Main content
EnglishRomână
Contact
EnglishRomână

Claude Code for teams: setup and training

Claude Code set up in the company's repositories, then taught to the engineers on real tickets. Every pull request gets a second reviewer; a person still merges.

Claude Code setup for a teamLifeOS case study

Our own setup

A second model reviews the studio's pull requests

Second-model reviews, 14 repositories

Review findings back in minutes

Median time per review

A shared skill library

Skills in the setup

The workflow engineers learn

A second model reviews every draft pull request, and nothing merges without a person.

FIG. 1One change, from ticket to main
Claude Code workflow: a ticket starts a session that reads the repository's CLAUDE.md, works in its own git worktree and opens a draft pull request; CI and a second model review it; a person approves before it merges into the protected main branch.Sessions run in parallel, one git worktree each. The permission file denies access to .env files and secrets. Review findings are verified before anyone acts on them.TicketTrackerCLAUDE.mdRulesSessionClaude CodeDraft pull requestOne concernCITests, lint, buildModel reviewSecond modelPerson approvesRequiredMainProtected

Four habits behind the workflow

Draft first
Every change opens as a draft pull request. Nothing merges until the checks pass and a person approves.
Adversarial review
A second model reviews each pull request for bugs, security issues and edge cases. Its findings are verified before anyone acts on them.
Parallel sessions
Large migrations and refactors split into sessions, each in its own git worktree, so they never step on each other.
Small pull requests
Changes stay small enough to read in one sitting. A branch that grows past one concern gets split.

Reviewers, secrets and code sent to the model

Every change is still reviewed
Branch protection keeps the rule: no merge without a human approval, whoever wrote the code. More on AI-written code in production.
Secrets stay out of reach
The permission file denies reading .env files and credential stores. Tokens reach the agent through the environment, never through the prompt.
Code sent to the model
Claude Code sends the files it reads to Anthropic's API. Under the commercial terms that data is not used for training, and teams that need EU processing can run Claude Code through Amazon Bedrock or Google Vertex AI.
An audit trail in git
Every agent change arrives as a commit in a pull request, with the review attached.

Rollout stages: assess, configure, train

One repository first, on the team's calendar.

Assess
Workflow, review times and bottlenecks, measured before anything changes.
Configure
Committed to the first repository through a normal pull request.
Train
Hands-on, on the team's own tickets. Then the next repositories.

What lands in the repository

Every rule is a file, reviewed like code.

FIG. 2Files added to one repository

  • repository/

  • CLAUDE.md

    Conventions and no-go zones

  • .claude/

  • settings.json

    Allowed commands, hooks

  • commands/

    Team workflows

  • skills/

    Repeatable tasks

  • .mcp.json

    Tracker, docs, read-only data

  • .github/workflows/claude-review.yml

    Review on every pull request

Our own pull requests get a second reviewer

A second model attacks each diff, Claude fixes what blocks, and a person merges.

LifeOS case study
FIG. 3The review loop
Review loop: a draft pull request goes to a second model in a read-only sandbox; Claude verifies each finding; blocking findings are fixed and reviewed again, at most two rounds; a clean pull request goes to a person, who merges it.CleanDraft pull requestOne concernSecond modelRead-onlyVerify findingsClaudeFix what blocksTwo roundsPerson mergesSuggestions too

Claude Code for teams: common questions

Rolling out Claude Code to a team: what setup includes

Setup covers the instruction file, permission rules, hooks for formatting and tests, team commands, connections to internal tools and automated pull request review, all committed to the first repository. Training follows on the team's own tickets, and the setup then spreads repository by repository.

Measuring developer productivity with AI tools

The useful measures are the ones the team already tracks, before and after: time from ticket to merged pull request, review turnaround, bugs that reach production. Lines of code and suggestion acceptance rates reward the wrong thing. The assess step records the baseline, so both sides of the comparison use the same numbers.

Code leaving the machine, and what is retained

Claude Code sends the files it reads and the output of the commands it runs to the model API. Under Anthropic's commercial terms that data is not used for training, and retention follows the terms of the plan the company uses. Teams that need EU processing can run Claude Code through Amazon Bedrock or Google Vertex AI in an EU region. Last checked 28 September 2026.

Teams already on GitHub Copilot

No switch is needed. Copilot stays in the editor for completions, and Claude Code takes the multi-file tasks in the terminal. Instruction files, review gates and small pull requests work with Copilot, Cursor or Codex too, so the practices outlast any one tool.

Review of AI-written code

AI-written code goes through the same review as any other code, plus an automated second review. More on AI-written code in production.

One repository first, then the whole team

Training on the team's own tickets, and every rule committed as a file.

Claude Code setup for a team
Custom AI agents
Email
contact@elastoapps.com
Studio
Cluj-Napoca, Romania, EU