Claire Vo: There are people out there definitely like me that really want to know the advanced techniques that can leverage the most powerful parts of these AI powered coding tools. Where do you want us to get started that you think many people don't think about in terms of how they can use these tools?

John Lindquist: Context and diagrams is a great place to start. They're definitely the best way to get AI to do what you want. So, they have what are called mermaid diagrams. This is a way of visualizing database operations and it's a way of essentially compressing your application down into very small lines of text that show how your application works. Now for a human to read this, this is a big challenge. But an AI can consume this easily. I could even just say, "Please explain the authentication flow." And because it already has it in the context, it's not going to have to do a bunch of file reads and codebase exploration to figure this out. It's going to come up with results much quicker.

Claire Vo: Welcome back to How I AI. I'm Claire Vo, product leader and AI obsessive here on a mission to help you build better with these new tools. Today we have John Linquist at egghead.io who is a super user of AI powered engineering tools like cursor and claude code. Now, I love all you nontechnical folks out there, but this is an episode for the senior software engineers who really want to understand how they can use the power features of some of these AI engineering tools to really both optimize the quality of code that they're generating, but also become more efficient as they use their IDE, terminal, and AI assistance to write, check, and deploy code.

John Lindquist: Yeah, I think context and diagrams is a great place to start for us. They're definitely the best way to get AI to do what you want. So we'll be using cloud code throughout. These diagrams are all generated from I can share a prompt with however you want to share with the audience that can walk through your codebase and generate diagrams based on user actions or user interactions the events the channels whatever happens in your code to help the AI understand the flow and how the pieces are connected. I think Windsurf recently came out with something called code maps, a similar concept. Essentially preloading valuable context so that you have to remember that every time an AI starts it has no memory, no idea of what's going on in your application.

John Lindquist: We can do that using diagrams. Uh, so for example, one of these diagrams um will have I I call they're markdown files with diagrams in them. So they have what are called mermaid diagrams and mermaid is a standard format for rendering diagrams inside of markdown. So this is a a way of visualizing uh database operations. And it's a way of essentially compressing your application down into very small lines of text that show how your application works. Now for a human to read this, this is a big challenge. But an AI can consume this easily and it's like a very compressed very um robust way of explaining application.

John Lindquist: The one we're going to focus on is called append system prompt. So in there before we load in uh any sort of user prompt or anything we're actually going to say claude append system prompt and then you can drop in a command and this command can read in from our memory from AI/diagrams and then this is a called a glob pattern read through all of the markdown files essentially force them into claude once I do this. So this is reading all the files all the markdown files and this is cat will kind of concatenate them all together into a single uh text read.

Claire Vo: One thing I want to call out for folks that are watching this is two things. It seems like, in your standard repos, you're creating a memory directory where you're going to structure some of the context and files you might want any of these AI tools to use. The other thing that I think a lot of people are quite lazy about is they haven't explored the surface area of all the system commands available in cloud code. And appending system prompt is one of those ones that I think people probably underuse.

John Lindquist: Yeah, absolutely. Um, it's one I use constantly. So when I let this run, you'll notice that it's now prompting the user to do something. And we don't have to try and reference all of the files, which you normally do with AT. I could even just say, like, I use dictation all the time. Please explain the authentication flow. And because it already has it in the context, it's not going to have to do a bunch of file reads and codebase exploration to figure this out. This does come at the cost of a lot more context, a lot more tokens being used up front, but the work that you do, the time that you spend on these tasks is more valuable than that to me.

Claire Vo: This is the era of the file type and I think so many people think about markdown and JSON files as effective ways to inject context into LLMs. In this one, you have mermaid diagrams, which again are hard to parse as a human. But to a machine, it's very effective. I think this is an interesting moment where we can all use different file types in a more extensive way than our human brains could because the machines are so good at using the different component structures or syntax of those files.

John Lindquist: Yeah, absolutely. There's a lot of research being done into how they can compress all of this information down into like a single image. I'm huge on video and using videos. Gemini being the best model for uploading video and understanding. I recently built a tool that can take one of my six-hour workshops and process the entire thing and take out notes and examples and thoughts and frequently asked questions.

Claire Vo: Where in your development process do you find that you generate those files? So for me, I actually have a GitHub action that generates files almost exactly like you have with documentation and diagrams for new features of a specific scope. And so I do it when a pull request is closed and then I go back and update our diagrams.

