Engineering with AI
Working with AI without a method is a gamble, and GATE is my attempt to stop gambling.
01 Definition
AI prepares Separates the problem from the solution the request came with, and classifies the data.
A person decides Product approves the definition.
02 Building
AI prepares Writes the change and checks it against the team's rules.
A person decides Whoever builds it understands it, fixes it and opens the merge request.
03 Review
AI prepares Summarises the change, its risks and the rules that apply.
A person decides Someone else approves it, never the person who wrote it.
04 Production
AI prepares Gets the deploy ready, along with the plan to roll back.
A person decides A person decides to go to production and takes responsibility.
05 Learning
AI prepares Turns what was corrected into a rule, with its origin.
A person decides The team approves it and it applies from the next change on.
This is the first chapter of a series about GATE, the framework I wrote to bring order to working with AI. I mark it as opinion: it is what I believe after several months of getting things wrong and correcting them.
Almost every software team uses AI assistants. Developers, leaders and product people use them, and work gets done faster from the first day. But almost no team has a method. Each person works with their own prompt, in their own window, and whatever they learned in one conversation is lost when they close it.
At first that works, but then the same problems repeat. Quality depends on who used the assistant that day and how well they knew how to ask. A discussion that seemed closed comes back a few weeks later in another change. Security gets reviewed once the code is already written. And if someone asks which standard was applied or who approved a change, nobody knows what to say. I saw it in teams I have led, and my only answer was to ask people to trust us.
The risk of vibecoding
There is a way of using AI that became fashionable, and I understand why, because it is fun. You ask the assistant to build something, check that it works and move on. For a weekend prototype that is fine. For a product that handles money or people's data, it is a gamble.
The assistant writes just as confidently when it is right as when it is wrong. If nobody reviews with judgement, the mistake reaches production without anyone noticing. A team that generates code without understanding it piles up technical debt faster than ever.
In an engineering team someone has to notice when something is off and take responsibility for what gets delivered. That cannot be handed to the tool, and I see more and more teams trying.
What GATE is
GATE came out of that concern. It is an open framework that organises work with AI into stages, and in each one it is clear what goes in, what comes out and who approves it. AI prepares and people decide. The assistant reads the code and the data, drafts, and checks what it says against the source. The person accepts or rejects, authorises any write to a shared system and takes responsibility for the result.
Every decision in a code review and every root cause of an incident becomes a team rule, written down and with its origin, that applies from the next change on. That way the team stops arguing about the same thing twice. I wish I had had this years ago, when everyone who left took what they knew with them.
Each company builds its own SDLC on top of GATE, with the assistant it prefers, whether Claude, Gemini, Kiro or another. How I built it and how it is implemented are the two chapters that follow.
What changes day to day
The first change is that the team gains time. AI does the mechanical work, such as writing tickets, rebuilding context or putting together the timeline of an incident, and the team spends that time deciding. A small group can cover what used to take several roles.
Continuous improvement takes longer, but it is worth more. Every review leaves rules, every incident leaves a root cause that prevents the next one, and security is reviewed before the code is written. None of that depends on someone remembering, because the cycle does it on its own.
Knowledge stops living in the heads of two or three people, so someone new settles in faster and a resignation stops being an emergency. And for any change you can know what was asked and who approved it, which is the first thing an auditor asks.
Not everything is working yet: some parts already run and others do not.
Who decides
Governing work with AI means deciding in advance where the machine stops and where a person decides, and letting the rest move fast. GATE sets those points and separates some roles on purpose: nobody reviews their own change or approves the design they proposed, and whoever builds does not validate the business result. In a small team with a powerful assistant it is easy to get things wrong very quickly, and those separations prevent it.
An assistant can prepare a flawless deploy, but the decision to go to production is made by a person, and that person is accountable. If a company expects the tool to be accountable, the problem shows up at its first audit or its first serious incident.
What experience brings
AI multiplies what its user does. If that person lacks judgement, it multiplies confusion. If they know what to ask and what cannot fail, they achieve far more than they could a few years ago.
If there is also a method that keeps what has been learned, every month the team starts from what it learned before, without rebuilding context from scratch. Each company has to build that method itself.
What comes next
Assistants will keep improving, and fast. They will handle longer tasks with less supervision, and that makes a bad decision more expensive.
Over time, AI will be able to do more without a person approving every step. I am fine with that, as long as it is decided with evidence and someone approves it. Models become a commodity within months, so I believe companies with their knowledge in order will do better than those that only have the best model.
GATE is public and free to use. I wrote it so that other companies can use it and to put in writing what I had in my head. If you try it with your team, I would like to know where it got stuck.