Skills and Sub-Agents

In the last few months I have been fully immersed in development, and I have had to play with every new model, especially Claude. I could write down some kind of workflows or skills, but my workflow changes every week.

From the start of the AI era until around December 2025, I was a very document-first person. I am not like that anymore. Now I build skills and files, I build sub-agents and assign skills to them, and I manage those sub-agents in a far more organized way than before.

Before, we used ultrathink mode for everything. Our usage ran out faster, and then we saw the consequences. Later we got the ability to manage sub-agents, which gave us more control over context. Caching was also clearly explained in educational videos by the AI providers. I was already confused about some of this, and the new updates fixed everything.

I still have issues with memory. We are still the same as humans, we do not fully understand why memory exists. Skills files and sub-agents feel good and manageable right now. But sometimes the skills files do not get updated as the project's direction changes, so we need some kind of monitoring.

That is what I feel, because this differs a lot from project to project. It is not just about the team, or about whatever the frontier models can do. They do not know our workflow. They have analytics, stats, and transcripts of everything we do, but that may not fully cover our personal projects. That is why we have to figure out the best way ourselves.

I am building two really different categories of products. The astrology products need a very different context and workflow, because they rely on traditional knowledge, and they need higher-thinking models than the usual groundwork models. We realized this while building it. I built it alone on purpose, because I wanted to understand every issue and circumstance myself, not learn what was going on from someone's blog.

I use API-based pricing, and I have used every subscription from Pro to Max 20x. Max 5x is my default for day-to-day work, because I felt a significant difference between Pro and Max in model selection and output quality. I am not going into more detail here.

Right now my workflow is scaffolding. First the CLAUDE.md file, then the skills, then I plan the sub-agents and assign models to them. Then we start working from the plan, or from whatever docs we have.

Usually I follow the docs folder inside my project, though not always in the same pattern. When the work relies on traditional knowledge or specific domain knowledge, the docs folder is where I go. Scaffold first, then let the agents work.