AI agents in practice2 min

Four rules for working with AI agents

How I build projects with AI tools: scoping tasks clearly, demanding proof, designing first and writing every decision down.

The task: build a web app with login, database and chat – alongside other projects. The solution: AI tools with clearly divided jobs and four firm rules. The result can be checked, because every step comes with proof.

How the work is divided

An AI assistant helps with planning. With it I work out what comes next, where the risks lie and what is deliberately left alone. In the end there is an assignment: a text with the goal, the limits and the answer to the question of how you can tell that the work is finished.

A coding agent turns this assignment into code, writes tests and puts the result online.

Architecture, the sequence of tasks and acceptance on the live site are part of my role as developer and integrator.

Four rules

One task, one chat. An agent that is asked to handle three things at once easily loses the thread. That is why every task gets its own chat – and only one agent ever works on the project at a time.

Done means: with proof. A passed test, a screenshot, a link to the live page. “It's done” without proof doesn't count.

Picture first, then code. Before a new view is built, there is a draft. Only when that fits does the programming start.

Decisions are written down. Every important decision gets a number and is recorded in a file. The agent reads it before it starts. That way no conversation begins from zero.

What goes wrong

Agents are fast and sound confident – even when they are wrong. Without the duty to provide proof, I would get answers that sound good and are not correct.

The second trap: with agents many things go quickly, so the plan grows faster than the result. A short list helps against that: what is on now – and what has to wait.

Why the log matters

I don't rely on “should be fine”. There are checklists, handovers and a log that records who checked what and when. The more independently an agent works, the more that log matters.

Back to the blog