Why Software Engineering Is Becoming: Plan and Review

Hi there. I'm Louis and I'm one of the co-founders of Vibe Kanban. And we make tools that help software engineers work with coding agents. And that means that we spend a lot of time thinking about how to effectively work with coding agents, be productive, get more done. And there's a lot that can be accomplished. And so if you can figure out ways to optimize your workflow, you can get a lot of advantage in a world where some engineers who figure out how to use AI are going to be getting two, three, four times as much shipped as those that don't.

Today, I'm going to focus on two aspects of this. One is planning and one is review, and how you can optimize your workflow around these two human elements of the process that are becoming more and more prevalent. So, one of the things that I think about a lot is that as coding agents' capabilities have become stronger, they are running for longer. They're taking more time to complete their work. If we look at the types of jobs that you might do on a day to day as a software engineer, you can bucket them into planning, writing code, and reviewing code, whether that's reviewing what you do or reviewing what your teammates do.

And if we tried to put a percentage of what amount of time we're spending doing those different jobs, I think before AI significantly started to improve people's productivity, it might have looked a bit like this, a significant amount of writing code, a decent amount of reviewing both your own and other people's code, and a little bit of planning as well. And as things start to get accelerated through AI, so first with GitHub Copilot, then through ChatGPT, later with Cursor, and of course, most recently, again with Cloud Code, you start to see that actually the writing code part has flipped from being the thing that I spend most of my day doing to the thing that I'm spending almost no time doing.

And so, I guess the question is, what do we do with all that extra time that we just got back? Well, one of the things you could do is get on TikTok, or play Candy Crush, or, you know, my vice, I don't mind admitting to, is doom-scrolling Twitter. And there are even IDs now that will help you procrastinate the hardest if that's what you want to do. However, I think there are actually plenty of things that you can do. And one of the things to think about is that actually, even though writing code is not really something that's even part of my workflow anymore, there are other areas of the workflow that seem to be taking up more time than before.

So, if before I was spending 10 minutes writing code, and thanks to AI assistance I'm now spending one minute writing code, I didn't just get nine minutes back. In fact, I think it probably added at least two minutes to the amount of time I had to spend planning, and at least two minutes to the amount of time I have to spend reviewing code. And so, it's a mixture of being both an accelerant and also displacing work to other areas of the workflow, creating work in those areas. And so, if that's true, then basically, if you want to ship more, you need to find ways to plan and review more quickly.

If you're sitting there thinking, I really want to, you know, ship something, and I'm just looking for the lever to pull that's going to help me get that out there even faster, essentially, if you can figure out how to plan more faster, review more faster, you can figure out how to ship more. That is the biggest lever you now have. So, I'm going to dive now into some of the tips that I have for getting more done or making your life, you know, more fun in a world where most of the job of software engineering is becoming planning and review.

And the first of those tips is, if you had to choose, and it is a choice, you should be choosing to spend more time planning than more time reviewing. And I actually think that if you spend more time planning, it will save time that you otherwise would have to spend reviewing. So, one of the things that I mean by planning is you need a system of record. So, with coding agents, everybody kind of has to be a bit of a manager. You need to find ways to structure your daily work so that you can easily see, okay, well, what tasks do I have in progress? What tasks are in review? And what tasks are there to do? And this is because you're going to be getting so much more done than you were before that it can actually become difficult to really keep track of everything that needs to be done when it's all happening in your brain. So, it doesn't have to be Vibe Kanban, the product we make. It can be Notion, it can be Google Docs, or whatever. But you need to be organized about your own personal productivity in a way that I think only teams really had to care about this before AI coding agents arrived on the scene.

My next tip is about actually the act of planning with an AI coding agent. So, you need to basically ask AI before it makes a change to take your description, and your description might be quite small, maybe contain references to what the problem is, but not the full detail of how to fix the task. And what we're going to do is we're going to work with the coding agent to turn that into a much more comprehensive plan. So, you can see here in the red box, I've just got a simple prompt that says, please plan how to execute this task. And you can use the flawed code, or the codex plan modes to achieve this as well. You don't have to just prompt it to create a plan. And what you'll get back from your coding agent is the outline of the changes that it's going to make. So, you can see here it's referencing different files, and it's being specific about some of the changes it might make. This is really your opportunity to ask for changes. And it's important that you ask for changes in the planning stage rather than at the changes stage.

Because as soon as the coding agent tries to make changes, it's very difficult to then undo those changes. AI coding agents are lazy. And so, if they make changes, and you later decide you don't like those changes, what you'll end up with is some compromise between the first attempt and the feedback that you gave it, rather than the clean change that you otherwise would have wanted. So, this plan is really your opportunity to kind of set the approach, and from there you're much more likely to have a successful outcome. So, visualizing how planning might affect review, I've got two timelines here. So, the top timeline is one where the person in the driving seat didn't spend a lot of time planning. And so, you can see that they did a small plan. AI went and made a change. It then came back to the human and said, human, please review this. The human had some feedback. And then, you can see this goes back and forth with AI making further changes, the human reviewing, AI making further changes, etc. The other option, option B, is you invest more time in planning.

