Getting into Flow with AI Coding

NatashaTheRobot Aug 30, 2026 I’ve been seeing a lot of posts lately saying that software engineers who use AI miss the flow state. This is a bit surprising to me because if anything, using AI puts me into an ULTRA INTENSE flow state haha. So I wanted to share how I code with AI in case it helps others. First, there are a lot of posts from developers saying they no longer write or review code and that they have 100s of agents running at the same time doing their work for them. While this type of content grabs attention, I personally cannot fathom such a system. I’m sure it’s possible… I just don’t see it as something that would build anything that could be shipped to production and maintained. If you’ve worked with LLMs for coding even a little bit, you’d know that they make A LOT of mistakes at every turn, especially in Swift. Having them go unmonitored will just create slop on top of slop on top of slop. But maybe that’s the goal for those doing it… I don’t know. Anyway, I keep my setup super simple and flexible. No need to complicate it. I love using Claude Code inside the Claude app. I appreciate all the developers out there who say they no longer need a UI for anything and use the Terminal version, but I LOVE having all the UI features. I also set up the Claude Agent inside of Xcode for minor issues. And as a browser extension for when I need it to work with super specific sites. The Fable model is a HUGE improvement over all the other models we’ve had so far. I tried the much hyped OpenAI’s Codex with Sol, and it’s just nowhere near the same. But first, the mindset. While I don’t mind outsourcing a lot of the work to a coding agent like Claude Code or OpenAI’s Codex, I do not outsource my control. I keep myself extremely involved in the whole process instead of just turning off my brain and letting the agent do it all. I think this is the main part that turns coding into AI from something super boring (mainly just waiting for the agent to finish) to something active that THROWS you into that EXTEME ULTRADEEP FLOW STATE. I start a complex task by describing it as best as I can to Fable in general terms, ask it to research relevant topics, and then have it write out an Artifact (a markdown file) that contains all the steps of the plan. I actively read as it thinks and that gives me ideas for more things to add / ask / decide that I can throw back at the agent. I also keep any extra ideas in a note in my Notes app so I don’t forget it if it’s not relevant at that moment. I stay super active the whole time! Once the plan is completed, I review it deeply and interact with the model as I read it. I say no to certain parts, modify others, add to others, and have it explain parts I don’t understand. As the model works on each part, I constantly ask it to update the Artifact with the latest decisions or modifications. This serves two purposes. It keeps the model on track. And if the session runs out of context, the same file can be given as the main context to the next session. Before the session context window ends, I have the LLM modify the file with more details about what has been done and what is still pending. I then go step-by-step from the file and have the model break down each step even further. The issue with agents is that they are very trigger-happy. They are ready to ahead and code EVERYTHING right away. So I have to explicitly ask them to NOT CODE first at almost every step. For each step in the plan, I re-read it, clarify, modify. Then I have the model break it down into clear commits so there is not one MEGA commit. For each commit, I make sure to review the code written and run the project at that step to make sure it works as envisioned. This is where I’m extra confused by how developers could outsource this process completely. For each commit, I find myself making a ton of small micro-decisions which are extremely important for my product. While the AI can make those micro-decisions for us, I find that I have to steer it pretty closely to make sure the app ends up with MY DECISIONS and not some random ones. I know some people will say to just write a SKILL file for the model, but these are 100s of small completely different types of decisions based exactly on the context of that one commit. It would be impossible to write that skill - that is literally the work of my whole brain! Another benefit of breaking down each step into smaller commits for a problem you haven’t worked on before is that you end up LEARNING A LOT. More than was ever possible before! In addition, smaller commits take a shorter time for the agent to complete, so this is where the FLOW STATE comes in easily. My attention is now fully occupied by reviewing the plan, breaking it down, questioning it, learning about it, reviewing the code, refactoring it if needed, running the app to see how the UI turned out, modifying the UI, noticing other bugs or issues I want to fix, writing those down in the Notes for later, and so on! It all goes so fast, and so many things are build with incredible polish, that it feels like an addicting high. There is never a full stop like before where you can’t figure out a bug and have to stop and walk away and hope you fix it the next day. It just keeps going and going! The hours just fly by! Remember, YOU are the agent of your agent. YOU control it. It does not control you! Never turn off your brain. Just keep learning and building!