How I AI: My Weekly Codex Experiments

People like to ask, Nate, how do you use AI? Tell us your secrets. Let us in. So, I'm just going to start doing weekly. This is how I've been learning about AI and using AI this week. And and I'm going to focus on the things that are surprising or interesting. This week, what's been standing out to me is I think assembling context windows has gotten a lot more interesting when it comes to how codeex works on your local file system. And I am saying codeex specifically versus clawed code or clawed co-work. I have tried both with this same workflow. It doesn't work. So what I do with codecs now when I want serious work done is I actually tell Codeex to look at my file system overall and to make copies of certain files that I describe in natural language. Not by describing the title or naming the specific subtitle in the text, but by saying this is what it's about. This is about when I made it. Can you please find it? And then it inevitably finds it. It's it's amazing. I don't have to keep track of my files anymore. and Codeex pulls all of those files in copies into a working folder and it's a very neat working folder. And then I open a new chat in Codeex and I say, "Hey, we've got a really clean context window here. Look at this particular folder and here is your task." And if I have detailed instructions, I copy those into the folder as a specific transcript as part of the task and it just goes to work. And I found that that enables me to do very long document work. I can do 30, 40, 50,000 words that way very easily with codecs. I also find that I can do complex document work, complex spreadsheet work, complex coding work that way because it's able to work across that folder structure easily. And that kind of makes sense because if you think about where Codeex was coming from, it came from a sandbox world where you put everything into the GitHub and everything into the repo and it can just read all the code files. Well, code files and text files are basically the same thing. Uh it's just whether it's reading code or text, it's looking at files in a folder and trying to figure out how they go together. And that's something that Codeex has been able to do. That may be related to Claude's compute shortage. It may be 4.7 itself just not being quite as solid as as 5.5 at its ability to do longunning tasks. I'm not sure. Uh but I've been loving that cuz I can assemble a context window on my local machine in folders and it's really clean and efficient to do. So that's one of the ways that I've been AIing this week. Another way that I feel like I've been evolving AI and this is not related to files per se. It's something that does apply to both claude and codecs. My prompting has shifted a lot even in the last couple three four weeks. Prompting, you know, before 2025 and the inflection point in December was very much about prompt engineering and how you structure the the prompt and making sure that you have the right stuff in the right order. That still matters for certain oneoff workflows. It's still a helpful skill to know. Between December and and we'll call it April of 2026, uh I feel like a lot of what I was doing with prompting was essentially giving models that had longunning agentic workflows tasks, pointing them at files and telling them to go get work done and telling them what good looks like, which you can do with an eval or you can do in small in small cases one-off cases maybe in the prompt itself. um you know since May uh since 4.7 dropped since 5.5 especially dropped and and codecs got refreshed I have found that I am shifting from here is your task please go and get that task done with these files and resources and here's how you know it's good is a set of meaningful questions I have that circle around the standards I want to meet when I get this job done. here's some files I think are relevant for the task. Help me to define the shape of this task first and then once we define it then we can go execute it aentically. I can be at that messy stage with these agents and I feel like it's truly a back and forth collaboriveness especially with 5.5. I feel like the model doesn't get lost when you then shift gears and say, "Now go do it. Now go get it done." And so when people say like, "Well, why are you talking about codec so much?" It's because my felt sense of what I'm able to accomplish has shifted in the last three or four weeks. And you know, I know the AI race by now. I I am it's going to shift again. I know it will. And that's okay. I don't need any particular team to win here. I just want to get more efficient at working. And I'm finding I can do things like simultaneous drafting in a local folder. Uh developing a series of eight or nine prompts to run at once in collaboration with a model that can then execute on those sequentially. There's all kinds of really fun things you can do that essentially are unlocked by codecs having a really efficient way to use files, a model that knows how to stay on task for a long period of time, and an excellent autoreview system where you can let it run on your computer and the auto review puts good guardrails around it. And so that's one of the things that's been really revolutionary for me because I feel like I'm at a point now where I can do multi-threading and it works really well in codecs. I can incubate multiple ideas at once and it's just so much fun. I feel like I can shape and direct my ideas more efficiently and I have no doubt that the team at Anthropic is busy shipping 4.8 or whatever they're going to call the next model and it's going to be amazing and I'm I'm very excited for that. Uh there's a reason why I don't pick a side. I just get excited about what AI can do for you. I I think that is the right place to be if you're interested in how AI changes our lives. I I think it's compelling. So, I hope this little peak into how I'm using AI this week has been helpful.