Transcription
Okay, so we're seeing a very interesting shift when it comes to software building with coding agents. The harness around your coding agent is probably more important than the raw intelligence of the model itself.
Now, the good news is that irrespective of what harness you use, we're seeing emergent workflows that seem to work really well. In this video, I want to walk you through a simple workflow that has been seen to work really well for building effective software. This is simple: plan, execute, test, and verify. This simple workflow, irrespective of the coding agent, will help you build better software.
Okay. So, let me show you a practical example of this workflow. Okay. So, the workflow I'm about to show you is going to work in any coding agent. It's basically plan, implementation, testing, and then adding more features. But for this specific video, I'm going to be using this new coding agent called Verdant. I was part of their early access program when it was initially released. I really enjoyed working with this specific coding system, and I'm really happy that they are also sponsoring this video.
Now, there are a couple of features which make Verdant different from some of the other agents that we have seen, and one of them is the ability to do parallel work. So, it uses Git trees where you can implement multiple different features in parallel without affecting the main code. It has a very clean UI compared to other coding IDEs, and I think it's a perfect system for VIP coding.
Now, it has two different modes. One is plan mode. The other one is agent mode. Agent mode is for implementation. Plan mode is for initial ideation. This is great for refining your ideas. I usually like to use these coding agents as brainstorming partners. So, we're going to start with a simple prompt, and then we're going to build on top of it.
Okay, so for the app, we're going to be building a text-to-image generator using Nano Banana, and we're going to incrementally add more features. If you're coding with coding agents, one thing you want to do is to use dictation software because right now, the input is actually the bottleneck. So, for this, I created my own dictation system. If you are interested, the link is going to be in the video description.
"Help me create a text-to-image generator. We're going to use the Google Nano Banana model to generate four different images."
Okay, so I'm happy with the first iteration. We're going to just send this in. Okay, so right now, it's using a search tool to look for the documentation. Usually, I recommend including specific links because you don't want the agent to be wasting a lot of tokens when it's using tools like search or different MCP servers to look up documentation. So, in the process, it found the correct API documentation, which is pretty neat. It also has a good understanding of what the return format is going to be. It does ask you some clarifying questions if it has any confusion, which is pretty neat because you want to provide this information upfront.
"Okay, I just want HTML with CSS and JavaScript." We're going to submit this. Right now, we're using Opus 4.5 for planning, and we're going to use the same model for implementation as well.
Okay, so it came up with a detailed plan of what exactly it's going to implement. So, it's going to use the API endpoint. I specifically asked it to use the REST API. I don't want to use any SDK. Next, I said, "Assume that it's designed by a billion-dollar design company." So, here's the type of aesthetic that it came up with. Later on, I'll show you how to further improve this design with skills. Next, it's talking about how the image generation and image display is going to look like. You want to go through the plan and make sure that you are aligned with the plan. Now, with the plan, it also put what the definition of done is going to look like. I'm happy with it, and we're going to let Opus 4.5 cook.
So, when you click on build, it automatically switches from plan mode to agent mode because now it's going to the implementation. Right now, it's running this one task, but you can create another task which is going to run in parallel. I'll show you an example later in the video. This is powered by new workspaces. If you create a new workspace, that will create an isolated folder from the branch using Git workspace. This will enable you to run parallel agents, which is going to make sure that they don't interfere with your existing implementation. You also get a view of the implementation itself. I think it's a more autonomous coding system. So, even if you're coding, you don't want to look at the code that is being implemented, I think it's still a really good system for that. And you can directly generate code reviews here. You can also open a specific file in the code editor if you want to look at what exactly is being implemented.
Okay, so here's what it implemented, and it looks like AI slump, but we're going to fix that, and I'll show you a very easy way of doing it.
Okay, so I got my API key from AI Studio. Now, let's see if this is going to work.
"Create an image of a llama wearing sunglasses standing in front of a snowy mountain."