John Lindquist: Yeah, usually I think pull request is a good paradigm there. As soon as you have something working where you want it to be working and then you can say okay now this is working as expected. Please diagram it. I think for a lot of the projects, we already have pre-existing code bases that don't have diagrams. And so that's been the major use case is taking existing stuff and diagramming all of that so that our AI development is accelerated.

Claire Vo: Yeah. And then I will give folks just a couple other use cases of generating mermaid diagrams from code. I use a lot of diagram generating out of our repo to answer very complex security and data flow requirements from our customers. This is a workflow that is actually pretty expensive if you ask an engineer to do it. Also if you're going through sock 2 compliance, these are assets that historically have been so tedious to create efficiently and now you can kind of generate them on demand.

Claire Vo: You showed us how to just pull all of these documents into a system prompt. And this seems like a command that you're using over and over again. And that's something you and I talked about before we started recording, which is how to alias and make more efficient your use of of different commands.

John Lindquist: So there are a lot of uh with with on Mac it's ZSH is the default shell. Depending on what tools you use the most, you can easily set up aliases for things like setting the default model for for claude. If I just type X now, anything I type has bypass permissions enabled. Or if I type H, this will be Haiku. Or if I type in the scenario CDI, this will do that diagram loading. I keep them in very short shortcuts.

John Lindquist: So I tend to build any idea I come up with. Um so for example this is one I'm working on called sketch and this feeds into the Gemini CLI. You can tell an AI like listen I want to use like this is a wrapper around Gemini where it will execute Gemini with specific prompts. You can have these little CLIs and these little projects that are be just for you. Because you just like build the tools that you need now.

Claire Vo: I think we're going to close out and spend a little bit of time on your workflow for when you're doing more complex coding projects or features, how you keep those really high quality using some advanced techniques in I think in cloud code and cursor.

John Lindquist: Often when the AI is generating code it'll often build out mistakes. So you would run something like bun type check and you would see that it has this error but your claude code and the other agents don't know that this error exists. What Claude has is the concept of hooks. I'm going to set what's called a stop hook.

John Lindquist: We're going to focus on is we're going to see step one were their files changed when we stopped. And a stop is once Claude has kind of finished its conversation. If there's files changed we're going to say okay then let's go ahead and run that bun type check. And if there is a type check then we can say back to claude hey there were typescript errors this is the report and then send them back the output. Otherwise, if there were no errors, then go ahead and commit.

Claire Vo: I like this combination of kind of like structured commands in the terminal combined with natural language calls back into Claude to then kind of put the bow on the end of any work that this AI system does. If you haven't created these hooks for key repos where everybody is benefiting from this when they're using something like claude code then you're missing out on some of the scaled leverage of these tools.

Claire Vo: I think there's interface wars happening right now. Are people going to love these terminal UIs and command line tools like Claude Code? Do people want the IDE? What do you think wins for real software engineers out there?

John Lindquist: Yeah, so I think you need both. I think you need an IDE and there are so many use cases for the CLIs. The reason being that the CLIs have a lot of configuration and a lot of settings. But if you have an IDE and you're reading through the files and you're selecting lines and you want to modify certain bits like focused work, there's so many use cases for IDE. For one IDE to stand out above the other they have to separate themselves like cursor is doing with their agent mode. You have to focus on the UX you have to make that experience better than everyone else.

Claire Vo: What would you tell kind of senior principal software engineers, engineering leaders? How do you make that pitch for the value proposition of these tools?

John Lindquist: The first thing that jumps to mind is anytime an issue is opened like you can set up streamlined workflows that someone opens an issue and you can have claude automatically tackle it. All of that busy drudgery that you're going through to even get started on the issue, it can wipe out so much of that. If you're not using it to like inspect and investigate and write orientation and all that stuff, then like you're really missing out.

Claire Vo: What I often tell people is a good way to think about how to design your AI workflows is do not think in a task level orientation like I'm going to write code. I say, think about if I gave you infinite junior to mid-career talent who is always available, who would do the work you would do if you had unlimited amount of time and no meetings. All of that could just become a prompt.

John Lindquist: And something as simple as the commit messages are so much better than they used to be because developers didn't have to write them. Commit messages used to be like second attempt or like please work or swear words.

Claire Vo: Yeah. Thank you so much for joining us and sharing your workflows. You can find John on egghead.io.