Optimizing AI Usage

Another AI post I know! I think after spending some time thoughtfully vibe coding though, I'm seeing a lot of benefits and getting better at the prompting side of things. I am particularly interested in having a really good AI workflow between different models. I used to stick with just one chat provider like openai's chatgpt or anthropic's claude but after being introduced to copilot, I'm playing around with the idea of figuring out what model is good at what and how to best work with their strengths

WHY AI?

As part of this approach is embracing AI as a manager because as I grow in my IT journey, I'll need to get better and better at approaching things from a managerial POV while understanding the POV of a "grunt". I'm hoping this makes me better at delivering high quality work efficiently.

Findings and workflow

I'm finding that I can put this into play a lot easier than I thought as long as I understand the strengths of the models which you can find out by using a lot of resources or listening to how people who know more than you describe their usage of AI.

Claude

Currently claude opus and sonnet(4.6 as of writing) are the brains of the operation, being really good at taking my idea and making that into a plan or prompt for other models. I can increase the effectiveness of this by feeding claude agent.md files which are now very easy to find since everyone has had time to develop different strategies. An agency agent.md file converted for my own projects works wonders. Sonnet is also a very good backend, powershell , python. batch and cmd programmer. I can make it more efficient by having opus plan as much as possible

GPT

Currently GPT 5.4 is my favorite model. With github's copilot, one prompt goes a looooooooooooooooooooong way. This model can use any effort of thinking and get a lot done in one go rather than needing to be approved for 50 different things in a way that wastes usage. It seems to intuit well and have very good tool usage. I use it for the same things I use sonnet for: backend, powershell , python. batch and cmd. I like it better than sonnet specifically because of how much it gets done. It gives me a better workflow than the approval hell

Gemini

Apparently Gemini 3/3.1 is the frontend engine that can. I didn't know how good this model was for frontend work, likely because I've been mostly doing powershell, batch and cmd scripting these days. Now that I've been told better by everyone on the internet, I can see that I get a lot of value from giving the backend and local scripts to gpt and sonnet while leaving the design work for gemini. It seems to handle react very well and that long context is overkill for that sort of work. I also get a lot out of using flash since I always have master prompt delivered from opus or handoff prompt from sonnet or gpt 5.4 which serves the same purpose

Takeway

Good Sh*t