What real aiassisted coding looks like in 2026 for existing or complex codebases Tips and tricks full walkthrough is up on my yt channel where I implement a feature from scratch At most Im usually working on 1 major thing then have 23 more Claude code sessions open for minor stuff usually design or research related coding software programming ai artificialintelligence

Here's seven things the best AI coders do every single time. According to Anthropic, 90% of cloud code is now written by AI, which is crazy. But Anthropic has some of the best engineering talents in the world, so there's an incredible amount of technique and oversight that is going into their process. Here's exactly how you should code with AI if you are building real products, if you have complex existing code base, if you have thousands or millions of users and code quality matters and refactoring matters to you. Number one is to clarify terms. If you expect AI, if you push a button and it just magically codes everything and it works, that is completely unrealistic, especially if you have a complex existing code base. I use the term bytecoding to refer to like zero supervision of AI code. In my day-to-day work of building a SaaS product with thousands of users, what I do every day, I call AI-assisted coding. The most important thing is I spend probably 90% of my time interacting with cloud code, planning what it's going to do. I have a super extensive cloud MD file that I'll show you and I have a shortcut Q plan, so we're going to plan how to implement this feature, especially reuse existing things in the code base and introduce minimal changes. I only work on one small chunk at a time, so iterative chunks here, one little feature at a time, then build upon that, give it as much information as it needs to complete the task successfully, so feed in context, edge cases I'm thinking about, all these different things. For model selection, sometimes I'll swap them, but I'm usually almost always on Opus or whatever is the best model at the time. Then I pair AI coding with sensible guardrails, such as linters and CICD. For example, cloud code can push to GitHub, but every time it pushes, it triggers a lint check. Just to give you an example of my very extensive cloud MD file, this is already outdated, almost a year old, and it's way, way, way, way longer than this. This is just the introduction, but basically you want to be continuously tuning your cloud MD file or your skills, depending on how you organize it. If you notice a cloud code struggled with a particular implementation, ask it to reflect on the conversation, what's happened, and to brainstorm something useful that can be added to your cloud MD file or your skill, so you can avoid whatever mess was created next time you implement something similar. Make sure you commit frequently and review the code. This is really important. I know you want to skip it, but it's so, so important, also because there needs to be someone who understands what's going on in your code base. It's going to matter when you have a critical bug, when you're trying to add something new, when you need to refactor it, because nobody understands what is going on in the spaghetti code. And as a result of AI coding, I spent a lot more time on system design, which honestly is fun for me, like thinking about trade-offs, thinking about what we could do here and there, and I spent a lot of time reading diffs. And so one skill to gain, and that's going to help you a lot with AI-assisted coding, is getting used to reading diffs really quickly, understanding what AI is trying to do, how it changed the system, what it did wrong, and giving that feedback, continuously improving your CloudMD file or skills.