You give it much more detail. And sometimes that means that AI is going to run for longer, because you're actually going to be making it work harder than it otherwise was going to, to achieve the change that you want. But the trade-off and the upside of this is that now, you probably only have to do one round of review. So, you check it. Maybe there's some small things that are off, but directionally, the work is correct. And that means that you only need to implement one round of feedback. So, the thing that really takes time is when you as a human have to get involved. And review is one of the parts of the process that is really human labor-intensive. So, if you can find ways to review the changes less frequently, you'll find that that has a huge impact on your ability to do more, because it's so time-consuming.

To recap, plan more and you will review less. Now, we're going to talk about review. And there's two parts to the review puzzle. One is reviewing the change itself as a user, maybe checking a browser to see if the features have been implemented correctly. And the other is actually reviewing code. So, first, we're going to talk about the first one of those options. So, how do you review things for their functional correctness? And basically, one of the biggest things and the biggest mistakes that I see people make is they have a really complicated dev setup. So, at the kind of extreme end of the spectrum, maybe you need to have a dev server running, but that's connected to a dev database. But that dev database doesn't really do much unless you seed it with some dev data, which nobody's written down anywhere. So, you need to do it manually yourself. And before you even are able to test the change, you realize, wow, I've got to go and click around and do ten minutes of work just to test this change.

So, what I really advocate for is making it as simple as possible to run dev servers, both so that you can test things, but in the near future, AI is going to be able to test work itself, but it's not going to be able to do that if it's spending ten minutes setting up your dev server and even getting it running in the first place. So, here you can see on the left I've got my chat, and on the right I've got my dev server. And I'm able to start that so easily because in one click, I can run a dev server from within Vibe Kanban. We also have a dev server set up where we actually copy all of the data from a seed directory into the database each time. So, you have this kind of fully verifiable environment that is set up that's pretty dependable each time, and there's zero setup needed other than me just hitting the start button to be able to test things.

My next tip is around code review, and the most important thing with code review at the moment is actually to make sure that you're getting AI to work for you to review the changes. And this isn't something that I actually would have recommended maybe even as soon as three months ago, but very recently AI code review tools have gotten really good. There's a lot of different startups out there that are tackling this as their sole mission, but I actually think that the out-of-the-box Claude code and Codex coding agents and the review feature of those coding agents are pretty much as good.

And so, what I do is in my editor, which is Vibe Kanban, I ask for a review which will run under the hood either the Codex or the Claude code coding agents in review mode. And you'll get back a list of changes that the agent has identified you should make. And often, they've got a priority like P1, P2, P3, and you can decide, okay, well, yes, this is important, please go and fix it. Or alternatively, that looks a bit like maybe something that's style. And actually, I prefer a different style, and you can skip them. But it saves me a lot of time.

I do this on all of my pieces of work now before I go and review the code myself. Which brings me on to my next point, which is you want to review the code, but you only want to do it once. And I'd recommend doing that right at the end of the process. So work with AI, get a plan, execute the plan, test the change functionally, maybe give feedback to it based on that. And only once everything is working, start to look at the code because it's incredibly time-consuming process to read AI generated code.

So again, you want to minimize the amount of times and the amount of time that you're spending doing that. This is how I review code quickly within the editor that I use. I have my logs on the left, but on the right, I'm able to pull up the code change that the editor at the coding agent made. There's a small diff here, a stylistic mistake, to import midway through the file, which I know is not totally illegal, but it's something I just don't really like. I can comment on these changes mid-file, sending feedback to the coding agent to make the change. Hopefully, there won't be any further changes after I've given that feedback.

This is a kind of thing we've touched on throughout the conversation, but I wanted to emphasize it as its own piece of advice. Working on multiple tasks in parallel is crucial. Coding agents can run for 5, 10, sometimes 15 minutes unaided. Running two or three tasks simultaneously keeps you constantly engaged. I illustrate this with an image of the board I was working on today, showing three different tasks. One running task keeps me busy while I unblock and edit others. This experimental time in coding requires adapting workflows to leverage AI effectively and automate tasks it requests.

It's your job to figure out how to automate tasks AI asks you to do. Keep a mental log of AI requests and find ways to automate or simplify them. Reduce setup time needed for tasks to maximize productivity. This is the key lever to getting more done efficiently. Thank you for listening. I dedicate considerable time to how we work with coding agents. You can follow my experimentation on Twitter at token gobbler or check out our work at vibekanban.com. Explore the Vibe Guide for tips on maximizing efficiency when working with coding agents.