AI-Assisted Feature Development for iOS

Welcome back. We now have a clear product direction for AI usage insights. Next, we’ll look at the experience from a design perspective, check how the feature fits the app technically, and then bring everything together in our feature brief. Let’s jump in.

Notice how it’s also giving us the token estimates for the prompt and the responses.

Next, let’s switch roles, and let’s take a look at this as a designer. The product review helped us decide what the feature is for. The design review should help us decide where it belongs, how it should feel, and how to keep the UI from becoming a giant wall of management text. And yes, giant wall of management text is a real design smell. Very technical term. Probably in Apple’s Human Interface Guidelines somewhere. Maybe, maybe not. Don’t quote me on it, though.

Anyways, I’ll open my prompts markdown file, and I’ll copy prompt two: Prompt 2: Designer Review for the designer review. Now I’ll switch back to ChatGPT and ask it to continue.

Excellent. Let’s see what came back. So we have a recommended user flow where it says, “A lead opens the existing narrative summary screen. We have a short AI usage insights section. Each row represents or presents one aggregate pattern. Selecting a row takes us to the most relevant existing detail chart, and the chart screen provides the underlying trend and comparison context.” Now, this is where I want to be very picky. AI tools love suggesting more UI, more screens, more cards, more badges, more everything. Sometimes in a useful way, but sometimes not really. And for this app, smaller and leaner is better.

The current app already uses a summary row that takes you to a detail screen. So the most natural design is one new summary row that opens one new AI usage insights screen. No tab bar, no modals, no complicated navigation. Just a clean new destination. What do you think? In this case, did ChatGPT give us what we’re looking for? I think it does. It mentions that we will have an AI usage insights section that appears amongst the existing insights style rows. But notice how it doesn’t say an AI usage insights row that appears among the existing rows. So just to be clear, I’ll have it adjust that for me.

Here I’ve told ChatGPT that I don’t really want to add any sections or tab bars or any different UI patterns than what it already has. We have a clear set of rows that navigate into details, and I want to keep that pattern. And as it returns, it says, “Agreed. The feature should just be one additional summary row, not a new section.” And it’s important that you also have your AI tool check out your code so that it can find existing patterns when creating a plan. That way, it can do a better plan for you and lead to more efficient results.

Now, and for the final prompt before we write all of this to a file, let’s review this feature but as an engineering manager. At this point, we’re still not asking for code implementation, any Swift files, or anything in Xcode. We are asking, what’s the smallest reasonable shape this feature could take in this app? That distinction is important. If you ask for code too early, the AI may start solving details before the boundaries are clear. Here, we want architecture pressure, not implementation.

I’ll go ahead and open my prompts file and copy prompt three: Prompt 3: Engineering Manager Review. Then I’ll paste it into ChatGPT. Notice how in prompt three, we have certain relevant information like the app architecture or any known decisions that have been captured in our Swift files or types. ChatGPT uses that and is able to provide us a recommended implementation shape, files to potentially add, modify, and a data flow. It also brings up some edge cases and acceptance criteria, which is excellent for us, as well as some open questions for human review.

Now, at this point, the review should start making the feature feel more concrete. We’re looking for things like models, navigation, view models, flows, services, views, etc. But we’re still not writing the Swift files yet. If the AI suggests adding networking, persistence, a backend, or a runtime AI call, that’s a good moment to push back and refine things to your liking. Now, while those may be interesting ideas for a real app later, for version one, they don’t really belong in the plan.

For the purposes of moving things along and to demonstrate things in this lesson, I’m just going to have ChatGPT come up with some sensible defaults and recommendations for the open questions for human review. In your case, if you have anything that the AI is telling you that you should review, take the time to do so, have a back and forth with it, and polish things to your liking.

Now, while it’s thinking about that and coming up with some recommendations and defaults, I’m going to go ahead and copy the final prompt, which is going to take all the work that we’ve done and put it into a feature brief markdown file with some rules and a specific structure so that we can standardize our output. So I’ll go ahead and copy this prompt, return to ChatGPT, and we can see that here are its recommended sensible defaults for our review and approval for version one. So I will just go ahead and say, “Perfect. Your recommendations are approved.” And I’m pasting the final prompt and letting ChatGPT work its magic for us.

Perfect. Let’s just make sure to see if it wrote it out to a file. It hasn’t, so I’ll just say, “Write it out to a file in our module three folder.” Excellent. We now have our feature brief markdown file. So let’s go ahead and take a look at it in VS Code.

And there we go. We have our AI usage insights feature brief with the name, the rough idea, the audience, the problem statement, version one scope, et cetera. Now, this gives us our feature brief, and it’s not huge, but it’s much better than just having a rough original idea that you write out in a sentence or two or a paragraph or two without any structure or rhyme or reason. We now know who the feature is for, what it should do, what it should avoid, where it belongs in the app, and what questions still need some human judgment and approval or review and refinement. And that’s the whole point of this lesson. It’s to show you how AI has helped us think from different angles. But it did not get the final vote. We did, and we guided it towards our goal.

So let’s quickly review what we have ended up with. Now, we started with the rough idea of, hey, I want some AI usage insights to the tokens dashboard app that I’m working with so that a team lead can understand really how the team is using AI and to find any usage patterns. From that rough idea, using some of the lengthier, more descriptive and verbose prompts, we have a feature brief with things like the audience, the scope, goals, non-goals, any open questions, et cetera. And that’s a much better starting point for development planning. And just as important, we have already avoided a few traps. We are not building a performance review tool. We’re not adding a runtime AI call. We’re not assuming model pricing, and we are not claiming that AI caused delivery improvements. So not bad for one planning pass.

Now, in this lesson, you used AI to review a feature idea from product, design, and engineering perspectives. The important part was not getting a fancy AI answer. The important part was using those answers to make better human decisions. You now have a feature brief markdown file that describes the AI usage insights feature clearly enough to plan from.

In the next lesson, we will take this brief and turn it into an iOS development plan. That is where we will define the proposed Swift types, data flow, edge cases, and acceptance criteria. So far, still no code, but we’re being very methodical, organized, and tidy. So with that, I’ll see you in the next lesson. Bye for now.