills.