Claude Code in a team codebase: what to agree on first

One developer using Claude Code produces inconsistencies in their own work. Multiple developers using Claude Code in the same codebase compounds this — different naming patterns, different error handling approaches, different testing styles, all generated at speed.

The fix is a shared CLAUDE.md, but what goes in it matters. Here's what's worth standardizing.

The shared CLAUDE.md

Put CLAUDE.md in the repo root and commit it. This is the source of truth for how Claude should behave in this codebase. Every developer's Claude sessions read it, so the defaults are consistent.

What to put in it: the conventions that people get wrong when they don't know the codebase. Naming patterns, error handling approach, which libraries to use for what, what not to do. Don't document things that are already obvious — document the things that require knowing the history or the decisions.

A task template in the team standards

When different developers give Claude different quality prompts, the output varies. Agree on a task template: minimum required fields (scope, done condition, do-not list) and the pre-task read instruction. This levels up the lower-quality prompts without requiring everyone to become expert prompters.

Review standards for AI-generated code

AI-generated code needs different review emphasis than human-written code. The typical human reviewer spends time on logic and architecture. For AI-generated code, the priorities shift: file scope (are unexpected files changed?), test quality (are these tests actually testing behavior?), and assumption verification (did Claude assume something about the codebase that isn't true?).

Add this to your PR template or code review guidelines. Otherwise reviewers treat AI-generated code like human-written code and miss the specific ways it tends to go wrong.

What to do about style drift

Over time, different Claude sessions will make different style choices in different parts of the codebase. This is inevitable. The mitigation: update CLAUDE.md when you notice a pattern drift. "We're seeing three different async patterns — standardize on this one, add it to CLAUDE.md."

CLAUDE.md maintenance becomes part of the engineering workflow, not a one-time setup.