Trying to sanity check how multi-step workflows are actually handled in production. The way I’m currently thinking about it: When building, it feels very agentic (using markdown like claude.md, rooms, skills, etc. to structure everything and have the AI help design the system). But in production, I’m trying to understand how control actually flows across multiple steps. Is it more like this: Option A (AI-led flow): - Trigger fires - Script collects data + relevant markdown - AI is called - AI responds and decides the next step - AI continues task-to-task from there OR Option B (Script-orchestrated flow): - Trigger fires - Script collects data + relevant markdown - AI is called for a specific task - AI returns output - Script takes that output, decides next step - Script loads new context/markdown - Script calls AI again (and this repeats for each step) So basically: 👉 Is the AI “driving” the workflow once triggered 👉 Or are scripts orchestrating each step and calling AI as needed Feels like it might be more B or somewhere in the middle, but not sure what’s actually best practice. Would love to know how others are structuring this in real systems.

A few people asked about my .md based task management system so sharing it here: https://github.com/rocleemusic/TaskSystem I should also mention I have my main workspace as a github repo because I work between 2 different computers at work and home, so I have a hook command, when i type /sync on a computer it pulls from github and then runs the task system. the manual call is : What's my focus for today? I'm also still tweaking the retrospective section since it's only come up once since starting to use it. There's a full breakdown and examples in the repo, feel free to ask questions or if you find something that's broken, it might be missing some context from my main Claude.md I've also included a build-instructions.md so you can give that to claude and have it built anywhere. it's large cuz it includes the schemas for the child folder files and i wanted to make it easy to use as a single file. But it was built thinking about how I would like to interact with Claude (like Jarvis). The main entry points for me are an ideas.md at root where I can throw any idea I suddenly have, and the main chat at start of session to get the 1-3 things most important to do today. It's grown a bit to include recurring tasks and flag high-priority things. The other goal was to offload the burden of remember what's important or to figure out what I have energy to do after work. I also wanted it as pure markdown incase an online service went down, I can still view the full list or just my focus, it's just a text file on my system. The thinking behind this was largely due to some realizations after watching Jake's videos. If we think of AI as a programming language, then what is it's infrastructure? What do I need to actually build a full app for something that I can describe with plain language and have AI serve as the interpretation layer to the computer below? Any workflow is really just a program codified into a step by step task list.

I'm curious what everyone here likes to do for fun (of course building stuff with Claude is fun too lol), and if you've applied any AI to your hobby. For me it's been super useful for DND planning and I find I get to stay in creative flow more. Curious what other people are doing