All notes

Current chapter · Six-part field guide

Part 1 of 6 Building an AI-first development system

Why I moved to AI-first development during a migration

During a cloud front-end migration, I chose to move from AI-assisted to AI-first development. Here's what I wanted that change to mean in practice.

5 min readAI EngineeringEngineering SystemsProduct Development

I wanted AI to take part in understanding and checking the work, as well as writing the code.

The decision I made during the migration

I was already in the middle of the cloud front-end migration when I decided to change how I was approaching the work. I wanted to move from AI-assisted development to AI-first development. The migration was still the job in front of me, but I also wanted to change the process I was using to do it.

AI-assisted development can fit quite comfortably into an existing way of working. I can ask for a component, use AI to understand a piece of code, or get help with a test. That's useful, but I may still be carrying the requirements, explaining the same context, deciding each next step and working out how to check the result. Getting more code written doesn't automatically change those parts of the job.

Choosing a different approach midway through the migration meant thinking about both things together. I still needed to move the product forward, and I wanted the development process to support that work more fully. I couldn't assume that a better way of generating code would, on its own, make the rest of the work easier to understand or check.

What I wanted AI to take part in

I wanted AI to take part in understanding and checking the work, as well as writing the code. That's what I mean by AI-first here. I wanted to think about where AI could usefully participate throughout the development process, starting before an implementation request and continuing through the checks needed to review the result.

Before changing something, I need to understand what it's supposed to do and which existing behaviour still matters. Where does that requirement come from? Which source describes the current behaviour? What am I assuming, and how will I check the change? I want those questions to be part of the work AI helps me prepare, with the answers available for implementation and review.

Once that context is clear, AI can help plan the change, carry out a defined part of it and help check the result. I want the useful context to carry through those steps. If a requirement changes or a check fails, the next step needs to take that into account. Otherwise, I'm still doing the work of keeping the process connected, even if individual tasks happen faster.

The migration still had to preserve behaviour

A migration gives this a practical context. The existing product is part of the evidence for what the new version needs to do. I may want to improve the implementation, but I still need to understand which behaviour people depend on. A cleaner component or a convincing screen doesn't answer that question by itself.

I also need to know which source I'm using for which decision. A design can help explain how something should look, while the existing application and its API behaviour may tell me something else I need to preserve. If those sources disagree, I need to resolve the difference or make it explicit. Asking AI to continue doesn't make an unclear requirement any clearer.

So the move towards AI-first development couldn't be a reason to loosen the migration work. It needed to help me make that work more explicit: the behaviour to preserve, the change being proposed, and the evidence I'd use to judge it. Those are things I need to be able to inspect together when reviewing a change.

What I still need to decide

AI-first doesn't mean I can leave every decision to an agent. I still need to decide what problem I'm solving, which constraints matter and when a proposed change is acceptable. AI can help me examine a choice, but I need to be clear about the decision I've actually made and the reasons behind it.

If a requirement is unclear, I want the uncertainty to stay visible until it's resolved. Sometimes the next useful step is a question, a check against the existing product, or a conversation about the intended behaviour. I don't want a plausible implementation to quietly become the answer to a product question that nobody has settled.

At the same time, I don't want to supervise every mechanical step. If I have to keep moving context between tools, repeat the same instructions and reconstruct what happened before I can review it, there's still work to do on the process. I want my attention to go towards the decisions and limitations that need it, with enough evidence to understand the result.

How I'd judge whether it's working

I wouldn't judge this only by how quickly code appears. I need to understand what changed, why it was changed, what was checked and what remains uncertain. Those answers help me decide whether to accept the work, and they should still be useful when I come back to change it again.

There's a cost to making the process more explicit. Preparing context and deciding what to verify takes time, so I don't think every small change needs the same amount of structure. A straightforward copy edit and a migration change affecting existing behaviour need different levels of checking. The process has to leave room for that judgement too.

That was the direction I chose in the middle of the migration. I wanted a development process in which AI could help carry the work forward, while I could still understand and take responsibility for the decisions. The name matters less to me than whether I can rely on that process when the next change comes along. That's what I wanted to work towards while continuing the migration.

Save thearnab

Your labs, one tap away. Add the app to your home screen to open it straight into the Lab.

In Chrome or Edge, use the install icon in the address bar or the browser’s app menu.

In Safari on a Mac, choose File → Add to Dock. On a phone, open this site in Safari or Chrome and save it to your home screen.

For music without a connection, use Prepare offline in the Songbook while online.