Making AI the everyday way of working for an engineering team

The challenge

The engineering team used AI now and then, mostly by pasting answers in from a chat window. They wanted it to be the default way of working across a TypeScript and Python monorepo of around 20 packages, with guardrails around secrets, data and production systems.

What we did

• Introduced Claude Code and paired with engineers on their real backlog, moving the team step by step from pasting AI answers, to AI writing code, to AI-drafted phased plans that the team shared, implemented and reviewed.
• Wrote layered project instructions (CLAUDE.md and AGENTS.md, at the root and per package) so the tools follow the team’s conventions.
• Built six skills for routine jobs, including opening pull requests, running affected tests, creating database migrations and hotfix releases, plus a command that turns a Jira ticket and its linked Confluence pages into a read-only implementation plan.
• Added two review agents on a cheaper model: one with no edit tools that flags destructive schema changes, breaking API changes, wildcard permissions and plaintext secrets; one that writes tests in each package’s style and flags likely bugs rather than weakening assertions.
• Added guard hooks that block edits to applied migrations, generated code and secrets files, and stop a push when the schema and generated code drift apart. Release and migration skills can only be started by a person.
• Gave engineers read-only access to production logs and traces for each environment, so they can investigate issues with no write access.

The result

Claude Code became how the team works every day. The setup lives in the repository and was delivered through normal pull-request review, so new joiners get the same tools and guardrails from day one.