All right, let's send this in. There's an animation going on, which is good news. It seems like only one image was generated. This is probably to do with rate limits, but still, we have a success. In this case, we had to manually verify the outputs, but it's 2026. We shouldn't be doing that. In this case, we're going to give the agent some tools to verify this web app. We're going to go to settings, then click on MCP. One of the MCP servers that I highly recommend for any web development is Chrome DevTools. You can also add something like Playwright for browser automation, but we're going to just enable this to use your Chrome DevTools to verify whether it's working properly or not and fix any issues that you encounter.
Okay, we're going to send this in. Right now, it's being completely controlled by the agent. Okay, so it just wrote a prompt. It is hitting the prompt. I think it made some code changes, and now everything seems to be working properly. For MCPs, it gives you a list of all the tools that are available. So, this is pretty good. You know exactly which tools are in this MCP server.
Okay. Now, I don't like the design of this page. So, let me show you what we can do here. Now, if you look at something like Cloud Code, it has skill integration where you can just point it to a custom skill, and the agent will be able to use it. As far as I know, Verdant does not offer that. But there's a very simple workaround. And this is going to work in any agent coding system where you can just use custom commands. So, this is the front-end design skill that Cloud Code uses. I copied all of this text. It's a single markdown file. Nothing complex. And I went to command, added a new command, which basically is that whole text. We're going to call it "front-end design," and here's a quick description of what exactly this skill does.
Now, in order to run that, I'm going to create a new workspace. Let's call it "design." I'm going to say, "Use the to make a modern UI for this web app." We're going to send this in.
Now, at the same time, I can create another task. Let's say I want to implement a dark theme. "Enable the user to switch between dark and light theme by adding a toggle." So, now we have multiple parallel agents running simultaneously. So, this one is going to be working with the existing app, and the second one is working on another branch.
Okay, it seems like it has completed the toggle implementation between dark and bright mode. This is in the original app. We can actually toggle between the two. Seems to be working fine. The renewed design is also complete. We went from this to this, which looks pretty awesome. Now, we can definitely use some work around the size and placement of the images. Well, let's see if this actually works. Rather than manually testing it, I'm going to ask it to use its Chrome DevTools.
"Use the Chrome DevTools to verify that everything worked."
Okay, so it opened up the app. It wrote a prompt, and it's verifying its work. Seems like everything is working. And it always amazes me that these agents are able to now easily use these UI verification tools. So, here are the different tests that it ran. Everything seems to be successful.
Now, since we have the basic layout working, this is where I'm going to start incrementally adding more and more features. Okay, so with Verdant, I have noticed something that it starts compacting the conversation pretty quickly because I think it's not using the full 200,000 tokens of the context window. Now, that means you want to incrementally build different features. Don't go all out. This also gives the model a lot better context, and it's going to result in better outputs as well.
Okay, so we went from something like this, which looks like a typical AI-generated website, to something like this, which is a lot more elegant design in a few steps. So, as you saw, my workflow is extremely simple. It includes three steps: plan every feature. I usually spend a lot of time planning with these coding agents, so I use them as brainstorming partners. Then, execute that plan, test it, verify it myself, and if I need to, completely automate it. And then, I will incrementally add more features.
Now, this might sound like a slower execution cycle, but in the long run, it actually is much faster because if you try to implement 10 different features simultaneously, you increase the surface area for failure. Also, I don't really use a lot of sub-agents. There might be one or two for verification and testing purposes, but that's pretty much it. One thing I do is I create custom slash commands. For example, the design skill was not available in Verdant, so I created a custom slash command for that. That's pretty much it.
And tools like Verdant make it very simple. One good thing about Verdant is that it's a very clean UI without a lot of distractions. So, you can just focus on your ideation and then work with the agent. It asks you these clarifying questions, which usually are really good for you to refine your thought process. Right? So, do check them out. The link is going to be in the video description. I hope you found this video useful.