What practices do you use for mixed teams working in a single repository? It seems like a good idea to implement a role-based system and tie specific scopes to those roles.
retained text
We have devs, QA and a business analyst working in the same repo (the BA keeps specs there). Everyone gets the same [CLAUDE.md](http://CLAUDE.md), skills and permissions, and that's where it goes wrong. The BA asks something like "does the implementation match this spec" and Claude, reading dev-oriented project memory, starts fixing the code. QA asks for a bug repro and it wants to run a migration on staging. **What I found so far:** * \`claude --agent qa\`, where qa is an agent you define yourself in \`.claude/agents/qa.md\`. The whole session then runs as that agent, with its tools, model, permission mode, skills, hooks and MCP servers * \`claude --settings .claude/roles/qa.json\` loads an extra settings file you can commit, with \`Edit(...)\` and \`Bash(...)\` rules Together that's maybe 70% of what I need. **The parts I couldn't figure out:** * a write allow-list. Rules go deny, then ask, then allow, so "this role can only edit docs/\*\* and specs/\*\*" turns into denying every other folder, and that list gets out of date fast. dontAsk mode sort of works, but it also auto-denies every Bash command that would normally prompt * one switch. Right now it's two flags every time you start a session, there's no "I'm QA" setting you set once * role specific instructions. The agent body replaces the default system prompt, and the shared [CLAUDE.md](http://CLAUDE.md) still loads for everyone Feature request if anyone wants to weigh in: [https://github.com/anthropics/claude-code/issues/97907](https://github.com/anthropics/claude-code/issues/97907) How are other teams handling this? Separate configs per person? Hooks?
document fetched 2026-09-29 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1wsmgs4/what_practices_do_you_use_for_mixed_teams_working