Free CLAUDE.md review

Claude Code · Codex · Cursor · for development teams

Help your team adopt Claude Code and coding agents.

VAI Soft Lab is a small software studio in Sydney, Australia, run by Vahid Alamipour, a lead engineer with 16+ years of building software, from junior developer to CTO. Vahid has led 30+ developers into AI engineering and works with coding agents every day. We help development teams use Claude Code, OpenAI Codex and Cursor well: one rules file, safe permissions, tests the agent can run, and a review flow that keeps up. Our main stack is TypeScript, React, React Native, Next.js, Node.js, PostgreSQL and AWS. How we work: we look at how your team works today, agree the scope, then set it up in your repo with your developers.

Teams in Australia and worldwide. English. Sydney time zone.

Problems teams hit

Everyone has the tool. Few get the results.

Buying seats is easy. The hard part is how the team uses the agents: what they may touch, what they must check, and who reads the result.

  • Everyone uses it differently.No shared rules file, so each agent guesses the naming, the libraries and the error style. We write one short CLAUDE.md or AGENTS.md that every agent and every tool reads.
  • Too much power, too few limits.Agents run with production keys, broad permissions or auto mode, and nobody knows what they may do. We set permissions, block the commands that must never run, and keep secrets out of reach.
  • Review is the bottleneck.Pull requests get bigger and wait longer. We cut the work into small tasks, give the agent tests it can run, and agree a short review checklist.
  • Agents clash with each other.Two agents in one folder write over each other's work. We set up git worktrees, clear task borders and one merge at a time.
  • The speed boost fades.Studies of open-source projects found the first speed-up fades while warnings and complexity stay higher. We add the checks that keep quality up: tests, linters and CI the agent must pass.
  • Usage limits and surprise costs.Sessions stop in the middle of a task, or the bill jumps. We help you pick plans by real use, and set rules for big jobs and parallel agents.

How it works

Four steps.

You decide after each step. Nothing starts before you agree.

  1. Free CLAUDE.md reviewSend your CLAUDE.md or AGENTS.md and five lines about your project. We send back the three changes we would make, free, within one working day. We do not need your code for this.
  2. Agree the scopeIf you want help, we look at how your team builds today: tools, repo, tests, reviews. We agree together what to set up first and how long it takes. Then you get the time and the price in writing.
  3. Set it up with your teamIn your repo, on a branch, under your access rules: the rules file, permissions and hooks, tests the agent can run, worktrees for parallel work, and a review checklist. Then we work through real tickets with your developers.
  4. Your team runs itYou keep everything in your repo, plus a short written guide: how to start a task, what to check, what never to let an agent do. When it runs on its own, we leave.
What we set up
  • One short rules file for all agents and tools.
  • Permissions, blocked commands and hooks.
  • Tests and checks the agent runs before it says done.
  • Worktrees and task borders for parallel agents.
  • A review checklist and a report format.
What we never do
  • Give an agent production keys or customer data.
  • Push straight to main, or merge without a person reading the diff.
  • Put your code into a consumer AI tool.

Questions about coding agents for teams

Questions.

Do we have to use Claude Code, or can you help with Codex and Cursor too?

All three. Many teams use two of them. Claude Code, Codex and Cursor all read AGENTS.md now, so one rules file and the same tests can serve the whole team. More in our article on which coding agent to use for which job.

Is our code safe with a coding agent?

Use a business plan or commercial terms: Anthropic does not train on Claude Code use under commercial terms, ChatGPT Business does not train on your data by default, and Cursor does not train when Privacy Mode is on. Then limit what the agent may do, and keep keys and customer data out of its reach.

Do you need access to our code?

Not for the free review: we only need your CLAUDE.md or AGENTS.md and five lines about the project. For the setup, yes, with the access you choose, on a branch, and you can remove it at any time.

How long does it take?

It depends on the team and the codebase. The free review takes one working day. A setup for one team is planned with you and written down before we start, with the time and the price.

Our developers already use AI chat or autocomplete. What changes with agents?

An agent does whole tasks: it reads the code, edits many files, runs commands and checks its own work. That is why it needs rules, limits, tests and review that chat never needed.

How many agents should each developer run?

Start with one, then two. Add a third only when your review keeps up. Our article on running several coding agents at once has the details.

Where are you, and do you work with teams outside Australia?

We are in Sydney, Australia. We work with teams in Australia and around the world, in English, by email and video call.

Next step

Get a free CLAUDE.md review.

Send your CLAUDE.md or AGENTS.md and five lines about your project to hello@vaisoftlab.com. I send back the three changes I would make. Free, within one working day.

Five lines: what the project is, the stack, team size, which agents you use, and what goes wrong most.

Who does the work
  • Vahid Alamipour: 16+ years building software, from junior developer to CTO, then lead engineer roles in Malaysia and Australia.
  • Has led 30+ developers into AI engineering. Over a year of daily work with coding agents.
  • Four Anthropic certificates in building with Claude.
  • Runs his own agent team every day in VShell, the Mac app he built for Claude Code, Codex and Cursor.
  • Other ways to work with us: all services.