Building software with an AI agent: what I hand over to Claude Code and what I never do

Over the last two years I built two production applications end to end with Claude Code, that is, together with an AI coding agent. One is an apartment allocation (lottery) platform for a housing cooperative with 639 members, the other is a financial audit system that processes about 15,000 bank transactions. In this article I explain what the agent actually does, what I do, and how quality is protected. When choosing a software team or developer, this division of labor matters more than the answer to "Do you use AI?"

The work I hand to the agent

The agent is very fast at well-defined tasks with clear boundaries: a new screen or component, an API endpoint, a SQL migration, refactoring a specific module, searching and analyzing existing code, updating documentation. For these tasks, the agent writes a significant part of the code I would otherwise write. The time saved goes to the decisions that really matter.

The work I never hand to the agent

• Architecture and data model. Which tables exist, how the permission model is built, which work is done on the server: these are my decisions.
• Correctness of business rules. In work where a mistake is expensive, such as a lottery algorithm or financial matching, I write the rules and the agent implements them.
• What goes live. I am responsible for every line that goes to production.

How is quality protected?

I treat the agent's output like a teammate's pull request. In practice this means:

• Written project rules. Every project has a rules file that the agent reads (CLAUDE.md) and per-module READMEs. Coding standards, folder structure, prohibitions and workflow are written down there. The agent reads them first in every session, so the output stays consistent.
• Automated gates. Code that fails type checking (TypeScript) and lint checks cannot be committed, and the same checks run again in the deployment pipeline.
• Browser verification and code review. I try every change in the browser, including "small" ones, and read it line by line.
• Splitting the task correctly. Work divided into pieces of a size the agent can handle reliably gives faster and less error-prone results. When the agent goes off track, I stop and redirect it, or write that part by hand.

What changes for the client?

When set up correctly, AI-assisted development delivers the same quality in less time. When set up wrongly, it leaves behind a codebase that was written fast but that nobody fully understands. Before working with a developer, I suggest asking these questions: Who checks the code the agent writes, and how? Are the project rules written down? When the project is finished, can another developer take over this code?

My answer: code, rules and documentation are delivered together. The agent brings speed, and the responsibility stays with me.

A concrete example: the division of labor on the lottery platform

The most sensitive part of the cooperative's apartment allocation platform was the lottery algorithm. A wrong draw means both a legal and a social problem. So I designed the algorithm's rules, how the rounds would work and how a draw would be undone. The AI agent implemented this design: it wrote the database functions, the server-side functions and the admin screens. I made the decision that the real draw and the trial (dry-run) simulation share the same core code. That way, a difference between the simulation result shown on screen and the real draw became structurally impossible.

In the financial audit system, I used AI to match bank transactions with members, but no match enters the data without human approval. Amounts are calculated with integer arithmetic, and a match that does not pass a series of checks is rejected.

When do I write by hand?

Not every job suits AI. When it is a change of a few lines that is very critical, finding the root cause of a complex bug, or a problem where the agent has not found the right approach in two or three attempts, writing the code myself is faster. The real skill in AI-assisted software development is choosing correctly who to give the job to.

In short, an AI agent, when managed well, works like a very fast developer who has joined the team. But like any developer, it needs clear instructions, written rules and careful review. If you are considering AI-assisted development for your project, ask whether these three things are really in place.