10 Pro Tips for AI Coding

Key Takeaways

AI coding has completely changed in the last few months. And the techniques that worked the best six months ago are no longer the best way to build software. So today, I'm going to share 10 tips with you that are going to help you take your AI coding to the next level and make software as quickly as possible with the highest level of quality. These tips will work with any AI coding platform, but they work best with something like Cursor or Clawed Code where you have direct control over the code and the files that are involved. My first tip is to just use agents. Use a tool like cursor or claw code where you can just have an agent that can read through all of your files, find the relevant context, and modify multiple files all at once to try to build your feature. Don't spend time copy pasting from Chad GPT or using autocomplete to complete one little function. Just launch an agent and just build whole features at a time. Sonnet 4 is an incredible model. It's perfectly capable of that. So now the bottleneck has shifted from trying to get really accurate low-level code to building whole features and matching the requirements that you give the AI. Speaking of which, my second tip is to use markdown files to plan your features before you have the agent implement them. This is going to help you split up the work so the agent doesn't have to figure out everything on the fly. You can plan the work ahead of time and you can review the plan and make sure that the way the feature is going to be implemented is done exactly the way that you expect and the way that you want. This should take up the bulk of your coding time nowadays. Figuring out exactly how features should work and then delegating that work to your AI coding agents. On a practical level, the way you would do this is you would describe the feature that you want to build to your AI agent and directly ask it to come up with a plan and create a markdown file to summarize all of the things that it needs to do. Then you can review that plan and ask for edits. And then you would actually start up a new conversation and have a separate AI agent review that plan and start working on it because that way you have a clear set of context for it. And your plan should include all of the things that it needs to do. By taking this approach, you're effectively writing a very complete prompt by creating that document and then giving that to the agent so that it can complete all of the work. Now, when the agent is complete, there's another step that you should take, which is my third tip. have AI review the code and write a status file. This is going to help you catch any sort of bugs or defects from the original plan. Essentially, once the agent has finished, just ask it to look over all of the code that it has changed and compare it to the original plan and then write a status document. Oftent times, you'll find minor little defects that the agent can then fix. And the great thing is you'll have this status document that you can pass to a future agent. So if you're trying to modify the feature or trying to iterate on it, you can simply pass the original documentation and the latest status to your new AI coding agent so that it can pick up the work, understand exactly where things are at, and then know what to do from there. My next tip is to run multiple AI agents at the same time. The thing is with these AI agents having the ability to open files and read them and modify multiple files to complete a feature, it just takes a while for them to finish their work. But you don't have to sit around just waiting for them. You can actually run multiple AI agents in parallel. In cursor, you can simply open a separate chat or even run a background agent. And in Claude Code, you can use tasks to do multiple things in parallel. By using multiple agents, you're going to be able to utilize all of your time by writing new prompts and reviewing results, and you really get into a flow instead of just sitting around waiting for the results. Practically speaking, the best way to actually do this is to have the AI agents work on unrelated features with files that don't interact with each other. That way they're going to be able to just keep working on whatever they're working on without trying to overwrite the other agents file which can lead to some conflicts and then you have to like manually approve that okay this agent can work on it now now this agent can work on it. So I recommend that you just have the agents focus on separate features. Another way to utilize this is in a new project when you're just setting things up. Have an agent build out all of the front end and have a separate agent build out all of the backend. I've had parallel agents write like 5,000 lines of code and most of it just worked out of the box. This separation helps the agents focus and deliver higher quality code and gets you that code more quickly. Speaking of new projects, my next tip is to use starter kits to kickstart your project instead of writing everything from scratch. You'd be surprised how many times AI agents get tripped up by silly stuff that you have to do for every project like configuring Tailwind or Shad CNN or your database integration or O. You're going to be doing all of this stuff for every project. So just use a starter kit that has all of this built in. I actually created one. It's free open source. It's create Vololo app. You can try it out. You run one command and you have a whole project with a front end, backend, database, authentication. It just works and it's a great way to get ahead and start working on features instead of initial configuration. If you want to use some other kind of Nex.js starter kit, there's plenty of other boilerplates out there, but create Vololo app should help you get started very quickly. The next tip is for troubleshooting, and it is what I call the beaver method. When you get stuck, it's hard for the AI to figure out exactly what is wrong by you just describing it. So, what I like to do is ask the AI to add a bunch of logs to my code so that you can run it and then you can copy paste those logs back to your AI agent and that will greatly help it figure out exactly where a problem is happening. You can actually take this a step further and have the AI think of the top five root causes for whatever issue you're facing and then add logs that help it identify which of those root causes is actually causing the problem. This approach has greatly simplified and accelerated my troubleshooting process. My next tip is that you should actively manage the context of your AI agents. It's easy to get carried away in a single chat and keep trying to fix an issue, but oftent times the agent can't get through something simply because it's already done so much and its context window is overloaded with a bunch of irrelevant information that is distracting it. You just got to ignore those token context windows that AI models have because agent performance goes down dramatically way before you get to the actual limit. So in practice, you really do want to have your agent only focused on a specific set of files. And that's why the documentation approach that we talked about earlier helps so much. It condenses a lot of information about your code in one document that the agent can use. So it's a really powerful way to actively manage context. In addition to the feature and status docs, you should also tag specific code files that are going to be relevant to whatever the agent is working on. And you can copy errors and logs. And if you need to, you can also tag at web to have the AI agent search for some additional information. The main point being keep your context low so that your agent can stay focused and start new conversations often. My next tip is that you shouldn't just build features. You should have your AI continuously looking for opportunities to improve your code quality, performance, and security. That way, as you're building, you're going to feel confident about your code and know that you're handling things in the right way, and there's going to be no nasty surprises down the line. But the truth is, it is really hard to anticipate every possible performance or security or code quality problem. Which is exactly why I'm building a new product called Vibe Scan that is going to help you analyze all of your code. It's going to scan all of your different files and evaluate it for all of these potential issues. That's going to help you ship your AI code confidently and know that there are no surprises in your code. Vibescan is launching next month and until the launch there is a 50% off for life discount if you become a founding member. So check the link in the description and I'd love to give you that discount and get your input as we build out the product. Tip number nine is to use rules files in your editor like cursor rules to help the AI agent run the correct commands. I resisted using rules for a very long time, but more recently I found a very minimal set of rules that really does help me get better performance. And I found that less is more. So I just keep my list of cursor rules to the specific technologies that I use and the specific types of commands that I need to run to add dependencies or install new components. I also have a rule reminding cursor that it doesn't have access to the env files because I keep those in my git ignore so that my secrets aren't shared with cursor. Some other builders have more sophisticated rules files and I've seen them have success with taking a test-driven development approach using rules to first write a bunch of tests and then implement a feature. I haven't tried it myself but it's definitely a good use of the rules files. Oh, also there is no standardized rules file name. So every editor is going to be a little bit different. So just make sure whatever editor you're using, you have the right file in your codebase. For cursor, it's cursor rules. Okay. My final tip is that you should be vigilant and don't let bad code take hold. What I mean by that is you'll oftent times ask an AI agent to build some feature and halfway through it as you're looking at what it's actually changing, you realize like this isn't exactly what I meant. Just stop the agent right there, roll back to the previous checkpoint, modify your prompt so that it's more clear and try again. If you let the agent implement everything and then you try to paper it over, some of those files might get forgotten. Some of the context might get written into your documentation or your status files and things just kind of rot over time. You don't want to allow that to happen. You just want to stop the AI agent in its tracks, go back, fix it, and make sure that your code is clean. This means that you should be looking at your code. You shouldn't just be completely vibe coding and ignoring the code completely. And this attention to detail is going to set you apart from a lot of people that try to build full stack applications and fail while you are building a real highquality piece of software just using the latest techniques. If you want to see me put all of these tips into action, check out this video where I build a full stack application in just a couple of hours using all of the approaches that I just talked about and going into a lot more detail. I'll see you there. Thanks for watching. Take care.